diff --git a/.ci/update_windows/update.py b/.ci/update_windows/update.py index ef9374c44..127247b2f 100755 --- a/.ci/update_windows/update.py +++ b/.ci/update_windows/update.py @@ -1,6 +1,9 @@ import pygit2 from datetime import datetime import sys +import os +import shutil +import filecmp def pull(repo, remote_name='origin', branch='master'): for remote in repo.remotes: @@ -42,7 +45,8 @@ def pull(repo, remote_name='origin', branch='master'): raise AssertionError('Unknown merge analysis result') pygit2.option(pygit2.GIT_OPT_SET_OWNER_VALIDATION, 0) -repo = pygit2.Repository(str(sys.argv[1])) +repo_path = str(sys.argv[1]) +repo = pygit2.Repository(repo_path) ident = pygit2.Signature('comfyui', 'comfy@ui') try: print("stashing current changes") @@ -51,7 +55,10 @@ except KeyError: print("nothing to stash") backup_branch_name = 'backup_branch_{}'.format(datetime.today().strftime('%Y-%m-%d_%H_%M_%S')) print("creating backup branch: {}".format(backup_branch_name)) -repo.branches.local.create(backup_branch_name, repo.head.peel()) +try: + repo.branches.local.create(backup_branch_name, repo.head.peel()) +except: + pass print("checking out master branch") branch = repo.lookup_branch('master') @@ -63,3 +70,41 @@ pull(repo) print("Done!") +self_update = True +if len(sys.argv) > 2: + self_update = '--skip_self_update' not in sys.argv + +update_py_path = os.path.realpath(__file__) +repo_update_py_path = os.path.join(repo_path, ".ci/update_windows/update.py") + +cur_path = os.path.dirname(update_py_path) + + +req_path = os.path.join(cur_path, "current_requirements.txt") +repo_req_path = os.path.join(repo_path, "requirements.txt") + + +def files_equal(file1, file2): + try: + return filecmp.cmp(file1, file2, shallow=False) + except: + return False + +def file_size(f): + try: + return os.path.getsize(f) + except: + return 0 + + +if self_update and not files_equal(update_py_path, repo_update_py_path) and file_size(repo_update_py_path) > 10: + shutil.copy(repo_update_py_path, os.path.join(cur_path, "update_new.py")) + exit() + +if not os.path.exists(req_path) or not files_equal(repo_req_path, req_path): + import subprocess + try: + subprocess.check_call([sys.executable, '-s', '-m', 'pip', 'install', '-r', repo_req_path]) + shutil.copy(repo_req_path, req_path) + except: + pass diff --git a/.ci/update_windows/update_comfyui.bat b/.ci/update_windows/update_comfyui.bat index 60d1e694f..bb08c0de0 100755 --- a/.ci/update_windows/update_comfyui.bat +++ b/.ci/update_windows/update_comfyui.bat @@ -1,2 +1,8 @@ +@echo off ..\python_embeded\python.exe .\update.py ..\ComfyUI\ -pause +if exist update_new.py ( + move /y update_new.py update.py + echo Running updater again since it got updated. + ..\python_embeded\python.exe .\update.py ..\ComfyUI\ --skip_self_update +) +if "%~1"=="" pause diff --git a/.ci/update_windows/update_comfyui_and_python_dependencies.bat b/.ci/update_windows/update_comfyui_and_python_dependencies.bat deleted file mode 100755 index b7308550d..000000000 --- a/.ci/update_windows/update_comfyui_and_python_dependencies.bat +++ /dev/null @@ -1,3 +0,0 @@ -..\python_embeded\python.exe .\update.py ..\ComfyUI\ -..\python_embeded\python.exe -s -m pip install --upgrade torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117 xformers -r ../ComfyUI/requirements.txt pygit2 -pause diff --git a/.ci/update_windows_cu118/update_comfyui_and_python_dependencies.bat b/.ci/update_windows_cu118/update_comfyui_and_python_dependencies.bat deleted file mode 100755 index c33adc0a7..000000000 --- a/.ci/update_windows_cu118/update_comfyui_and_python_dependencies.bat +++ /dev/null @@ -1,11 +0,0 @@ -@echo off -..\python_embeded\python.exe .\update.py ..\ComfyUI\ -echo -echo This will try to update pytorch and all python dependencies, if you get an error wait for pytorch/xformers to fix their stuff -echo You should not be running this anyways unless you really have to -echo -echo If you just want to update normally, close this and run update_comfyui.bat instead. -echo -pause -..\python_embeded\python.exe -s -m pip install --upgrade torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118 xformers -r ../ComfyUI/requirements.txt pygit2 -pause diff --git a/.github/workflows/windows_release_cu118_dependencies.yml b/.github/workflows/windows_release_cu118_dependencies.yml deleted file mode 100644 index 75c42b624..000000000 --- a/.github/workflows/windows_release_cu118_dependencies.yml +++ /dev/null @@ -1,71 +0,0 @@ -name: "Windows Release cu118 dependencies" - -on: - workflow_dispatch: -# push: -# branches: -# - master - -jobs: - build_dependencies: - env: - # you need at least cuda 5.0 for some of the stuff compiled here. - TORCH_CUDA_ARCH_LIST: "5.0+PTX 6.0 6.1 7.0 7.5 8.0 8.6 8.9" - FORCE_CUDA: 1 - MAX_JOBS: 1 # will crash otherwise - DISTUTILS_USE_SDK: 1 # otherwise distutils will complain on windows about multiple versions of msvc - XFORMERS_BUILD_TYPE: "Release" - runs-on: windows-latest - steps: - - name: Cache Built Dependencies - uses: actions/cache@v3 - id: cache-cu118_python_stuff - with: - path: cu118_python_deps.tar - key: ${{ runner.os }}-build-cu118 - - - if: steps.cache-cu118_python_stuff.outputs.cache-hit != 'true' - uses: actions/checkout@v3 - - - if: steps.cache-cu118_python_stuff.outputs.cache-hit != 'true' - uses: actions/setup-python@v4 - with: - python-version: '3.10.9' - - - if: steps.cache-cu118_python_stuff.outputs.cache-hit != 'true' - uses: comfyanonymous/cuda-toolkit@test - id: cuda-toolkit - with: - cuda: '11.8.0' - # copied from xformers github - - name: Setup MSVC - uses: ilammy/msvc-dev-cmd@v1 - - name: Configure Pagefile - # windows runners will OOM with many CUDA architectures - # we cheat here with a page file - uses: al-cheb/configure-pagefile-action@v1.3 - with: - minimum-size: 2GB - # really unfortunate: https://github.com/ilammy/msvc-dev-cmd#name-conflicts-with-shell-bash - - name: Remove link.exe - shell: bash - run: rm /usr/bin/link - - - if: steps.cache-cu118_python_stuff.outputs.cache-hit != 'true' - shell: bash - run: | - python -m pip wheel --no-cache-dir torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118 -r requirements.txt pygit2 -w ./temp_wheel_dir - python -m pip install --no-cache-dir ./temp_wheel_dir/* - echo installed basic - git clone --recurse-submodules https://github.com/facebookresearch/xformers.git - cd xformers - python -m pip install --no-cache-dir wheel setuptools twine - echo building xformers - python setup.py bdist_wheel -d ../temp_wheel_dir/ - cd .. - rm -rf xformers - ls -lah temp_wheel_dir - mv temp_wheel_dir cu118_python_deps - tar cf cu118_python_deps.tar cu118_python_deps - - diff --git a/.github/workflows/windows_release_cu118_dependencies_2.yml b/.github/workflows/windows_release_cu118_dependencies_2.yml deleted file mode 100644 index a7760b21e..000000000 --- a/.github/workflows/windows_release_cu118_dependencies_2.yml +++ /dev/null @@ -1,37 +0,0 @@ -name: "Windows Release cu118 dependencies 2" - -on: - workflow_dispatch: - inputs: - xformers: - description: 'xformers version' - required: true - type: string - default: "xformers" - -# push: -# branches: -# - master - -jobs: - build_dependencies: - runs-on: windows-latest - steps: - - uses: actions/checkout@v3 - - uses: actions/setup-python@v4 - with: - python-version: '3.10.9' - - - shell: bash - run: | - python -m pip wheel --no-cache-dir torch torchvision torchaudio ${{ inputs.xformers }} --extra-index-url https://download.pytorch.org/whl/cu118 -r requirements.txt pygit2 -w ./temp_wheel_dir - python -m pip install --no-cache-dir ./temp_wheel_dir/* - echo installed basic - ls -lah temp_wheel_dir - mv temp_wheel_dir cu118_python_deps - tar cf cu118_python_deps.tar cu118_python_deps - - - uses: actions/cache/save@v3 - with: - path: cu118_python_deps.tar - key: ${{ runner.os }}-build-cu118 diff --git a/.github/workflows/windows_release_cu118_package.yml b/.github/workflows/windows_release_cu118_package.yml deleted file mode 100644 index 0f0fbf280..000000000 --- a/.github/workflows/windows_release_cu118_package.yml +++ /dev/null @@ -1,79 +0,0 @@ -name: "Windows Release cu118 packaging" - -on: - workflow_dispatch: -# push: -# branches: -# - master - -jobs: - package_comfyui: - permissions: - contents: "write" - packages: "write" - pull-requests: "read" - runs-on: windows-latest - steps: - - uses: actions/cache/restore@v3 - id: cache - with: - path: cu118_python_deps.tar - key: ${{ runner.os }}-build-cu118 - - shell: bash - run: | - mv cu118_python_deps.tar ../ - cd .. - tar xf cu118_python_deps.tar - pwd - ls - - - uses: actions/checkout@v3 - with: - fetch-depth: 0 - persist-credentials: false - - shell: bash - run: | - cd .. - cp -r ComfyUI ComfyUI_copy - curl https://www.python.org/ftp/python/3.10.9/python-3.10.9-embed-amd64.zip -o python_embeded.zip - unzip python_embeded.zip -d python_embeded - cd python_embeded - echo 'import site' >> ./python310._pth - curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py - ./python.exe get-pip.py - ./python.exe -s -m pip install ../cu118_python_deps/* - sed -i '1i../ComfyUI' ./python310._pth - cd .. - - git clone https://github.com/comfyanonymous/taesd - cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/ - - mkdir ComfyUI_windows_portable - mv python_embeded ComfyUI_windows_portable - mv ComfyUI_copy ComfyUI_windows_portable/ComfyUI - - cd ComfyUI_windows_portable - - mkdir update - cp -r ComfyUI/.ci/update_windows/* ./update/ - cp -r ComfyUI/.ci/update_windows_cu118/* ./update/ - cp -r ComfyUI/.ci/windows_base_files/* ./ - - cd .. - - "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma -mx=8 -mfb=64 -md=32m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable - mv ComfyUI_windows_portable.7z ComfyUI/new_ComfyUI_windows_portable_nvidia_cu118_or_cpu.7z - - cd ComfyUI_windows_portable - python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu - - ls - - - name: Upload binaries to release - uses: svenstaro/upload-release-action@v2 - with: - repo_token: ${{ secrets.GITHUB_TOKEN }} - file: new_ComfyUI_windows_portable_nvidia_cu118_or_cpu.7z - tag: "latest" - overwrite: true - diff --git a/.github/workflows/windows_release_dependencies.yml b/.github/workflows/windows_release_dependencies.yml index aafe8a214..ffd3e2216 100644 --- a/.github/workflows/windows_release_dependencies.yml +++ b/.github/workflows/windows_release_dependencies.yml @@ -24,7 +24,7 @@ on: description: 'python patch version' required: true type: string - default: "6" + default: "8" # push: # branches: # - master @@ -41,10 +41,9 @@ jobs: - shell: bash run: | echo "@echo off - ..\python_embeded\python.exe .\update.py ..\ComfyUI\\ + call update_comfyui.bat nopause echo - - echo This will try to update pytorch and all python dependencies, if you get an error wait for pytorch/xformers to fix their stuff - echo You should not be running this anyways unless you really have to + echo This will try to update pytorch and all python dependencies. echo - echo If you just want to update normally, close this and run update_comfyui.bat instead. echo - diff --git a/.github/workflows/windows_release_nightly_pytorch.yml b/.github/workflows/windows_release_nightly_pytorch.yml index 90e09d27a..fa24a985c 100644 --- a/.github/workflows/windows_release_nightly_pytorch.yml +++ b/.github/workflows/windows_release_nightly_pytorch.yml @@ -7,7 +7,7 @@ on: description: 'cuda version' required: true type: string - default: "121" + default: "124" python_minor: description: 'python minor version' @@ -19,7 +19,7 @@ on: description: 'python patch version' required: true type: string - default: "1" + default: "3" # push: # branches: # - master @@ -49,7 +49,7 @@ jobs: echo 'import site' >> ./python3${{ inputs.python_minor }}._pth curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py ./python.exe get-pip.py - python -m pip wheel torch torchvision torchaudio --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu${{ inputs.cu }} -r ../ComfyUI/requirements.txt pygit2 -w ../temp_wheel_dir + python -m pip wheel torch torchvision torchaudio mpmath==1.3.0 numpy==1.26.4 --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu${{ inputs.cu }} -r ../ComfyUI/requirements.txt pygit2 -w ../temp_wheel_dir ls ../temp_wheel_dir ./python.exe -s -m pip install --pre ../temp_wheel_dir/* sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth @@ -68,7 +68,7 @@ jobs: cp -r ComfyUI/.ci/update_windows/* ./update/ cp -r ComfyUI/.ci/windows_base_files/* ./ - echo "..\python_embeded\python.exe .\update.py ..\ComfyUI\\ + echo "call update_comfyui.bat nopause ..\python_embeded\python.exe -s -m pip install --upgrade --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cu${{ inputs.cu }} -r ../ComfyUI/requirements.txt pygit2 pause" > ./update/update_comfyui_and_python_dependencies.bat cd .. diff --git a/.github/workflows/windows_release_package.yml b/.github/workflows/windows_release_package.yml index 87d37c24d..4e3cdabd2 100644 --- a/.github/workflows/windows_release_package.yml +++ b/.github/workflows/windows_release_package.yml @@ -19,7 +19,7 @@ on: description: 'python patch version' required: true type: string - default: "6" + default: "8" # push: # branches: # - master diff --git a/README.md b/README.md index 77347240d..8caf71842 100644 --- a/README.md +++ b/README.md @@ -11,7 +11,7 @@ This ui will let you design and execute advanced stable diffusion pipelines usin ## Features - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything. -- Fully supports SD1.x, SD2.x, [SDXL](https://comfyanonymous.github.io/ComfyUI_examples/sdxl/) and [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/) +- Fully supports SD1.x, SD2.x, [SDXL](https://comfyanonymous.github.io/ComfyUI_examples/sdxl/), [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/) and [Stable Cascade](https://comfyanonymous.github.io/ComfyUI_examples/stable_cascade/) - Asynchronous Queue system - Many optimizations: Only re-executes the parts of the workflow that changes between executions. - Command line option: ```--lowvram``` to make it work on GPUs with less than 3GB vram (enabled automatically on GPUs with low vram) @@ -46,7 +46,7 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git | Ctrl + Enter | Queue up current graph for generation | | Ctrl + Shift + Enter | Queue up current graph as first for generation | | Ctrl + Alt + Enter | Stop the current generation | -| Ctrl + Shift + Alt + Enter| Clear the generation queue, then stop the current generation | +| Ctrl + Shift + Alt + Enter| Clear the generation queue, then stop the current generation | | Ctrl + Z/Ctrl + Y | Undo/Redo | | Ctrl + S | Save workflow | | Ctrl + O | Load workflow | @@ -79,6 +79,8 @@ There is a portable standalone build for Windows that should work for running on Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you put your Stable Diffusion checkpoints/models (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints +If you have trouble extracting it, right click the file -> properties -> unblock + #### How do I share models between another UI and ComfyUI? See the [Config file](extra_model_paths.yaml.example) to set the search paths for models. In the standalone windows build you can find this file in the ComfyUI directory. Rename this file to extra_model_paths.yaml and edit it with your favorite text editor. @@ -95,16 +97,15 @@ Put your SD checkpoints (the huge ckpt/safetensors files) in: models/checkpoints Put your VAE in: models/vae -Note: pytorch stable does not support python 3.12 yet. If you have python 3.12 you will have to use the nightly version of pytorch. If you run into issues you should try python 3.11 instead. ### AMD GPUs (Linux only) AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version: -```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.6``` +```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.0``` -This is the command to install the nightly with ROCm 5.7 which has a python 3.12 package and might have some performance improvements: +This is the command to install the nightly with ROCm 6.0 which might have some performance improvements: -```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm5.7``` +```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.1``` ### NVIDIA @@ -112,9 +113,9 @@ Nvidia users should install stable pytorch using this command: ```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121``` -This is the command to install pytorch nightly instead which has a python 3.12 package and might have performance improvements: +This is the command to install pytorch nightly instead which might have performance improvements: -```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121``` +```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124``` #### Troubleshooting @@ -134,7 +135,16 @@ After this you should have everything installed and can proceed to running Comfy ### Others: -#### [Intel Arc](https://github.com/comfyanonymous/ComfyUI/discussions/476) +#### Intel GPUs + +Intel GPU support is available for all Intel GPUs supported by Intel's Extension for Pytorch (IPEX) with the support requirements listed in the [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) page. Choose your platform and method of install and follow the instructions. The steps are as follows: + +1. Start by installing the drivers or kernel listed or newer in the Installation page of IPEX linked above for Windows and Linux if needed. +1. Follow the instructions to install [Intel's oneAPI Basekit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit-download.html) for your platform. +1. Install the packages for IPEX using the instructions provided in the Installation page for your platform. +1. Follow the [ComfyUI manual installation](#manual-install-windows-linux) instructions for Windows and Linux and run ComfyUI normally as described above after everything is installed. + +Additional discussion and help can be found [here](https://github.com/comfyanonymous/ComfyUI/discussions/476). #### Apple Mac silicon @@ -143,7 +153,7 @@ You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS ve 1. Install pytorch nightly. For instructions, read the [Accelerated PyTorch training on Mac](https://developer.apple.com/metal/pytorch/) Apple Developer guide (make sure to install the latest pytorch nightly). 1. Follow the [ComfyUI manual installation](#manual-install-windows-linux) instructions for Windows and Linux. 1. Install the ComfyUI [dependencies](#dependencies). If you have another Stable Diffusion UI [you might be able to reuse the dependencies](#i-already-have-another-ui-for-stable-diffusion-installed-do-i-really-have-to-install-all-of-these-dependencies). -1. Launch ComfyUI by running `python main.py --force-fp16`. Note that --force-fp16 will only work if you installed the latest pytorch nightly. +1. Launch ComfyUI by running `python main.py` > **Note**: Remember to add your models, VAE, LoRAs etc. to the corresponding Comfy folders, as discussed in [ComfyUI manual installation](#manual-install-windows-linux). @@ -196,20 +206,20 @@ To use a textual inversion concepts/embeddings in a text prompt put them in the ```embedding:embedding_filename.pt``` -## How to increase generation speed? - -Make sure you use the regular loaders/Load Checkpoint node to load checkpoints. It will auto pick the right settings depending on your GPU. - -You can set this command line setting to disable the upcasting to fp32 in some cross attention operations which will increase your speed. Note that this will very likely give you black images on SD2.x models. If you use xformers or pytorch attention this option does not do anything. - -```--dont-upcast-attention``` - ## How to show high-quality previews? Use ```--preview-method auto``` to enable previews. The default installation includes a fast latent preview method that's low-resolution. To enable higher-quality previews with [TAESD](https://github.com/madebyollin/taesd), download the [taesd_decoder.pth](https://github.com/madebyollin/taesd/raw/main/taesd_decoder.pth) (for SD1.x and SD2.x) and [taesdxl_decoder.pth](https://github.com/madebyollin/taesd/raw/main/taesdxl_decoder.pth) (for SDXL) models and place them in the `models/vae_approx` folder. Once they're installed, restart ComfyUI to enable high-quality previews. +## How to use TLS/SSL? +Generate a self-signed certificate (not appropriate for shared/production use) and key by running the command: `openssl req -x509 -newkey rsa:4096 -keyout key.pem -out cert.pem -sha256 -days 3650 -nodes -subj "/C=XX/ST=StateName/L=CityName/O=CompanyName/OU=CompanySectionName/CN=CommonNameOrHostname"` + +Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app will now be accessible with `https://...` instead of `http://...`. + +> Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above. +

If you use a container, note that the volume mount `-v` can be a relative path so `... -v ".\:/openssl-certs" ...` would create the key & cert files in the current directory of your command prompt or powershell terminal. + ## Support and dev channel [Matrix space: #comfyui_space:matrix.org](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) (it's like discord but open source). diff --git a/comfy/cldm/cldm.py b/comfy/cldm/cldm.py index 5eee5a51c..28076dd92 100644 --- a/comfy/cldm/cldm.py +++ b/comfy/cldm/cldm.py @@ -52,6 +52,7 @@ class ControlNet(nn.Module): adm_in_channels=None, transformer_depth_middle=None, transformer_depth_output=None, + attn_precision=None, device=None, operations=comfy.ops.disable_weight_init, **kwargs, @@ -202,7 +203,7 @@ class ControlNet(nn.Module): SpatialTransformer( ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim, disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, - use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations + use_checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations ) ) self.input_blocks.append(TimestepEmbedSequential(*layers)) @@ -262,7 +263,7 @@ class ControlNet(nn.Module): mid_block += [SpatialTransformer( # always uses a self-attn ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim, disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer, - use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations + use_checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations ), ResBlock( ch, diff --git a/comfy/cli_args.py b/comfy/cli_args.py index b4bbfbfab..b8ac9bc69 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -35,6 +35,8 @@ parser = argparse.ArgumentParser() parser.add_argument("--listen", type=str, default="127.0.0.1", metavar="IP", nargs="?", const="0.0.0.0", help="Specify the IP address to listen on (default: 127.0.0.1). If --listen is provided without an argument, it defaults to 0.0.0.0. (listens on all)") parser.add_argument("--port", type=int, default=8188, help="Set the listen port.") +parser.add_argument("--tls-keyfile", type=str, help="Path to TLS (SSL) key file. Enables TLS, makes app accessible at https://... requires --tls-certfile to function") +parser.add_argument("--tls-certfile", type=str, help="Path to TLS (SSL) certificate file. Enables TLS, makes app accessible at https://... requires --tls-keyfile to function") parser.add_argument("--enable-cors-header", type=str, default=None, metavar="ORIGIN", nargs="?", const="*", help="Enable CORS (Cross-Origin Resource Sharing) with optional origin or allow all with default '*'.") parser.add_argument("--max-upload-size", type=float, default=100, help="Set the maximum upload size in MB.") @@ -49,7 +51,6 @@ cm_group = parser.add_mutually_exclusive_group() cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).") cm_group.add_argument("--disable-cuda-malloc", action="store_true", help="Disable cudaMallocAsync.") -parser.add_argument("--dont-upcast-attention", action="store_true", help="Disable upcasting of attention. Can boost speed but increase the chances of black images.") fp_group = parser.add_mutually_exclusive_group() fp_group.add_argument("--force-fp32", action="store_true", help="Force fp32 (If this makes your GPU work better please report it).") @@ -94,6 +95,11 @@ attn_group.add_argument("--use-pytorch-cross-attention", action="store_true", he parser.add_argument("--disable-xformers", action="store_true", help="Disable xformers.") +upcast = parser.add_mutually_exclusive_group() +upcast.add_argument("--force-upcast-attention", action="store_true", help="Force enable attention upcasting, please report if it fixes black images.") +upcast.add_argument("--dont-upcast-attention", action="store_true", help="Disable all upcasting of attention. Should be unnecessary except for debugging.") + + vram_group = parser.add_mutually_exclusive_group() vram_group.add_argument("--gpu-only", action="store_true", help="Store and run everything (text encoders/CLIP models, etc... on the GPU).") vram_group.add_argument("--highvram", action="store_true", help="By default models will be unloaded to CPU memory after being used. This option keeps them in GPU memory.") @@ -114,6 +120,9 @@ parser.add_argument("--disable-metadata", action="store_true", help="Disable sav parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.") +parser.add_argument("--verbose", action="store_true", help="Enables more debug prints.") + + if comfy.options.args_parsing: args = parser.parse_args() else: @@ -124,3 +133,10 @@ if args.windows_standalone_build: if args.disable_auto_launch: args.auto_launch = False + +import logging +logging_level = logging.INFO +if args.verbose: + logging_level = logging.DEBUG + +logging.basicConfig(format="%(message)s", level=logging_level) diff --git a/comfy/clip_model.py b/comfy/clip_model.py index 09e7bbca1..14f43c568 100644 --- a/comfy/clip_model.py +++ b/comfy/clip_model.py @@ -97,7 +97,7 @@ class CLIPTextModel_(torch.nn.Module): x = self.embeddings(input_tokens) mask = None if attention_mask is not None: - mask = 1.0 - attention_mask.to(x.dtype).unsqueeze(1).unsqueeze(1).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1]) + mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1]) mask = mask.masked_fill(mask.to(torch.bool), float("-inf")) causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1) @@ -119,6 +119,9 @@ class CLIPTextModel(torch.nn.Module): super().__init__() self.num_layers = config_dict["num_hidden_layers"] self.text_model = CLIPTextModel_(config_dict, dtype, device, operations) + embed_dim = config_dict["hidden_size"] + self.text_projection = operations.Linear(embed_dim, embed_dim, bias=False, dtype=dtype, device=device) + self.text_projection.weight.copy_(torch.eye(embed_dim)) self.dtype = dtype def get_input_embeddings(self): @@ -128,7 +131,10 @@ class CLIPTextModel(torch.nn.Module): self.text_model.embeddings.token_embedding = embeddings def forward(self, *args, **kwargs): - return self.text_model(*args, **kwargs) + x = self.text_model(*args, **kwargs) + out = self.text_projection(x[2]) + return (x[0], x[1], out, x[2]) + class CLIPVisionEmbeddings(torch.nn.Module): def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, dtype=None, device=None, operations=None): diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py index 4564fcfb2..acc86be85 100644 --- a/comfy/clip_vision.py +++ b/comfy/clip_vision.py @@ -1,8 +1,8 @@ -from .utils import load_torch_file, transformers_convert, common_upscale +from .utils import load_torch_file, transformers_convert, state_dict_prefix_replace import os import torch -import contextlib import json +import logging import comfy.ops import comfy.model_patcher @@ -41,9 +41,13 @@ class ClipVisionModel(): self.model.eval() self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) + def load_sd(self, sd): return self.model.load_state_dict(sd, strict=False) + def get_sd(self): + return self.model.state_dict() + def encode_image(self, image): comfy.model_management.load_model_gpu(self.patcher) pixel_values = clip_preprocess(image.to(self.load_device)).float() @@ -76,6 +80,9 @@ def convert_to_transformers(sd, prefix): sd['visual_projection.weight'] = sd.pop("{}proj".format(prefix)).transpose(0, 1) sd = transformers_convert(sd, prefix, "vision_model.", 48) + else: + replace_prefix = {prefix: ""} + sd = state_dict_prefix_replace(sd, replace_prefix) return sd def load_clipvision_from_sd(sd, prefix="", convert_keys=False): @@ -93,7 +100,7 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False): clip = ClipVisionModel(json_config) m, u = clip.load_sd(sd) if len(m) > 0: - print("missing clip vision:", m) + logging.warning("missing clip vision: {}".format(m)) u = set(u) keys = list(sd.keys()) for k in keys: diff --git a/comfy/conds.py b/comfy/conds.py index 6cff25184..23fa48872 100644 --- a/comfy/conds.py +++ b/comfy/conds.py @@ -1,4 +1,3 @@ -import enum import torch import math import comfy.utils diff --git a/comfy/controlnet.py b/comfy/controlnet.py index df474201c..8cf4a61a6 100644 --- a/comfy/controlnet.py +++ b/comfy/controlnet.py @@ -1,7 +1,7 @@ import torch import math import os -import contextlib +import logging import comfy.utils import comfy.model_management import comfy.model_detection @@ -10,6 +10,7 @@ import comfy.ops import comfy.cldm.cldm import comfy.t2i_adapter.adapter +import comfy.ldm.cascade.controlnet def broadcast_image_to(tensor, target_batch_size, batched_number): @@ -38,6 +39,8 @@ class ControlBase: self.timestep_percent_range = (0.0, 1.0) self.global_average_pooling = False self.timestep_range = None + self.compression_ratio = 8 + self.upscale_algorithm = 'nearest-exact' if device is None: device = comfy.model_management.get_torch_device() @@ -78,6 +81,8 @@ class ControlBase: c.strength = self.strength c.timestep_percent_range = self.timestep_percent_range c.global_average_pooling = self.global_average_pooling + c.compression_ratio = self.compression_ratio + c.upscale_algorithm = self.upscale_algorithm def inference_memory_requirements(self, dtype): if self.previous_controlnet is not None: @@ -133,11 +138,13 @@ class ControlBase: return out class ControlNet(ControlBase): - def __init__(self, control_model, global_average_pooling=False, device=None, load_device=None, manual_cast_dtype=None): + def __init__(self, control_model=None, global_average_pooling=False, device=None, load_device=None, manual_cast_dtype=None): super().__init__(device) self.control_model = control_model self.load_device = load_device - self.control_model_wrapped = comfy.model_patcher.ModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device()) + if control_model is not None: + self.control_model_wrapped = comfy.model_patcher.ModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device()) + self.global_average_pooling = global_average_pooling self.model_sampling_current = None self.manual_cast_dtype = manual_cast_dtype @@ -159,15 +166,15 @@ class ControlNet(ControlBase): dtype = self.manual_cast_dtype output_dtype = x_noisy.dtype - if self.cond_hint is None or x_noisy.shape[2] * 8 != self.cond_hint.shape[2] or x_noisy.shape[3] * 8 != self.cond_hint.shape[3]: + if self.cond_hint is None or x_noisy.shape[2] * self.compression_ratio != self.cond_hint.shape[2] or x_noisy.shape[3] * self.compression_ratio != self.cond_hint.shape[3]: if self.cond_hint is not None: del self.cond_hint self.cond_hint = None - self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * 8, x_noisy.shape[2] * 8, 'nearest-exact', "center").to(dtype).to(self.device) + self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * self.compression_ratio, x_noisy.shape[2] * self.compression_ratio, self.upscale_algorithm, "center").to(dtype).to(self.device) if x_noisy.shape[0] != self.cond_hint.shape[0]: self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number) - context = cond['c_crossattn'] + context = cond.get('crossattn_controlnet', cond['c_crossattn']) y = cond.get('y', None) if y is not None: y = y.to(dtype) @@ -178,7 +185,9 @@ class ControlNet(ControlBase): return self.control_merge(None, control, control_prev, output_dtype) def copy(self): - c = ControlNet(self.control_model, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype) + c = ControlNet(None, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype) + c.control_model = self.control_model + c.control_model_wrapped = self.control_model_wrapped self.copy_to(c) return c @@ -196,7 +205,7 @@ class ControlNet(ControlBase): super().cleanup() class ControlLoraOps: - class Linear(torch.nn.Module): + class Linear(torch.nn.Module, comfy.ops.CastWeightBiasOp): def __init__(self, in_features: int, out_features: int, bias: bool = True, device=None, dtype=None) -> None: factory_kwargs = {'device': device, 'dtype': dtype} @@ -215,7 +224,7 @@ class ControlLoraOps: else: return torch.nn.functional.linear(input, weight, bias) - class Conv2d(torch.nn.Module): + class Conv2d(torch.nn.Module, comfy.ops.CastWeightBiasOp): def __init__( self, in_channels, @@ -288,13 +297,13 @@ class ControlLora(ControlNet): for k in sd: weight = sd[k] try: - comfy.utils.set_attr(self.control_model, k, weight) + comfy.utils.set_attr_param(self.control_model, k, weight) except: pass for k in self.control_weights: if k not in {"lora_controlnet"}: - comfy.utils.set_attr(self.control_model, k, self.control_weights[k].to(dtype).to(comfy.model_management.get_torch_device())) + comfy.utils.set_attr_param(self.control_model, k, self.control_weights[k].to(dtype).to(comfy.model_management.get_torch_device())) def copy(self): c = ControlLora(self.control_weights, global_average_pooling=self.global_average_pooling) @@ -319,9 +328,10 @@ def load_controlnet(ckpt_path, model=None): return ControlLora(controlnet_data) controlnet_config = None + supported_inference_dtypes = None + if "controlnet_cond_embedding.conv_in.weight" in controlnet_data: #diffusers format - unet_dtype = comfy.model_management.unet_dtype() - controlnet_config = comfy.model_detection.unet_config_from_diffusers_unet(controlnet_data, unet_dtype) + controlnet_config = comfy.model_detection.unet_config_from_diffusers_unet(controlnet_data) diffusers_keys = comfy.utils.unet_to_diffusers(controlnet_config) diffusers_keys["controlnet_mid_block.weight"] = "middle_block_out.0.weight" diffusers_keys["controlnet_mid_block.bias"] = "middle_block_out.0.bias" @@ -362,7 +372,7 @@ def load_controlnet(ckpt_path, model=None): leftover_keys = controlnet_data.keys() if len(leftover_keys) > 0: - print("leftover keys:", leftover_keys) + logging.warning("leftover keys: {}".format(leftover_keys)) controlnet_data = new_sd pth_key = 'control_model.zero_convs.0.0.weight' @@ -377,16 +387,24 @@ def load_controlnet(ckpt_path, model=None): else: net = load_t2i_adapter(controlnet_data) if net is None: - print("error checkpoint does not contain controlnet or t2i adapter data", ckpt_path) + logging.error("error checkpoint does not contain controlnet or t2i adapter data {}".format(ckpt_path)) return net if controlnet_config is None: - unet_dtype = comfy.model_management.unet_dtype() - controlnet_config = comfy.model_detection.model_config_from_unet(controlnet_data, prefix, unet_dtype, True).unet_config + model_config = comfy.model_detection.model_config_from_unet(controlnet_data, prefix, True) + supported_inference_dtypes = model_config.supported_inference_dtypes + controlnet_config = model_config.unet_config + load_device = comfy.model_management.get_torch_device() + if supported_inference_dtypes is None: + unet_dtype = comfy.model_management.unet_dtype() + else: + unet_dtype = comfy.model_management.unet_dtype(supported_dtypes=supported_inference_dtypes) + manual_cast_dtype = comfy.model_management.unet_manual_cast(unet_dtype, load_device) if manual_cast_dtype is not None: controlnet_config["operations"] = comfy.ops.manual_cast + controlnet_config["dtype"] = unet_dtype controlnet_config.pop("out_channels") controlnet_config["hint_channels"] = controlnet_data["{}input_hint_block.0.weight".format(prefix)].shape[1] control_model = comfy.cldm.cldm.ControlNet(**controlnet_config) @@ -404,7 +422,7 @@ def load_controlnet(ckpt_path, model=None): cd = controlnet_data[x] cd += model_sd[sd_key].type(cd.dtype).to(cd.device) else: - print("WARNING: Loaded a diff controlnet without a model. It will very likely not work.") + logging.warning("WARNING: Loaded a diff controlnet without a model. It will very likely not work.") class WeightsLoader(torch.nn.Module): pass @@ -413,7 +431,12 @@ def load_controlnet(ckpt_path, model=None): missing, unexpected = w.load_state_dict(controlnet_data, strict=False) else: missing, unexpected = control_model.load_state_dict(controlnet_data, strict=False) - print(missing, unexpected) + + if len(missing) > 0: + logging.warning("missing controlnet keys: {}".format(missing)) + + if len(unexpected) > 0: + logging.debug("unexpected controlnet keys: {}".format(unexpected)) global_average_pooling = False filename = os.path.splitext(ckpt_path)[0] @@ -424,11 +447,13 @@ def load_controlnet(ckpt_path, model=None): return control class T2IAdapter(ControlBase): - def __init__(self, t2i_model, channels_in, device=None): + def __init__(self, t2i_model, channels_in, compression_ratio, upscale_algorithm, device=None): super().__init__(device) self.t2i_model = t2i_model self.channels_in = channels_in self.control_input = None + self.compression_ratio = compression_ratio + self.upscale_algorithm = upscale_algorithm def scale_image_to(self, width, height): unshuffle_amount = self.t2i_model.unshuffle_amount @@ -448,13 +473,13 @@ class T2IAdapter(ControlBase): else: return None - if self.cond_hint is None or x_noisy.shape[2] * 8 != self.cond_hint.shape[2] or x_noisy.shape[3] * 8 != self.cond_hint.shape[3]: + if self.cond_hint is None or x_noisy.shape[2] * self.compression_ratio != self.cond_hint.shape[2] or x_noisy.shape[3] * self.compression_ratio != self.cond_hint.shape[3]: if self.cond_hint is not None: del self.cond_hint self.control_input = None self.cond_hint = None - width, height = self.scale_image_to(x_noisy.shape[3] * 8, x_noisy.shape[2] * 8) - self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, width, height, 'nearest-exact', "center").float().to(self.device) + width, height = self.scale_image_to(x_noisy.shape[3] * self.compression_ratio, x_noisy.shape[2] * self.compression_ratio) + self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, width, height, self.upscale_algorithm, "center").float().to(self.device) if self.channels_in == 1 and self.cond_hint.shape[1] > 1: self.cond_hint = torch.mean(self.cond_hint, 1, keepdim=True) if x_noisy.shape[0] != self.cond_hint.shape[0]: @@ -473,11 +498,14 @@ class T2IAdapter(ControlBase): return self.control_merge(control_input, mid, control_prev, x_noisy.dtype) def copy(self): - c = T2IAdapter(self.t2i_model, self.channels_in) + c = T2IAdapter(self.t2i_model, self.channels_in, self.compression_ratio, self.upscale_algorithm) self.copy_to(c) return c def load_t2i_adapter(t2i_data): + compression_ratio = 8 + upscale_algorithm = 'nearest-exact' + if 'adapter' in t2i_data: t2i_data = t2i_data['adapter'] if 'adapter.body.0.resnets.0.block1.weight' in t2i_data: #diffusers format @@ -505,13 +533,22 @@ def load_t2i_adapter(t2i_data): if cin == 256 or cin == 768: xl = True model_ad = comfy.t2i_adapter.adapter.Adapter(cin=cin, channels=[channel, channel*2, channel*4, channel*4][:4], nums_rb=2, ksize=ksize, sk=True, use_conv=use_conv, xl=xl) + elif "backbone.0.0.weight" in keys: + model_ad = comfy.ldm.cascade.controlnet.ControlNet(c_in=t2i_data['backbone.0.0.weight'].shape[1], proj_blocks=[0, 4, 8, 12, 51, 55, 59, 63]) + compression_ratio = 32 + upscale_algorithm = 'bilinear' + elif "backbone.10.blocks.0.weight" in keys: + model_ad = comfy.ldm.cascade.controlnet.ControlNet(c_in=t2i_data['backbone.0.weight'].shape[1], bottleneck_mode="large", proj_blocks=[0, 4, 8, 12, 51, 55, 59, 63]) + compression_ratio = 1 + upscale_algorithm = 'nearest-exact' else: return None + missing, unexpected = model_ad.load_state_dict(t2i_data) if len(missing) > 0: - print("t2i missing", missing) + logging.warning("t2i missing {}".format(missing)) if len(unexpected) > 0: - print("t2i unexpected", unexpected) + logging.debug("t2i unexpected {}".format(unexpected)) - return T2IAdapter(model_ad, model_ad.input_channels) + return T2IAdapter(model_ad, model_ad.input_channels, compression_ratio, upscale_algorithm) diff --git a/comfy/diffusers_convert.py b/comfy/diffusers_convert.py index a9eb9302f..ed2a45fea 100644 --- a/comfy/diffusers_convert.py +++ b/comfy/diffusers_convert.py @@ -1,5 +1,6 @@ import re import torch +import logging # conversion code from https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_stable_diffusion.py @@ -177,7 +178,7 @@ def convert_vae_state_dict(vae_state_dict): for k, v in new_state_dict.items(): for weight_name in weights_to_convert: if f"mid.attn_1.{weight_name}.weight" in k: - print(f"Reshaping {k} for SD format") + logging.debug(f"Reshaping {k} for SD format") new_state_dict[k] = reshape_weight_for_sd(v) return new_state_dict @@ -205,6 +206,21 @@ textenc_pattern = re.compile("|".join(protected.keys())) # Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp code2idx = {"q": 0, "k": 1, "v": 2} +# This function exists because at the time of writing torch.cat can't do fp8 with cuda +def cat_tensors(tensors): + x = 0 + for t in tensors: + x += t.shape[0] + + shape = [x] + list(tensors[0].shape)[1:] + out = torch.empty(shape, device=tensors[0].device, dtype=tensors[0].dtype) + + x = 0 + for t in tensors: + out[x:x + t.shape[0]] = t + x += t.shape[0] + + return out def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""): new_state_dict = {} @@ -237,20 +253,24 @@ def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""): capture_qkv_bias[k_pre][code2idx[k_code]] = v continue - relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k) - new_state_dict[relabelled_key] = v + text_proj = "transformer.text_projection.weight" + if k.endswith(text_proj): + new_state_dict[k.replace(text_proj, "text_projection")] = v.transpose(0, 1).contiguous() + else: + relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k) + new_state_dict[relabelled_key] = v for k_pre, tensors in capture_qkv_weight.items(): if None in tensors: raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing") relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre) - new_state_dict[relabelled_key + ".in_proj_weight"] = torch.cat(tensors) + new_state_dict[relabelled_key + ".in_proj_weight"] = cat_tensors(tensors) for k_pre, tensors in capture_qkv_bias.items(): if None in tensors: raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing") relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre) - new_state_dict[relabelled_key + ".in_proj_bias"] = torch.cat(tensors) + new_state_dict[relabelled_key + ".in_proj_bias"] = cat_tensors(tensors) return new_state_dict diff --git a/comfy/diffusers_load.py b/comfy/diffusers_load.py index c0b420e79..98b888a19 100644 --- a/comfy/diffusers_load.py +++ b/comfy/diffusers_load.py @@ -1,4 +1,3 @@ -import json import os import comfy.sd diff --git a/comfy/extra_samplers/uni_pc.py b/comfy/extra_samplers/uni_pc.py index 08bf0fc9e..a30d1d03f 100644 --- a/comfy/extra_samplers/uni_pc.py +++ b/comfy/extra_samplers/uni_pc.py @@ -358,9 +358,6 @@ class UniPC: thresholding=False, max_val=1., variant='bh1', - noise_mask=None, - masked_image=None, - noise=None, ): """Construct a UniPC. @@ -372,9 +369,6 @@ class UniPC: self.predict_x0 = predict_x0 self.thresholding = thresholding self.max_val = max_val - self.noise_mask = noise_mask - self.masked_image = masked_image - self.noise = noise def dynamic_thresholding_fn(self, x0, t=None): """ @@ -391,10 +385,7 @@ class UniPC: """ Return the noise prediction model. """ - if self.noise_mask is not None: - return self.model(x, t) * self.noise_mask - else: - return self.model(x, t) + return self.model(x, t) def data_prediction_fn(self, x, t): """ @@ -409,8 +400,6 @@ class UniPC: s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1) s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims) x0 = torch.clamp(x0, -s, s) / s - if self.noise_mask is not None: - x0 = x0 * self.noise_mask + (1. - self.noise_mask) * self.masked_image return x0 def model_fn(self, x, t): @@ -723,8 +712,6 @@ class UniPC: assert timesteps.shape[0] - 1 == steps # with torch.no_grad(): for step_index in trange(steps, disable=disable_pbar): - if self.noise_mask is not None: - x = x * self.noise_mask + (1. - self.noise_mask) * (self.masked_image * self.noise_schedule.marginal_alpha(timesteps[step_index]) + self.noise * self.noise_schedule.marginal_std(timesteps[step_index])) if step_index == 0: vec_t = timesteps[0].expand((x.shape[0])) model_prev_list = [self.model_fn(x, vec_t)] @@ -766,7 +753,7 @@ class UniPC: model_x = self.model_fn(x, vec_t) model_prev_list[-1] = model_x if callback is not None: - callback(step_index, model_prev_list[-1], x, steps) + callback({'x': x, 'i': step_index, 'denoised': model_prev_list[-1]}) else: raise NotImplementedError() # if denoise_to_zero: @@ -858,7 +845,7 @@ def predict_eps_sigma(model, input, sigma_in, **kwargs): return (input - model(input, sigma_in, **kwargs)) / sigma -def sample_unipc(model, noise, image, sigmas, max_denoise, extra_args=None, callback=None, disable=False, noise_mask=None, variant='bh1'): +def sample_unipc(model, noise, sigmas, extra_args=None, callback=None, disable=False, variant='bh1'): timesteps = sigmas.clone() if sigmas[-1] == 0: timesteps = sigmas[:] @@ -867,16 +854,7 @@ def sample_unipc(model, noise, image, sigmas, max_denoise, extra_args=None, call timesteps = sigmas.clone() ns = SigmaConvert() - if image is not None: - img = image * ns.marginal_alpha(timesteps[0]) - if max_denoise: - noise_mult = 1.0 - else: - noise_mult = ns.marginal_std(timesteps[0]) - img += noise * noise_mult - else: - img = noise - + noise = noise / torch.sqrt(1.0 + timesteps[0] ** 2.0) model_type = "noise" model_fn = model_wrapper( @@ -888,7 +866,10 @@ def sample_unipc(model, noise, image, sigmas, max_denoise, extra_args=None, call ) order = min(3, len(timesteps) - 2) - uni_pc = UniPC(model_fn, ns, predict_x0=True, thresholding=False, noise_mask=noise_mask, masked_image=image, noise=noise, variant=variant) - x = uni_pc.sample(img, timesteps=timesteps, skip_type="time_uniform", method="multistep", order=order, lower_order_final=True, callback=callback, disable_pbar=disable) + uni_pc = UniPC(model_fn, ns, predict_x0=True, thresholding=False, variant=variant) + x = uni_pc.sample(noise, timesteps=timesteps, skip_type="time_uniform", method="multistep", order=order, lower_order_final=True, callback=callback, disable_pbar=disable) x /= ns.marginal_alpha(timesteps[-1]) return x + +def sample_unipc_bh2(model, noise, sigmas, extra_args=None, callback=None, disable=False): + return sample_unipc(model, noise, sigmas, extra_args, callback, disable, variant='bh2') \ No newline at end of file diff --git a/comfy/gligen.py b/comfy/gligen.py index 8d182839e..592522767 100644 --- a/comfy/gligen.py +++ b/comfy/gligen.py @@ -1,8 +1,9 @@ import torch -from torch import nn, einsum +from torch import nn from .ldm.modules.attention import CrossAttention from inspect import isfunction - +import comfy.ops +ops = comfy.ops.manual_cast def exists(val): return val is not None @@ -22,7 +23,7 @@ def default(val, d): class GEGLU(nn.Module): def __init__(self, dim_in, dim_out): super().__init__() - self.proj = nn.Linear(dim_in, dim_out * 2) + self.proj = ops.Linear(dim_in, dim_out * 2) def forward(self, x): x, gate = self.proj(x).chunk(2, dim=-1) @@ -35,14 +36,14 @@ class FeedForward(nn.Module): inner_dim = int(dim * mult) dim_out = default(dim_out, dim) project_in = nn.Sequential( - nn.Linear(dim, inner_dim), + ops.Linear(dim, inner_dim), nn.GELU() ) if not glu else GEGLU(dim, inner_dim) self.net = nn.Sequential( project_in, nn.Dropout(dropout), - nn.Linear(inner_dim, dim_out) + ops.Linear(inner_dim, dim_out) ) def forward(self, x): @@ -57,11 +58,12 @@ class GatedCrossAttentionDense(nn.Module): query_dim=query_dim, context_dim=context_dim, heads=n_heads, - dim_head=d_head) + dim_head=d_head, + operations=ops) self.ff = FeedForward(query_dim, glu=True) - self.norm1 = nn.LayerNorm(query_dim) - self.norm2 = nn.LayerNorm(query_dim) + self.norm1 = ops.LayerNorm(query_dim) + self.norm2 = ops.LayerNorm(query_dim) self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.))) self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.))) @@ -87,17 +89,18 @@ class GatedSelfAttentionDense(nn.Module): # we need a linear projection since we need cat visual feature and obj # feature - self.linear = nn.Linear(context_dim, query_dim) + self.linear = ops.Linear(context_dim, query_dim) self.attn = CrossAttention( query_dim=query_dim, context_dim=query_dim, heads=n_heads, - dim_head=d_head) + dim_head=d_head, + operations=ops) self.ff = FeedForward(query_dim, glu=True) - self.norm1 = nn.LayerNorm(query_dim) - self.norm2 = nn.LayerNorm(query_dim) + self.norm1 = ops.LayerNorm(query_dim) + self.norm2 = ops.LayerNorm(query_dim) self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.))) self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.))) @@ -126,14 +129,14 @@ class GatedSelfAttentionDense2(nn.Module): # we need a linear projection since we need cat visual feature and obj # feature - self.linear = nn.Linear(context_dim, query_dim) + self.linear = ops.Linear(context_dim, query_dim) self.attn = CrossAttention( - query_dim=query_dim, context_dim=query_dim, dim_head=d_head) + query_dim=query_dim, context_dim=query_dim, dim_head=d_head, operations=ops) self.ff = FeedForward(query_dim, glu=True) - self.norm1 = nn.LayerNorm(query_dim) - self.norm2 = nn.LayerNorm(query_dim) + self.norm1 = ops.LayerNorm(query_dim) + self.norm2 = ops.LayerNorm(query_dim) self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.))) self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.))) @@ -201,11 +204,11 @@ class PositionNet(nn.Module): self.position_dim = fourier_freqs * 2 * 4 # 2 is sin&cos, 4 is xyxy self.linears = nn.Sequential( - nn.Linear(self.in_dim + self.position_dim, 512), + ops.Linear(self.in_dim + self.position_dim, 512), nn.SiLU(), - nn.Linear(512, 512), + ops.Linear(512, 512), nn.SiLU(), - nn.Linear(512, out_dim), + ops.Linear(512, out_dim), ) self.null_positive_feature = torch.nn.Parameter( @@ -215,16 +218,15 @@ class PositionNet(nn.Module): def forward(self, boxes, masks, positive_embeddings): B, N, _ = boxes.shape - dtype = self.linears[0].weight.dtype - masks = masks.unsqueeze(-1).to(dtype) - positive_embeddings = positive_embeddings.to(dtype) + masks = masks.unsqueeze(-1) + positive_embeddings = positive_embeddings # embedding position (it may includes padding as placeholder) - xyxy_embedding = self.fourier_embedder(boxes.to(dtype)) # B*N*4 --> B*N*C + xyxy_embedding = self.fourier_embedder(boxes) # B*N*4 --> B*N*C # learnable null embedding - positive_null = self.null_positive_feature.view(1, 1, -1) - xyxy_null = self.null_position_feature.view(1, 1, -1) + positive_null = self.null_positive_feature.to(device=boxes.device, dtype=boxes.dtype).view(1, 1, -1) + xyxy_null = self.null_position_feature.to(device=boxes.device, dtype=boxes.dtype).view(1, 1, -1) # replace padding with learnable null embedding positive_embeddings = positive_embeddings * \ @@ -251,7 +253,7 @@ class Gligen(nn.Module): def func(x, extra_options): key = extra_options["transformer_index"] module = self.module_list[key] - return module(x, objs) + return module(x, objs.to(device=x.device, dtype=x.dtype)) return func def set_position(self, latent_image_shape, position_params, device): diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index 761c2e0ef..f9b281894 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -527,6 +527,9 @@ def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None, @torch.no_grad() def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2): """DPM-Solver++ (stochastic).""" + if len(sigmas) <= 1: + return x + sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max() seed = extra_args.get("seed", None) noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler @@ -595,6 +598,8 @@ def sample_dpmpp_2m(model, x, sigmas, extra_args=None, callback=None, disable=No @torch.no_grad() def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'): """DPM-Solver++(2M) SDE.""" + if len(sigmas) <= 1: + return x if solver_type not in {'heun', 'midpoint'}: raise ValueError('solver_type must be \'heun\' or \'midpoint\'') @@ -642,6 +647,9 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None): """DPM-Solver++(3M) SDE.""" + if len(sigmas) <= 1: + return x + seed = extra_args.get("seed", None) sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max() noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler @@ -690,18 +698,27 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl @torch.no_grad() def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None): + if len(sigmas) <= 1: + return x + sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max() noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler) @torch.no_grad() def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'): + if len(sigmas) <= 1: + return x + sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max() noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type) @torch.no_grad() def sample_dpmpp_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2): + if len(sigmas) <= 1: + return x + sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max() noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler return sample_dpmpp_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=r) @@ -748,7 +765,7 @@ def sample_lcm(model, x, sigmas, extra_args=None, callback=None, disable=None, n x = denoised if sigmas[i + 1] > 0: - x += sigmas[i + 1] * noise_sampler(sigmas[i], sigmas[i + 1]) + x = model.inner_model.inner_model.model_sampling.noise_scaling(sigmas[i + 1], noise_sampler(sigmas[i], sigmas[i + 1]), x) return x diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 2252a075e..4ca466d9a 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -1,3 +1,4 @@ +import torch class LatentFormat: scale_factor = 1.0 @@ -34,6 +35,70 @@ class SDXL(LatentFormat): ] self.taesd_decoder_name = "taesdxl_decoder" +class SDXL_Playground_2_5(LatentFormat): + def __init__(self): + self.scale_factor = 0.5 + self.latents_mean = torch.tensor([-1.6574, 1.886, -1.383, 2.5155]).view(1, 4, 1, 1) + self.latents_std = torch.tensor([8.4927, 5.9022, 6.5498, 5.2299]).view(1, 4, 1, 1) + + self.latent_rgb_factors = [ + # R G B + [ 0.3920, 0.4054, 0.4549], + [-0.2634, -0.0196, 0.0653], + [ 0.0568, 0.1687, -0.0755], + [-0.3112, -0.2359, -0.2076] + ] + self.taesd_decoder_name = "taesdxl_decoder" + + def process_in(self, latent): + latents_mean = self.latents_mean.to(latent.device, latent.dtype) + latents_std = self.latents_std.to(latent.device, latent.dtype) + return (latent - latents_mean) * self.scale_factor / latents_std + + def process_out(self, latent): + latents_mean = self.latents_mean.to(latent.device, latent.dtype) + latents_std = self.latents_std.to(latent.device, latent.dtype) + return latent * latents_std / self.scale_factor + latents_mean + + class SD_X4(LatentFormat): def __init__(self): self.scale_factor = 0.08333 + self.latent_rgb_factors = [ + [-0.2340, -0.3863, -0.3257], + [ 0.0994, 0.0885, -0.0908], + [-0.2833, -0.2349, -0.3741], + [ 0.2523, -0.0055, -0.1651] + ] + +class SC_Prior(LatentFormat): + def __init__(self): + self.scale_factor = 1.0 + self.latent_rgb_factors = [ + [-0.0326, -0.0204, -0.0127], + [-0.1592, -0.0427, 0.0216], + [ 0.0873, 0.0638, -0.0020], + [-0.0602, 0.0442, 0.1304], + [ 0.0800, -0.0313, -0.1796], + [-0.0810, -0.0638, -0.1581], + [ 0.1791, 0.1180, 0.0967], + [ 0.0740, 0.1416, 0.0432], + [-0.1745, -0.1888, -0.1373], + [ 0.2412, 0.1577, 0.0928], + [ 0.1908, 0.0998, 0.0682], + [ 0.0209, 0.0365, -0.0092], + [ 0.0448, -0.0650, -0.1728], + [-0.1658, -0.1045, -0.1308], + [ 0.0542, 0.1545, 0.1325], + [-0.0352, -0.1672, -0.2541] + ] + +class SC_B(LatentFormat): + def __init__(self): + self.scale_factor = 1.0 / 0.43 + self.latent_rgb_factors = [ + [ 0.1121, 0.2006, 0.1023], + [-0.2093, -0.0222, -0.0195], + [-0.3087, -0.1535, 0.0366], + [ 0.0290, -0.1574, -0.4078] + ] diff --git a/comfy/ldm/cascade/common.py b/comfy/ldm/cascade/common.py new file mode 100644 index 000000000..124902c09 --- /dev/null +++ b/comfy/ldm/cascade/common.py @@ -0,0 +1,161 @@ +""" + This file is part of ComfyUI. + Copyright (C) 2024 Stability AI + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +""" + +import torch +import torch.nn as nn +from comfy.ldm.modules.attention import optimized_attention + +class Linear(torch.nn.Linear): + def reset_parameters(self): + return None + +class Conv2d(torch.nn.Conv2d): + def reset_parameters(self): + return None + +class OptimizedAttention(nn.Module): + def __init__(self, c, nhead, dropout=0.0, dtype=None, device=None, operations=None): + super().__init__() + self.heads = nhead + + self.to_q = operations.Linear(c, c, bias=True, dtype=dtype, device=device) + self.to_k = operations.Linear(c, c, bias=True, dtype=dtype, device=device) + self.to_v = operations.Linear(c, c, bias=True, dtype=dtype, device=device) + + self.out_proj = operations.Linear(c, c, bias=True, dtype=dtype, device=device) + + def forward(self, q, k, v): + q = self.to_q(q) + k = self.to_k(k) + v = self.to_v(v) + + out = optimized_attention(q, k, v, self.heads) + + return self.out_proj(out) + +class Attention2D(nn.Module): + def __init__(self, c, nhead, dropout=0.0, dtype=None, device=None, operations=None): + super().__init__() + self.attn = OptimizedAttention(c, nhead, dtype=dtype, device=device, operations=operations) + # self.attn = nn.MultiheadAttention(c, nhead, dropout=dropout, bias=True, batch_first=True, dtype=dtype, device=device) + + def forward(self, x, kv, self_attn=False): + orig_shape = x.shape + x = x.view(x.size(0), x.size(1), -1).permute(0, 2, 1) # Bx4xHxW -> Bx(HxW)x4 + if self_attn: + kv = torch.cat([x, kv], dim=1) + # x = self.attn(x, kv, kv, need_weights=False)[0] + x = self.attn(x, kv, kv) + x = x.permute(0, 2, 1).view(*orig_shape) + return x + + +def LayerNorm2d_op(operations): + class LayerNorm2d(operations.LayerNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, x): + return super().forward(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) + return LayerNorm2d + +class GlobalResponseNorm(nn.Module): + "from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105" + def __init__(self, dim, dtype=None, device=None): + super().__init__() + self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim, dtype=dtype, device=device)) + self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim, dtype=dtype, device=device)) + + def forward(self, x): + Gx = torch.norm(x, p=2, dim=(1, 2), keepdim=True) + Nx = Gx / (Gx.mean(dim=-1, keepdim=True) + 1e-6) + return self.gamma.to(device=x.device, dtype=x.dtype) * (x * Nx) + self.beta.to(device=x.device, dtype=x.dtype) + x + + +class ResBlock(nn.Module): + def __init__(self, c, c_skip=0, kernel_size=3, dropout=0.0, dtype=None, device=None, operations=None): # , num_heads=4, expansion=2): + super().__init__() + self.depthwise = operations.Conv2d(c, c, kernel_size=kernel_size, padding=kernel_size // 2, groups=c, dtype=dtype, device=device) + # self.depthwise = SAMBlock(c, num_heads, expansion) + self.norm = LayerNorm2d_op(operations)(c, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.channelwise = nn.Sequential( + operations.Linear(c + c_skip, c * 4, dtype=dtype, device=device), + nn.GELU(), + GlobalResponseNorm(c * 4, dtype=dtype, device=device), + nn.Dropout(dropout), + operations.Linear(c * 4, c, dtype=dtype, device=device) + ) + + def forward(self, x, x_skip=None): + x_res = x + x = self.norm(self.depthwise(x)) + if x_skip is not None: + x = torch.cat([x, x_skip], dim=1) + x = self.channelwise(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) + return x + x_res + + +class AttnBlock(nn.Module): + def __init__(self, c, c_cond, nhead, self_attn=True, dropout=0.0, dtype=None, device=None, operations=None): + super().__init__() + self.self_attn = self_attn + self.norm = LayerNorm2d_op(operations)(c, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.attention = Attention2D(c, nhead, dropout, dtype=dtype, device=device, operations=operations) + self.kv_mapper = nn.Sequential( + nn.SiLU(), + operations.Linear(c_cond, c, dtype=dtype, device=device) + ) + + def forward(self, x, kv): + kv = self.kv_mapper(kv) + x = x + self.attention(self.norm(x), kv, self_attn=self.self_attn) + return x + + +class FeedForwardBlock(nn.Module): + def __init__(self, c, dropout=0.0, dtype=None, device=None, operations=None): + super().__init__() + self.norm = LayerNorm2d_op(operations)(c, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.channelwise = nn.Sequential( + operations.Linear(c, c * 4, dtype=dtype, device=device), + nn.GELU(), + GlobalResponseNorm(c * 4, dtype=dtype, device=device), + nn.Dropout(dropout), + operations.Linear(c * 4, c, dtype=dtype, device=device) + ) + + def forward(self, x): + x = x + self.channelwise(self.norm(x).permute(0, 2, 3, 1)).permute(0, 3, 1, 2) + return x + + +class TimestepBlock(nn.Module): + def __init__(self, c, c_timestep, conds=['sca'], dtype=None, device=None, operations=None): + super().__init__() + self.mapper = operations.Linear(c_timestep, c * 2, dtype=dtype, device=device) + self.conds = conds + for cname in conds: + setattr(self, f"mapper_{cname}", operations.Linear(c_timestep, c * 2, dtype=dtype, device=device)) + + def forward(self, x, t): + t = t.chunk(len(self.conds) + 1, dim=1) + a, b = self.mapper(t[0])[:, :, None, None].chunk(2, dim=1) + for i, c in enumerate(self.conds): + ac, bc = getattr(self, f"mapper_{c}")(t[i + 1])[:, :, None, None].chunk(2, dim=1) + a, b = a + ac, b + bc + return x * (1 + a) + b diff --git a/comfy/ldm/cascade/controlnet.py b/comfy/ldm/cascade/controlnet.py new file mode 100644 index 000000000..5dac59394 --- /dev/null +++ b/comfy/ldm/cascade/controlnet.py @@ -0,0 +1,93 @@ +""" + This file is part of ComfyUI. + Copyright (C) 2024 Stability AI + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +""" + +import torch +import torchvision +from torch import nn +from .common import LayerNorm2d_op + + +class CNetResBlock(nn.Module): + def __init__(self, c, dtype=None, device=None, operations=None): + super().__init__() + self.blocks = nn.Sequential( + LayerNorm2d_op(operations)(c, dtype=dtype, device=device), + nn.GELU(), + operations.Conv2d(c, c, kernel_size=3, padding=1), + LayerNorm2d_op(operations)(c, dtype=dtype, device=device), + nn.GELU(), + operations.Conv2d(c, c, kernel_size=3, padding=1), + ) + + def forward(self, x): + return x + self.blocks(x) + + +class ControlNet(nn.Module): + def __init__(self, c_in=3, c_proj=2048, proj_blocks=None, bottleneck_mode=None, dtype=None, device=None, operations=nn): + super().__init__() + if bottleneck_mode is None: + bottleneck_mode = 'effnet' + self.proj_blocks = proj_blocks + if bottleneck_mode == 'effnet': + embd_channels = 1280 + self.backbone = torchvision.models.efficientnet_v2_s().features.eval() + if c_in != 3: + in_weights = self.backbone[0][0].weight.data + self.backbone[0][0] = operations.Conv2d(c_in, 24, kernel_size=3, stride=2, bias=False, dtype=dtype, device=device) + if c_in > 3: + # nn.init.constant_(self.backbone[0][0].weight, 0) + self.backbone[0][0].weight.data[:, :3] = in_weights[:, :3].clone() + else: + self.backbone[0][0].weight.data = in_weights[:, :c_in].clone() + elif bottleneck_mode == 'simple': + embd_channels = c_in + self.backbone = nn.Sequential( + operations.Conv2d(embd_channels, embd_channels * 4, kernel_size=3, padding=1, dtype=dtype, device=device), + nn.LeakyReLU(0.2, inplace=True), + operations.Conv2d(embd_channels * 4, embd_channels, kernel_size=3, padding=1, dtype=dtype, device=device), + ) + elif bottleneck_mode == 'large': + self.backbone = nn.Sequential( + operations.Conv2d(c_in, 4096 * 4, kernel_size=1, dtype=dtype, device=device), + nn.LeakyReLU(0.2, inplace=True), + operations.Conv2d(4096 * 4, 1024, kernel_size=1, dtype=dtype, device=device), + *[CNetResBlock(1024, dtype=dtype, device=device, operations=operations) for _ in range(8)], + operations.Conv2d(1024, 1280, kernel_size=1, dtype=dtype, device=device), + ) + embd_channels = 1280 + else: + raise ValueError(f'Unknown bottleneck mode: {bottleneck_mode}') + self.projections = nn.ModuleList() + for _ in range(len(proj_blocks)): + self.projections.append(nn.Sequential( + operations.Conv2d(embd_channels, embd_channels, kernel_size=1, bias=False, dtype=dtype, device=device), + nn.LeakyReLU(0.2, inplace=True), + operations.Conv2d(embd_channels, c_proj, kernel_size=1, bias=False, dtype=dtype, device=device), + )) + # nn.init.constant_(self.projections[-1][-1].weight, 0) # zero output projection + self.xl = False + self.input_channels = c_in + self.unshuffle_amount = 8 + + def forward(self, x): + x = self.backbone(x) + proj_outputs = [None for _ in range(max(self.proj_blocks) + 1)] + for i, idx in enumerate(self.proj_blocks): + proj_outputs[idx] = self.projections[i](x) + return proj_outputs diff --git a/comfy/ldm/cascade/stage_a.py b/comfy/ldm/cascade/stage_a.py new file mode 100644 index 000000000..ca8867eaf --- /dev/null +++ b/comfy/ldm/cascade/stage_a.py @@ -0,0 +1,255 @@ +""" + This file is part of ComfyUI. + Copyright (C) 2024 Stability AI + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +""" + +import torch +from torch import nn +from torch.autograd import Function + +class vector_quantize(Function): + @staticmethod + def forward(ctx, x, codebook): + with torch.no_grad(): + codebook_sqr = torch.sum(codebook ** 2, dim=1) + x_sqr = torch.sum(x ** 2, dim=1, keepdim=True) + + dist = torch.addmm(codebook_sqr + x_sqr, x, codebook.t(), alpha=-2.0, beta=1.0) + _, indices = dist.min(dim=1) + + ctx.save_for_backward(indices, codebook) + ctx.mark_non_differentiable(indices) + + nn = torch.index_select(codebook, 0, indices) + return nn, indices + + @staticmethod + def backward(ctx, grad_output, grad_indices): + grad_inputs, grad_codebook = None, None + + if ctx.needs_input_grad[0]: + grad_inputs = grad_output.clone() + if ctx.needs_input_grad[1]: + # Gradient wrt. the codebook + indices, codebook = ctx.saved_tensors + + grad_codebook = torch.zeros_like(codebook) + grad_codebook.index_add_(0, indices, grad_output) + + return (grad_inputs, grad_codebook) + + +class VectorQuantize(nn.Module): + def __init__(self, embedding_size, k, ema_decay=0.99, ema_loss=False): + """ + Takes an input of variable size (as long as the last dimension matches the embedding size). + Returns one tensor containing the nearest neigbour embeddings to each of the inputs, + with the same size as the input, vq and commitment components for the loss as a touple + in the second output and the indices of the quantized vectors in the third: + quantized, (vq_loss, commit_loss), indices + """ + super(VectorQuantize, self).__init__() + + self.codebook = nn.Embedding(k, embedding_size) + self.codebook.weight.data.uniform_(-1./k, 1./k) + self.vq = vector_quantize.apply + + self.ema_decay = ema_decay + self.ema_loss = ema_loss + if ema_loss: + self.register_buffer('ema_element_count', torch.ones(k)) + self.register_buffer('ema_weight_sum', torch.zeros_like(self.codebook.weight)) + + def _laplace_smoothing(self, x, epsilon): + n = torch.sum(x) + return ((x + epsilon) / (n + x.size(0) * epsilon) * n) + + def _updateEMA(self, z_e_x, indices): + mask = nn.functional.one_hot(indices, self.ema_element_count.size(0)).float() + elem_count = mask.sum(dim=0) + weight_sum = torch.mm(mask.t(), z_e_x) + + self.ema_element_count = (self.ema_decay * self.ema_element_count) + ((1-self.ema_decay) * elem_count) + self.ema_element_count = self._laplace_smoothing(self.ema_element_count, 1e-5) + self.ema_weight_sum = (self.ema_decay * self.ema_weight_sum) + ((1-self.ema_decay) * weight_sum) + + self.codebook.weight.data = self.ema_weight_sum / self.ema_element_count.unsqueeze(-1) + + def idx2vq(self, idx, dim=-1): + q_idx = self.codebook(idx) + if dim != -1: + q_idx = q_idx.movedim(-1, dim) + return q_idx + + def forward(self, x, get_losses=True, dim=-1): + if dim != -1: + x = x.movedim(dim, -1) + z_e_x = x.contiguous().view(-1, x.size(-1)) if len(x.shape) > 2 else x + z_q_x, indices = self.vq(z_e_x, self.codebook.weight.detach()) + vq_loss, commit_loss = None, None + if self.ema_loss and self.training: + self._updateEMA(z_e_x.detach(), indices.detach()) + # pick the graded embeddings after updating the codebook in order to have a more accurate commitment loss + z_q_x_grd = torch.index_select(self.codebook.weight, dim=0, index=indices) + if get_losses: + vq_loss = (z_q_x_grd - z_e_x.detach()).pow(2).mean() + commit_loss = (z_e_x - z_q_x_grd.detach()).pow(2).mean() + + z_q_x = z_q_x.view(x.shape) + if dim != -1: + z_q_x = z_q_x.movedim(-1, dim) + return z_q_x, (vq_loss, commit_loss), indices.view(x.shape[:-1]) + + +class ResBlock(nn.Module): + def __init__(self, c, c_hidden): + super().__init__() + # depthwise/attention + self.norm1 = nn.LayerNorm(c, elementwise_affine=False, eps=1e-6) + self.depthwise = nn.Sequential( + nn.ReplicationPad2d(1), + nn.Conv2d(c, c, kernel_size=3, groups=c) + ) + + # channelwise + self.norm2 = nn.LayerNorm(c, elementwise_affine=False, eps=1e-6) + self.channelwise = nn.Sequential( + nn.Linear(c, c_hidden), + nn.GELU(), + nn.Linear(c_hidden, c), + ) + + self.gammas = nn.Parameter(torch.zeros(6), requires_grad=True) + + # Init weights + def _basic_init(module): + if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d): + torch.nn.init.xavier_uniform_(module.weight) + if module.bias is not None: + nn.init.constant_(module.bias, 0) + + self.apply(_basic_init) + + def _norm(self, x, norm): + return norm(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) + + def forward(self, x): + mods = self.gammas + + x_temp = self._norm(x, self.norm1) * (1 + mods[0]) + mods[1] + try: + x = x + self.depthwise(x_temp) * mods[2] + except: #operation not implemented for bf16 + x_temp = self.depthwise[0](x_temp.float()).to(x.dtype) + x = x + self.depthwise[1](x_temp) * mods[2] + + x_temp = self._norm(x, self.norm2) * (1 + mods[3]) + mods[4] + x = x + self.channelwise(x_temp.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * mods[5] + + return x + + +class StageA(nn.Module): + def __init__(self, levels=2, bottleneck_blocks=12, c_hidden=384, c_latent=4, codebook_size=8192): + super().__init__() + self.c_latent = c_latent + c_levels = [c_hidden // (2 ** i) for i in reversed(range(levels))] + + # Encoder blocks + self.in_block = nn.Sequential( + nn.PixelUnshuffle(2), + nn.Conv2d(3 * 4, c_levels[0], kernel_size=1) + ) + down_blocks = [] + for i in range(levels): + if i > 0: + down_blocks.append(nn.Conv2d(c_levels[i - 1], c_levels[i], kernel_size=4, stride=2, padding=1)) + block = ResBlock(c_levels[i], c_levels[i] * 4) + down_blocks.append(block) + down_blocks.append(nn.Sequential( + nn.Conv2d(c_levels[-1], c_latent, kernel_size=1, bias=False), + nn.BatchNorm2d(c_latent), # then normalize them to have mean 0 and std 1 + )) + self.down_blocks = nn.Sequential(*down_blocks) + self.down_blocks[0] + + self.codebook_size = codebook_size + self.vquantizer = VectorQuantize(c_latent, k=codebook_size) + + # Decoder blocks + up_blocks = [nn.Sequential( + nn.Conv2d(c_latent, c_levels[-1], kernel_size=1) + )] + for i in range(levels): + for j in range(bottleneck_blocks if i == 0 else 1): + block = ResBlock(c_levels[levels - 1 - i], c_levels[levels - 1 - i] * 4) + up_blocks.append(block) + if i < levels - 1: + up_blocks.append( + nn.ConvTranspose2d(c_levels[levels - 1 - i], c_levels[levels - 2 - i], kernel_size=4, stride=2, + padding=1)) + self.up_blocks = nn.Sequential(*up_blocks) + self.out_block = nn.Sequential( + nn.Conv2d(c_levels[0], 3 * 4, kernel_size=1), + nn.PixelShuffle(2), + ) + + def encode(self, x, quantize=False): + x = self.in_block(x) + x = self.down_blocks(x) + if quantize: + qe, (vq_loss, commit_loss), indices = self.vquantizer.forward(x, dim=1) + return qe, x, indices, vq_loss + commit_loss * 0.25 + else: + return x + + def decode(self, x): + x = self.up_blocks(x) + x = self.out_block(x) + return x + + def forward(self, x, quantize=False): + qe, x, _, vq_loss = self.encode(x, quantize) + x = self.decode(qe) + return x, vq_loss + + +class Discriminator(nn.Module): + def __init__(self, c_in=3, c_cond=0, c_hidden=512, depth=6): + super().__init__() + d = max(depth - 3, 3) + layers = [ + nn.utils.spectral_norm(nn.Conv2d(c_in, c_hidden // (2 ** d), kernel_size=3, stride=2, padding=1)), + nn.LeakyReLU(0.2), + ] + for i in range(depth - 1): + c_in = c_hidden // (2 ** max((d - i), 0)) + c_out = c_hidden // (2 ** max((d - 1 - i), 0)) + layers.append(nn.utils.spectral_norm(nn.Conv2d(c_in, c_out, kernel_size=3, stride=2, padding=1))) + layers.append(nn.InstanceNorm2d(c_out)) + layers.append(nn.LeakyReLU(0.2)) + self.encoder = nn.Sequential(*layers) + self.shuffle = nn.Conv2d((c_hidden + c_cond) if c_cond > 0 else c_hidden, 1, kernel_size=1) + self.logits = nn.Sigmoid() + + def forward(self, x, cond=None): + x = self.encoder(x) + if cond is not None: + cond = cond.view(cond.size(0), cond.size(1), 1, 1, ).expand(-1, -1, x.size(-2), x.size(-1)) + x = torch.cat([x, cond], dim=1) + x = self.shuffle(x) + x = self.logits(x) + return x diff --git a/comfy/ldm/cascade/stage_b.py b/comfy/ldm/cascade/stage_b.py new file mode 100644 index 000000000..7c3d8feab --- /dev/null +++ b/comfy/ldm/cascade/stage_b.py @@ -0,0 +1,256 @@ +""" + This file is part of ComfyUI. + Copyright (C) 2024 Stability AI + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +""" + +import math +import torch +from torch import nn +from .common import AttnBlock, LayerNorm2d_op, ResBlock, FeedForwardBlock, TimestepBlock + +class StageB(nn.Module): + def __init__(self, c_in=4, c_out=4, c_r=64, patch_size=2, c_cond=1280, c_hidden=[320, 640, 1280, 1280], + nhead=[-1, -1, 20, 20], blocks=[[2, 6, 28, 6], [6, 28, 6, 2]], + block_repeat=[[1, 1, 1, 1], [3, 3, 2, 2]], level_config=['CT', 'CT', 'CTA', 'CTA'], c_clip=1280, + c_clip_seq=4, c_effnet=16, c_pixels=3, kernel_size=3, dropout=[0, 0, 0.0, 0.0], self_attn=True, + t_conds=['sca'], stable_cascade_stage=None, dtype=None, device=None, operations=None): + super().__init__() + self.dtype = dtype + self.c_r = c_r + self.t_conds = t_conds + self.c_clip_seq = c_clip_seq + if not isinstance(dropout, list): + dropout = [dropout] * len(c_hidden) + if not isinstance(self_attn, list): + self_attn = [self_attn] * len(c_hidden) + + # CONDITIONING + self.effnet_mapper = nn.Sequential( + operations.Conv2d(c_effnet, c_hidden[0] * 4, kernel_size=1, dtype=dtype, device=device), + nn.GELU(), + operations.Conv2d(c_hidden[0] * 4, c_hidden[0], kernel_size=1, dtype=dtype, device=device), + LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + ) + self.pixels_mapper = nn.Sequential( + operations.Conv2d(c_pixels, c_hidden[0] * 4, kernel_size=1, dtype=dtype, device=device), + nn.GELU(), + operations.Conv2d(c_hidden[0] * 4, c_hidden[0], kernel_size=1, dtype=dtype, device=device), + LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + ) + self.clip_mapper = operations.Linear(c_clip, c_cond * c_clip_seq, dtype=dtype, device=device) + self.clip_norm = operations.LayerNorm(c_cond, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + + self.embedding = nn.Sequential( + nn.PixelUnshuffle(patch_size), + operations.Conv2d(c_in * (patch_size ** 2), c_hidden[0], kernel_size=1, dtype=dtype, device=device), + LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + ) + + def get_block(block_type, c_hidden, nhead, c_skip=0, dropout=0, self_attn=True): + if block_type == 'C': + return ResBlock(c_hidden, c_skip, kernel_size=kernel_size, dropout=dropout, dtype=dtype, device=device, operations=operations) + elif block_type == 'A': + return AttnBlock(c_hidden, c_cond, nhead, self_attn=self_attn, dropout=dropout, dtype=dtype, device=device, operations=operations) + elif block_type == 'F': + return FeedForwardBlock(c_hidden, dropout=dropout, dtype=dtype, device=device, operations=operations) + elif block_type == 'T': + return TimestepBlock(c_hidden, c_r, conds=t_conds, dtype=dtype, device=device, operations=operations) + else: + raise Exception(f'Block type {block_type} not supported') + + # BLOCKS + # -- down blocks + self.down_blocks = nn.ModuleList() + self.down_downscalers = nn.ModuleList() + self.down_repeat_mappers = nn.ModuleList() + for i in range(len(c_hidden)): + if i > 0: + self.down_downscalers.append(nn.Sequential( + LayerNorm2d_op(operations)(c_hidden[i - 1], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device), + operations.Conv2d(c_hidden[i - 1], c_hidden[i], kernel_size=2, stride=2, dtype=dtype, device=device), + )) + else: + self.down_downscalers.append(nn.Identity()) + down_block = nn.ModuleList() + for _ in range(blocks[0][i]): + for block_type in level_config[i]: + block = get_block(block_type, c_hidden[i], nhead[i], dropout=dropout[i], self_attn=self_attn[i]) + down_block.append(block) + self.down_blocks.append(down_block) + if block_repeat is not None: + block_repeat_mappers = nn.ModuleList() + for _ in range(block_repeat[0][i] - 1): + block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device)) + self.down_repeat_mappers.append(block_repeat_mappers) + + # -- up blocks + self.up_blocks = nn.ModuleList() + self.up_upscalers = nn.ModuleList() + self.up_repeat_mappers = nn.ModuleList() + for i in reversed(range(len(c_hidden))): + if i > 0: + self.up_upscalers.append(nn.Sequential( + LayerNorm2d_op(operations)(c_hidden[i], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device), + operations.ConvTranspose2d(c_hidden[i], c_hidden[i - 1], kernel_size=2, stride=2, dtype=dtype, device=device), + )) + else: + self.up_upscalers.append(nn.Identity()) + up_block = nn.ModuleList() + for j in range(blocks[1][::-1][i]): + for k, block_type in enumerate(level_config[i]): + c_skip = c_hidden[i] if i < len(c_hidden) - 1 and j == k == 0 else 0 + block = get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i], + self_attn=self_attn[i]) + up_block.append(block) + self.up_blocks.append(up_block) + if block_repeat is not None: + block_repeat_mappers = nn.ModuleList() + for _ in range(block_repeat[1][::-1][i] - 1): + block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device)) + self.up_repeat_mappers.append(block_repeat_mappers) + + # OUTPUT + self.clf = nn.Sequential( + LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device), + operations.Conv2d(c_hidden[0], c_out * (patch_size ** 2), kernel_size=1, dtype=dtype, device=device), + nn.PixelShuffle(patch_size), + ) + + # --- WEIGHT INIT --- + # self.apply(self._init_weights) # General init + # nn.init.normal_(self.clip_mapper.weight, std=0.02) # conditionings + # nn.init.normal_(self.effnet_mapper[0].weight, std=0.02) # conditionings + # nn.init.normal_(self.effnet_mapper[2].weight, std=0.02) # conditionings + # nn.init.normal_(self.pixels_mapper[0].weight, std=0.02) # conditionings + # nn.init.normal_(self.pixels_mapper[2].weight, std=0.02) # conditionings + # torch.nn.init.xavier_uniform_(self.embedding[1].weight, 0.02) # inputs + # nn.init.constant_(self.clf[1].weight, 0) # outputs + # + # # blocks + # for level_block in self.down_blocks + self.up_blocks: + # for block in level_block: + # if isinstance(block, ResBlock) or isinstance(block, FeedForwardBlock): + # block.channelwise[-1].weight.data *= np.sqrt(1 / sum(blocks[0])) + # elif isinstance(block, TimestepBlock): + # for layer in block.modules(): + # if isinstance(layer, nn.Linear): + # nn.init.constant_(layer.weight, 0) + # + # def _init_weights(self, m): + # if isinstance(m, (nn.Conv2d, nn.Linear)): + # torch.nn.init.xavier_uniform_(m.weight) + # if m.bias is not None: + # nn.init.constant_(m.bias, 0) + + def gen_r_embedding(self, r, max_positions=10000): + r = r * max_positions + half_dim = self.c_r // 2 + emb = math.log(max_positions) / (half_dim - 1) + emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp() + emb = r[:, None] * emb[None, :] + emb = torch.cat([emb.sin(), emb.cos()], dim=1) + if self.c_r % 2 == 1: # zero pad + emb = nn.functional.pad(emb, (0, 1), mode='constant') + return emb + + def gen_c_embeddings(self, clip): + if len(clip.shape) == 2: + clip = clip.unsqueeze(1) + clip = self.clip_mapper(clip).view(clip.size(0), clip.size(1) * self.c_clip_seq, -1) + clip = self.clip_norm(clip) + return clip + + def _down_encode(self, x, r_embed, clip): + level_outputs = [] + block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers) + for down_block, downscaler, repmap in block_group: + x = downscaler(x) + for i in range(len(repmap) + 1): + for block in down_block: + if isinstance(block, ResBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + ResBlock)): + x = block(x) + elif isinstance(block, AttnBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + AttnBlock)): + x = block(x, clip) + elif isinstance(block, TimestepBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + TimestepBlock)): + x = block(x, r_embed) + else: + x = block(x) + if i < len(repmap): + x = repmap[i](x) + level_outputs.insert(0, x) + return level_outputs + + def _up_decode(self, level_outputs, r_embed, clip): + x = level_outputs[0] + block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers) + for i, (up_block, upscaler, repmap) in enumerate(block_group): + for j in range(len(repmap) + 1): + for k, block in enumerate(up_block): + if isinstance(block, ResBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + ResBlock)): + skip = level_outputs[i] if k == 0 and i > 0 else None + if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)): + x = torch.nn.functional.interpolate(x, skip.shape[-2:], mode='bilinear', + align_corners=True) + x = block(x, skip) + elif isinstance(block, AttnBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + AttnBlock)): + x = block(x, clip) + elif isinstance(block, TimestepBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + TimestepBlock)): + x = block(x, r_embed) + else: + x = block(x) + if j < len(repmap): + x = repmap[j](x) + x = upscaler(x) + return x + + def forward(self, x, r, effnet, clip, pixels=None, **kwargs): + if pixels is None: + pixels = x.new_zeros(x.size(0), 3, 8, 8) + + # Process the conditioning embeddings + r_embed = self.gen_r_embedding(r).to(dtype=x.dtype) + for c in self.t_conds: + t_cond = kwargs.get(c, torch.zeros_like(r)) + r_embed = torch.cat([r_embed, self.gen_r_embedding(t_cond).to(dtype=x.dtype)], dim=1) + clip = self.gen_c_embeddings(clip) + + # Model Blocks + x = self.embedding(x) + x = x + self.effnet_mapper( + nn.functional.interpolate(effnet, size=x.shape[-2:], mode='bilinear', align_corners=True)) + x = x + nn.functional.interpolate(self.pixels_mapper(pixels), size=x.shape[-2:], mode='bilinear', + align_corners=True) + level_outputs = self._down_encode(x, r_embed, clip) + x = self._up_decode(level_outputs, r_embed, clip) + return self.clf(x) + + def update_weights_ema(self, src_model, beta=0.999): + for self_params, src_params in zip(self.parameters(), src_model.parameters()): + self_params.data = self_params.data * beta + src_params.data.clone().to(self_params.device) * (1 - beta) + for self_buffers, src_buffers in zip(self.buffers(), src_model.buffers()): + self_buffers.data = self_buffers.data * beta + src_buffers.data.clone().to(self_buffers.device) * (1 - beta) diff --git a/comfy/ldm/cascade/stage_c.py b/comfy/ldm/cascade/stage_c.py new file mode 100644 index 000000000..c85da1f01 --- /dev/null +++ b/comfy/ldm/cascade/stage_c.py @@ -0,0 +1,273 @@ +""" + This file is part of ComfyUI. + Copyright (C) 2024 Stability AI + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +""" + +import torch +from torch import nn +import math +from .common import AttnBlock, LayerNorm2d_op, ResBlock, FeedForwardBlock, TimestepBlock +# from .controlnet import ControlNetDeliverer + +class UpDownBlock2d(nn.Module): + def __init__(self, c_in, c_out, mode, enabled=True, dtype=None, device=None, operations=None): + super().__init__() + assert mode in ['up', 'down'] + interpolation = nn.Upsample(scale_factor=2 if mode == 'up' else 0.5, mode='bilinear', + align_corners=True) if enabled else nn.Identity() + mapping = operations.Conv2d(c_in, c_out, kernel_size=1, dtype=dtype, device=device) + self.blocks = nn.ModuleList([interpolation, mapping] if mode == 'up' else [mapping, interpolation]) + + def forward(self, x): + for block in self.blocks: + x = block(x) + return x + + +class StageC(nn.Module): + def __init__(self, c_in=16, c_out=16, c_r=64, patch_size=1, c_cond=2048, c_hidden=[2048, 2048], nhead=[32, 32], + blocks=[[8, 24], [24, 8]], block_repeat=[[1, 1], [1, 1]], level_config=['CTA', 'CTA'], + c_clip_text=1280, c_clip_text_pooled=1280, c_clip_img=768, c_clip_seq=4, kernel_size=3, + dropout=[0.0, 0.0], self_attn=True, t_conds=['sca', 'crp'], switch_level=[False], stable_cascade_stage=None, + dtype=None, device=None, operations=None): + super().__init__() + self.dtype = dtype + self.c_r = c_r + self.t_conds = t_conds + self.c_clip_seq = c_clip_seq + if not isinstance(dropout, list): + dropout = [dropout] * len(c_hidden) + if not isinstance(self_attn, list): + self_attn = [self_attn] * len(c_hidden) + + # CONDITIONING + self.clip_txt_mapper = operations.Linear(c_clip_text, c_cond, dtype=dtype, device=device) + self.clip_txt_pooled_mapper = operations.Linear(c_clip_text_pooled, c_cond * c_clip_seq, dtype=dtype, device=device) + self.clip_img_mapper = operations.Linear(c_clip_img, c_cond * c_clip_seq, dtype=dtype, device=device) + self.clip_norm = operations.LayerNorm(c_cond, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + + self.embedding = nn.Sequential( + nn.PixelUnshuffle(patch_size), + operations.Conv2d(c_in * (patch_size ** 2), c_hidden[0], kernel_size=1, dtype=dtype, device=device), + LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6) + ) + + def get_block(block_type, c_hidden, nhead, c_skip=0, dropout=0, self_attn=True): + if block_type == 'C': + return ResBlock(c_hidden, c_skip, kernel_size=kernel_size, dropout=dropout, dtype=dtype, device=device, operations=operations) + elif block_type == 'A': + return AttnBlock(c_hidden, c_cond, nhead, self_attn=self_attn, dropout=dropout, dtype=dtype, device=device, operations=operations) + elif block_type == 'F': + return FeedForwardBlock(c_hidden, dropout=dropout, dtype=dtype, device=device, operations=operations) + elif block_type == 'T': + return TimestepBlock(c_hidden, c_r, conds=t_conds, dtype=dtype, device=device, operations=operations) + else: + raise Exception(f'Block type {block_type} not supported') + + # BLOCKS + # -- down blocks + self.down_blocks = nn.ModuleList() + self.down_downscalers = nn.ModuleList() + self.down_repeat_mappers = nn.ModuleList() + for i in range(len(c_hidden)): + if i > 0: + self.down_downscalers.append(nn.Sequential( + LayerNorm2d_op(operations)(c_hidden[i - 1], elementwise_affine=False, eps=1e-6), + UpDownBlock2d(c_hidden[i - 1], c_hidden[i], mode='down', enabled=switch_level[i - 1], dtype=dtype, device=device, operations=operations) + )) + else: + self.down_downscalers.append(nn.Identity()) + down_block = nn.ModuleList() + for _ in range(blocks[0][i]): + for block_type in level_config[i]: + block = get_block(block_type, c_hidden[i], nhead[i], dropout=dropout[i], self_attn=self_attn[i]) + down_block.append(block) + self.down_blocks.append(down_block) + if block_repeat is not None: + block_repeat_mappers = nn.ModuleList() + for _ in range(block_repeat[0][i] - 1): + block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device)) + self.down_repeat_mappers.append(block_repeat_mappers) + + # -- up blocks + self.up_blocks = nn.ModuleList() + self.up_upscalers = nn.ModuleList() + self.up_repeat_mappers = nn.ModuleList() + for i in reversed(range(len(c_hidden))): + if i > 0: + self.up_upscalers.append(nn.Sequential( + LayerNorm2d_op(operations)(c_hidden[i], elementwise_affine=False, eps=1e-6), + UpDownBlock2d(c_hidden[i], c_hidden[i - 1], mode='up', enabled=switch_level[i - 1], dtype=dtype, device=device, operations=operations) + )) + else: + self.up_upscalers.append(nn.Identity()) + up_block = nn.ModuleList() + for j in range(blocks[1][::-1][i]): + for k, block_type in enumerate(level_config[i]): + c_skip = c_hidden[i] if i < len(c_hidden) - 1 and j == k == 0 else 0 + block = get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i], + self_attn=self_attn[i]) + up_block.append(block) + self.up_blocks.append(up_block) + if block_repeat is not None: + block_repeat_mappers = nn.ModuleList() + for _ in range(block_repeat[1][::-1][i] - 1): + block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device)) + self.up_repeat_mappers.append(block_repeat_mappers) + + # OUTPUT + self.clf = nn.Sequential( + LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device), + operations.Conv2d(c_hidden[0], c_out * (patch_size ** 2), kernel_size=1, dtype=dtype, device=device), + nn.PixelShuffle(patch_size), + ) + + # --- WEIGHT INIT --- + # self.apply(self._init_weights) # General init + # nn.init.normal_(self.clip_txt_mapper.weight, std=0.02) # conditionings + # nn.init.normal_(self.clip_txt_pooled_mapper.weight, std=0.02) # conditionings + # nn.init.normal_(self.clip_img_mapper.weight, std=0.02) # conditionings + # torch.nn.init.xavier_uniform_(self.embedding[1].weight, 0.02) # inputs + # nn.init.constant_(self.clf[1].weight, 0) # outputs + # + # # blocks + # for level_block in self.down_blocks + self.up_blocks: + # for block in level_block: + # if isinstance(block, ResBlock) or isinstance(block, FeedForwardBlock): + # block.channelwise[-1].weight.data *= np.sqrt(1 / sum(blocks[0])) + # elif isinstance(block, TimestepBlock): + # for layer in block.modules(): + # if isinstance(layer, nn.Linear): + # nn.init.constant_(layer.weight, 0) + # + # def _init_weights(self, m): + # if isinstance(m, (nn.Conv2d, nn.Linear)): + # torch.nn.init.xavier_uniform_(m.weight) + # if m.bias is not None: + # nn.init.constant_(m.bias, 0) + + def gen_r_embedding(self, r, max_positions=10000): + r = r * max_positions + half_dim = self.c_r // 2 + emb = math.log(max_positions) / (half_dim - 1) + emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp() + emb = r[:, None] * emb[None, :] + emb = torch.cat([emb.sin(), emb.cos()], dim=1) + if self.c_r % 2 == 1: # zero pad + emb = nn.functional.pad(emb, (0, 1), mode='constant') + return emb + + def gen_c_embeddings(self, clip_txt, clip_txt_pooled, clip_img): + clip_txt = self.clip_txt_mapper(clip_txt) + if len(clip_txt_pooled.shape) == 2: + clip_txt_pooled = clip_txt_pooled.unsqueeze(1) + if len(clip_img.shape) == 2: + clip_img = clip_img.unsqueeze(1) + clip_txt_pool = self.clip_txt_pooled_mapper(clip_txt_pooled).view(clip_txt_pooled.size(0), clip_txt_pooled.size(1) * self.c_clip_seq, -1) + clip_img = self.clip_img_mapper(clip_img).view(clip_img.size(0), clip_img.size(1) * self.c_clip_seq, -1) + clip = torch.cat([clip_txt, clip_txt_pool, clip_img], dim=1) + clip = self.clip_norm(clip) + return clip + + def _down_encode(self, x, r_embed, clip, cnet=None): + level_outputs = [] + block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers) + for down_block, downscaler, repmap in block_group: + x = downscaler(x) + for i in range(len(repmap) + 1): + for block in down_block: + if isinstance(block, ResBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + ResBlock)): + if cnet is not None: + next_cnet = cnet.pop() + if next_cnet is not None: + x = x + nn.functional.interpolate(next_cnet, size=x.shape[-2:], mode='bilinear', + align_corners=True).to(x.dtype) + x = block(x) + elif isinstance(block, AttnBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + AttnBlock)): + x = block(x, clip) + elif isinstance(block, TimestepBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + TimestepBlock)): + x = block(x, r_embed) + else: + x = block(x) + if i < len(repmap): + x = repmap[i](x) + level_outputs.insert(0, x) + return level_outputs + + def _up_decode(self, level_outputs, r_embed, clip, cnet=None): + x = level_outputs[0] + block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers) + for i, (up_block, upscaler, repmap) in enumerate(block_group): + for j in range(len(repmap) + 1): + for k, block in enumerate(up_block): + if isinstance(block, ResBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + ResBlock)): + skip = level_outputs[i] if k == 0 and i > 0 else None + if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)): + x = torch.nn.functional.interpolate(x, skip.shape[-2:], mode='bilinear', + align_corners=True) + if cnet is not None: + next_cnet = cnet.pop() + if next_cnet is not None: + x = x + nn.functional.interpolate(next_cnet, size=x.shape[-2:], mode='bilinear', + align_corners=True).to(x.dtype) + x = block(x, skip) + elif isinstance(block, AttnBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + AttnBlock)): + x = block(x, clip) + elif isinstance(block, TimestepBlock) or ( + hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module, + TimestepBlock)): + x = block(x, r_embed) + else: + x = block(x) + if j < len(repmap): + x = repmap[j](x) + x = upscaler(x) + return x + + def forward(self, x, r, clip_text, clip_text_pooled, clip_img, control=None, **kwargs): + # Process the conditioning embeddings + r_embed = self.gen_r_embedding(r).to(dtype=x.dtype) + for c in self.t_conds: + t_cond = kwargs.get(c, torch.zeros_like(r)) + r_embed = torch.cat([r_embed, self.gen_r_embedding(t_cond).to(dtype=x.dtype)], dim=1) + clip = self.gen_c_embeddings(clip_text, clip_text_pooled, clip_img) + + if control is not None: + cnet = control.get("input") + else: + cnet = None + + # Model Blocks + x = self.embedding(x) + level_outputs = self._down_encode(x, r_embed, clip, cnet) + x = self._up_decode(level_outputs, r_embed, clip, cnet) + return self.clf(x) + + def update_weights_ema(self, src_model, beta=0.999): + for self_params, src_params in zip(self.parameters(), src_model.parameters()): + self_params.data = self_params.data * beta + src_params.data.clone().to(self_params.device) * (1 - beta) + for self_buffers, src_buffers in zip(self.buffers(), src_model.buffers()): + self_buffers.data = self_buffers.data * beta + src_buffers.data.clone().to(self_buffers.device) * (1 - beta) diff --git a/comfy/ldm/cascade/stage_c_coder.py b/comfy/ldm/cascade/stage_c_coder.py new file mode 100644 index 000000000..0cb7c49fc --- /dev/null +++ b/comfy/ldm/cascade/stage_c_coder.py @@ -0,0 +1,95 @@ +""" + This file is part of ComfyUI. + Copyright (C) 2024 Stability AI + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +""" +import torch +import torchvision +from torch import nn + + +# EfficientNet +class EfficientNetEncoder(nn.Module): + def __init__(self, c_latent=16): + super().__init__() + self.backbone = torchvision.models.efficientnet_v2_s().features.eval() + self.mapper = nn.Sequential( + nn.Conv2d(1280, c_latent, kernel_size=1, bias=False), + nn.BatchNorm2d(c_latent, affine=False), # then normalize them to have mean 0 and std 1 + ) + self.mean = nn.Parameter(torch.tensor([0.485, 0.456, 0.406])) + self.std = nn.Parameter(torch.tensor([0.229, 0.224, 0.225])) + + def forward(self, x): + x = x * 0.5 + 0.5 + x = (x - self.mean.view([3,1,1])) / self.std.view([3,1,1]) + o = self.mapper(self.backbone(x)) + return o + + +# Fast Decoder for Stage C latents. E.g. 16 x 24 x 24 -> 3 x 192 x 192 +class Previewer(nn.Module): + def __init__(self, c_in=16, c_hidden=512, c_out=3): + super().__init__() + self.blocks = nn.Sequential( + nn.Conv2d(c_in, c_hidden, kernel_size=1), # 16 channels to 512 channels + nn.GELU(), + nn.BatchNorm2d(c_hidden), + + nn.Conv2d(c_hidden, c_hidden, kernel_size=3, padding=1), + nn.GELU(), + nn.BatchNorm2d(c_hidden), + + nn.ConvTranspose2d(c_hidden, c_hidden // 2, kernel_size=2, stride=2), # 16 -> 32 + nn.GELU(), + nn.BatchNorm2d(c_hidden // 2), + + nn.Conv2d(c_hidden // 2, c_hidden // 2, kernel_size=3, padding=1), + nn.GELU(), + nn.BatchNorm2d(c_hidden // 2), + + nn.ConvTranspose2d(c_hidden // 2, c_hidden // 4, kernel_size=2, stride=2), # 32 -> 64 + nn.GELU(), + nn.BatchNorm2d(c_hidden // 4), + + nn.Conv2d(c_hidden // 4, c_hidden // 4, kernel_size=3, padding=1), + nn.GELU(), + nn.BatchNorm2d(c_hidden // 4), + + nn.ConvTranspose2d(c_hidden // 4, c_hidden // 4, kernel_size=2, stride=2), # 64 -> 128 + nn.GELU(), + nn.BatchNorm2d(c_hidden // 4), + + nn.Conv2d(c_hidden // 4, c_hidden // 4, kernel_size=3, padding=1), + nn.GELU(), + nn.BatchNorm2d(c_hidden // 4), + + nn.Conv2d(c_hidden // 4, c_out, kernel_size=1), + ) + + def forward(self, x): + return (self.blocks(x) - 0.5) * 2.0 + +class StageC_coder(nn.Module): + def __init__(self): + super().__init__() + self.previewer = Previewer() + self.encoder = EfficientNetEncoder() + + def encode(self, x): + return self.encoder(x) + + def decode(self, x): + return self.previewer(x) diff --git a/comfy/ldm/models/autoencoder.py b/comfy/ldm/models/autoencoder.py index b91ec3249..f5f4de288 100644 --- a/comfy/ldm/models/autoencoder.py +++ b/comfy/ldm/models/autoencoder.py @@ -1,6 +1,4 @@ import torch -# import pytorch_lightning as pl -import torch.nn.functional as F from contextlib import contextmanager from typing import Any, Dict, List, Optional, Tuple, Union diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index fd8888d0e..da9f7aab7 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -1,14 +1,12 @@ -from inspect import isfunction import math import torch import torch.nn.functional as F from torch import nn, einsum from einops import rearrange, repeat -from typing import Optional, Any -from functools import partial +from typing import Optional +import logging - -from .diffusionmodules.util import checkpoint, AlphaBlender, timestep_embedding +from .diffusionmodules.util import AlphaBlender, timestep_embedding from .sub_quadratic_attention import efficient_dot_product_attention from comfy import model_management @@ -21,13 +19,14 @@ from comfy.cli_args import args import comfy.ops ops = comfy.ops.disable_weight_init -# CrossAttn precision handling -if args.dont_upcast_attention: - print("disabling upcasting of attention") - _ATTN_PRECISION = "fp16" -else: - _ATTN_PRECISION = "fp32" +FORCE_UPCAST_ATTENTION_DTYPE = model_management.force_upcast_attention_dtype() +def get_attn_precision(attn_precision): + if args.dont_upcast_attention: + return None + if FORCE_UPCAST_ATTENTION_DTYPE is not None: + return FORCE_UPCAST_ATTENTION_DTYPE + return attn_precision def exists(val): return val is not None @@ -87,7 +86,9 @@ class FeedForward(nn.Module): def Normalize(in_channels, dtype=None, device=None): return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device) -def attention_basic(q, k, v, heads, mask=None): +def attention_basic(q, k, v, heads, mask=None, attn_precision=None): + attn_precision = get_attn_precision(attn_precision) + b, _, dim_head = q.shape dim_head //= heads scale = dim_head ** -0.5 @@ -103,7 +104,7 @@ def attention_basic(q, k, v, heads, mask=None): ) # force cast to fp32 to avoid overflowing - if _ATTN_PRECISION =="fp32": + if attn_precision == torch.float32: sim = einsum('b i d, b j d -> b i j', q.float(), k.float()) * scale else: sim = einsum('b i d, b j d -> b i j', q, k) * scale @@ -117,7 +118,12 @@ def attention_basic(q, k, v, heads, mask=None): mask = repeat(mask, 'b j -> (b h) () j', h=h) sim.masked_fill_(~mask, max_neg_value) else: - sim += mask + if len(mask.shape) == 2: + bs = 1 + else: + bs = mask.shape[0] + mask = mask.reshape(bs, -1, mask.shape[-2], mask.shape[-1]).expand(b, heads, -1, -1).reshape(-1, mask.shape[-2], mask.shape[-1]) + sim.add_(mask) # attention, what we cannot get enough of sim = sim.softmax(dim=-1) @@ -132,7 +138,9 @@ def attention_basic(q, k, v, heads, mask=None): return out -def attention_sub_quad(query, key, value, heads, mask=None): +def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None): + attn_precision = get_attn_precision(attn_precision) + b, _, dim_head = query.shape dim_head //= heads @@ -143,7 +151,7 @@ def attention_sub_quad(query, key, value, heads, mask=None): key = key.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 3, 1).reshape(b * heads, dim_head, -1) dtype = query.dtype - upcast_attention = _ATTN_PRECISION =="fp32" and query.dtype != torch.float32 + upcast_attention = attn_precision == torch.float32 and query.dtype != torch.float32 if upcast_attention: bytes_per_token = torch.finfo(torch.float32).bits//8 else: @@ -168,6 +176,13 @@ def attention_sub_quad(query, key, value, heads, mask=None): if query_chunk_size is None: query_chunk_size = 512 + if mask is not None: + if len(mask.shape) == 2: + bs = 1 + else: + bs = mask.shape[0] + mask = mask.reshape(bs, -1, mask.shape[-2], mask.shape[-1]).expand(b, heads, -1, -1).reshape(-1, mask.shape[-2], mask.shape[-1]) + hidden_states = efficient_dot_product_attention( query, key, @@ -185,7 +200,9 @@ def attention_sub_quad(query, key, value, heads, mask=None): hidden_states = hidden_states.unflatten(0, (-1, heads)).transpose(1,2).flatten(start_dim=2) return hidden_states -def attention_split(q, k, v, heads, mask=None): +def attention_split(q, k, v, heads, mask=None, attn_precision=None): + attn_precision = get_attn_precision(attn_precision) + b, _, dim_head = q.shape dim_head //= heads scale = dim_head ** -0.5 @@ -204,10 +221,12 @@ def attention_split(q, k, v, heads, mask=None): mem_free_total = model_management.get_free_memory(q.device) - if _ATTN_PRECISION =="fp32": + if attn_precision == torch.float32: element_size = 4 + upcast = True else: element_size = q.element_size() + upcast = False gb = 1024 ** 3 tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * element_size @@ -226,6 +245,13 @@ def attention_split(q, k, v, heads, mask=None): raise RuntimeError(f'Not enough memory, use lower resolution (max approx. {max_res}x{max_res}). ' f'Need: {mem_required/64/gb:0.1f}GB free, Have:{mem_free_total/gb:0.1f}GB free') + if mask is not None: + if len(mask.shape) == 2: + bs = 1 + else: + bs = mask.shape[0] + mask = mask.reshape(bs, -1, mask.shape[-2], mask.shape[-1]).expand(b, heads, -1, -1).reshape(-1, mask.shape[-2], mask.shape[-1]) + # print("steps", steps, mem_required, mem_free_total, modifier, q.element_size(), tensor_size) first_op_done = False cleared_cache = False @@ -234,7 +260,7 @@ def attention_split(q, k, v, heads, mask=None): slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1] for i in range(0, q.shape[1], slice_size): end = i + slice_size - if _ATTN_PRECISION =="fp32": + if upcast: with torch.autocast(enabled=False, device_type = 'cuda'): s1 = einsum('b i d, b j d -> b i j', q[:, i:end].float(), k.float()) * scale else: @@ -258,12 +284,12 @@ def attention_split(q, k, v, heads, mask=None): model_management.soft_empty_cache(True) if cleared_cache == False: cleared_cache = True - print("out of memory error, emptying cache and trying again") + logging.warning("out of memory error, emptying cache and trying again") continue steps *= 2 if steps > 64: raise e - print("out of memory error, increasing steps and trying again", steps) + logging.warning("out of memory error, increasing steps and trying again {}".format(steps)) else: raise e @@ -280,24 +306,30 @@ def attention_split(q, k, v, heads, mask=None): BROKEN_XFORMERS = False try: x_vers = xformers.__version__ - #I think 0.0.23 is also broken (q with bs bigger than 65535 gives CUDA error) - BROKEN_XFORMERS = x_vers.startswith("0.0.21") or x_vers.startswith("0.0.22") or x_vers.startswith("0.0.23") + # XFormers bug confirmed on all versions from 0.0.21 to 0.0.26 (q with bs bigger than 65535 gives CUDA error) + BROKEN_XFORMERS = x_vers.startswith("0.0.2") and not x_vers.startswith("0.0.20") except: pass -def attention_xformers(q, k, v, heads, mask=None): +def attention_xformers(q, k, v, heads, mask=None, attn_precision=None): b, _, dim_head = q.shape dim_head //= heads + + disabled_xformers = False + if BROKEN_XFORMERS: if b * heads > 65535: - return attention_pytorch(q, k, v, heads, mask) + disabled_xformers = True + + if not disabled_xformers: + if torch.jit.is_tracing() or torch.jit.is_scripting(): + disabled_xformers = True + + if disabled_xformers: + return attention_pytorch(q, k, v, heads, mask) q, k, v = map( - lambda t: t.unsqueeze(3) - .reshape(b, -1, heads, dim_head) - .permute(0, 2, 1, 3) - .reshape(b * heads, -1, dim_head) - .contiguous(), + lambda t: t.reshape(b, -1, heads, dim_head), (q, k, v), ) @@ -310,14 +342,11 @@ def attention_xformers(q, k, v, heads, mask=None): out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=mask) out = ( - out.unsqueeze(0) - .reshape(b, heads, -1, dim_head) - .permute(0, 2, 1, 3) - .reshape(b, -1, heads * dim_head) + out.reshape(b, -1, heads * dim_head) ) return out -def attention_pytorch(q, k, v, heads, mask=None): +def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None): b, _, dim_head = q.shape dim_head //= heads q, k, v = map( @@ -335,17 +364,17 @@ def attention_pytorch(q, k, v, heads, mask=None): optimized_attention = attention_basic if model_management.xformers_enabled(): - print("Using xformers cross attention") + logging.info("Using xformers cross attention") optimized_attention = attention_xformers elif model_management.pytorch_attention_enabled(): - print("Using pytorch cross attention") + logging.info("Using pytorch cross attention") optimized_attention = attention_pytorch else: if args.use_split_cross_attention: - print("Using split optimization for cross attention") + logging.info("Using split optimization for cross attention") optimized_attention = attention_split else: - print("Using sub quadratic optimization for cross attention, if you have memory or speed issues try using: --use-split-cross-attention") + logging.info("Using sub quadratic optimization for cross attention, if you have memory or speed issues try using: --use-split-cross-attention") optimized_attention = attention_sub_quad optimized_attention_masked = optimized_attention @@ -367,10 +396,11 @@ def optimized_attention_for_device(device, mask=False, small_input=False): class CrossAttention(nn.Module): - def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., dtype=None, device=None, operations=ops): + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=ops): super().__init__() inner_dim = dim_head * heads context_dim = default(context_dim, query_dim) + self.attn_precision = attn_precision self.heads = heads self.dim_head = dim_head @@ -392,15 +422,15 @@ class CrossAttention(nn.Module): v = self.to_v(context) if mask is None: - out = optimized_attention(q, k, v, self.heads) + out = optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision) else: - out = optimized_attention_masked(q, k, v, self.heads, mask) + out = optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision) return self.to_out(out) class BasicTransformerBlock(nn.Module): def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, ff_in=False, inner_dim=None, - disable_self_attn=False, disable_temporal_crossattention=False, switch_temporal_ca_to_sa=False, dtype=None, device=None, operations=ops): + disable_self_attn=False, disable_temporal_crossattention=False, switch_temporal_ca_to_sa=False, attn_precision=None, dtype=None, device=None, operations=ops): super().__init__() self.ff_in = ff_in or inner_dim is not None @@ -408,6 +438,7 @@ class BasicTransformerBlock(nn.Module): inner_dim = dim self.is_res = inner_dim == dim + self.attn_precision = attn_precision if self.ff_in: self.norm_in = operations.LayerNorm(dim, dtype=dtype, device=device) @@ -415,7 +446,7 @@ class BasicTransformerBlock(nn.Module): self.disable_self_attn = disable_self_attn self.attn1 = CrossAttention(query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout, - context_dim=context_dim if self.disable_self_attn else None, dtype=dtype, device=device, operations=operations) # is a self-attention if not self.disable_self_attn + context_dim=context_dim if self.disable_self_attn else None, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations) # is a self-attention if not self.disable_self_attn self.ff = FeedForward(inner_dim, dim_out=dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations) if disable_temporal_crossattention: @@ -429,20 +460,16 @@ class BasicTransformerBlock(nn.Module): context_dim_attn2 = context_dim self.attn2 = CrossAttention(query_dim=inner_dim, context_dim=context_dim_attn2, - heads=n_heads, dim_head=d_head, dropout=dropout, dtype=dtype, device=device, operations=operations) # is self-attn if context is none + heads=n_heads, dim_head=d_head, dropout=dropout, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations) # is self-attn if context is none self.norm2 = operations.LayerNorm(inner_dim, dtype=dtype, device=device) self.norm1 = operations.LayerNorm(inner_dim, dtype=dtype, device=device) self.norm3 = operations.LayerNorm(inner_dim, dtype=dtype, device=device) - self.checkpoint = checkpoint self.n_heads = n_heads self.d_head = d_head self.switch_temporal_ca_to_sa = switch_temporal_ca_to_sa def forward(self, x, context=None, transformer_options={}): - return checkpoint(self._forward, (x, context, transformer_options), self.parameters(), self.checkpoint) - - def _forward(self, x, context=None, transformer_options={}): extra_options = {} block = transformer_options.get("block", None) block_index = transformer_options.get("block_index", 0) @@ -459,6 +486,7 @@ class BasicTransformerBlock(nn.Module): extra_options["n_heads"] = self.n_heads extra_options["dim_head"] = self.d_head + extra_options["attn_precision"] = self.attn_precision if self.ff_in: x_skip = x @@ -569,7 +597,7 @@ class SpatialTransformer(nn.Module): def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None, disable_self_attn=False, use_linear=False, - use_checkpoint=True, dtype=None, device=None, operations=ops): + use_checkpoint=True, attn_precision=None, dtype=None, device=None, operations=ops): super().__init__() if exists(context_dim) and not isinstance(context_dim, list): context_dim = [context_dim] * depth @@ -587,7 +615,7 @@ class SpatialTransformer(nn.Module): self.transformer_blocks = nn.ModuleList( [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d], - disable_self_attn=disable_self_attn, checkpoint=use_checkpoint, dtype=dtype, device=device, operations=operations) + disable_self_attn=disable_self_attn, checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=dtype, device=device, operations=operations) for d in range(depth)] ) if not use_linear: @@ -608,7 +636,7 @@ class SpatialTransformer(nn.Module): x = self.norm(x) if not self.use_linear: x = self.proj_in(x) - x = rearrange(x, 'b c h w -> b (h w) c').contiguous() + x = x.movedim(1, 3).flatten(1, 2).contiguous() if self.use_linear: x = self.proj_in(x) for i, block in enumerate(self.transformer_blocks): @@ -616,7 +644,7 @@ class SpatialTransformer(nn.Module): x = block(x, context=context[i], transformer_options=transformer_options) if self.use_linear: x = self.proj_out(x) - x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous() + x = x.reshape(x.shape[0], h, w, x.shape[-1]).movedim(3, 1).contiguous() if not self.use_linear: x = self.proj_out(x) return x + x_in @@ -643,6 +671,7 @@ class SpatialVideoTransformer(SpatialTransformer): disable_self_attn=False, disable_temporal_crossattention=False, max_time_embed_period: int = 10000, + attn_precision=None, dtype=None, device=None, operations=ops ): super().__init__( @@ -655,6 +684,7 @@ class SpatialVideoTransformer(SpatialTransformer): context_dim=context_dim, use_linear=use_linear, disable_self_attn=disable_self_attn, + attn_precision=attn_precision, dtype=dtype, device=device, operations=operations ) self.time_depth = time_depth @@ -684,6 +714,7 @@ class SpatialVideoTransformer(SpatialTransformer): inner_dim=time_mix_inner_dim, disable_self_attn=disable_self_attn, disable_temporal_crossattention=disable_temporal_crossattention, + attn_precision=attn_precision, dtype=dtype, device=device, operations=operations ) for _ in range(self.depth) diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.py index cc81c1f23..04eb83b21 100644 --- a/comfy/ldm/modules/diffusionmodules/model.py +++ b/comfy/ldm/modules/diffusionmodules/model.py @@ -3,8 +3,8 @@ import math import torch import torch.nn as nn import numpy as np -from einops import rearrange from typing import Optional, Any +import logging from comfy import model_management import comfy.ops @@ -190,7 +190,7 @@ def slice_attention(q, k, v): steps *= 2 if steps > 128: raise e - print("out of memory error, increasing steps and trying again", steps) + logging.warning("out of memory error, increasing steps and trying again {}".format(steps)) return r1 @@ -235,7 +235,7 @@ def pytorch_attention(q, k, v): out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False) out = out.transpose(2, 3).reshape(B, C, H, W) except model_management.OOM_EXCEPTION as e: - print("scaled_dot_product_attention OOMed: switched to slice attention") + logging.warning("scaled_dot_product_attention OOMed: switched to slice attention") out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(B, C, H, W) return out @@ -268,13 +268,13 @@ class AttnBlock(nn.Module): padding=0) if model_management.xformers_enabled_vae(): - print("Using xformers attention in VAE") + logging.info("Using xformers attention in VAE") self.optimized_attention = xformers_attention elif model_management.pytorch_attention_enabled(): - print("Using pytorch attention in VAE") + logging.info("Using pytorch attention in VAE") self.optimized_attention = pytorch_attention else: - print("Using split attention in VAE") + logging.info("Using split attention in VAE") self.optimized_attention = normal_attention def forward(self, x): @@ -562,7 +562,7 @@ class Decoder(nn.Module): block_in = ch*ch_mult[self.num_resolutions-1] curr_res = resolution // 2**(self.num_resolutions-1) self.z_shape = (1,z_channels,curr_res,curr_res) - print("Working with z of shape {} = {} dimensions.".format( + logging.debug("Working with z of shape {} = {} dimensions.".format( self.z_shape, np.prod(self.z_shape))) # z to block_in diff --git a/comfy/ldm/modules/diffusionmodules/openaimodel.py b/comfy/ldm/modules/diffusionmodules/openaimodel.py index ea936e066..ba8fc2c4a 100644 --- a/comfy/ldm/modules/diffusionmodules/openaimodel.py +++ b/comfy/ldm/modules/diffusionmodules/openaimodel.py @@ -1,12 +1,10 @@ from abc import abstractmethod -import math -import numpy as np import torch as th import torch.nn as nn import torch.nn.functional as F from einops import rearrange -from functools import partial +import logging from .util import ( checkpoint, @@ -260,7 +258,7 @@ class ResBlock(TimestepBlock): else: if emb_out is not None: if self.exchange_temb_dims: - emb_out = rearrange(emb_out, "b t c ... -> b c t ...") + emb_out = emb_out.movedim(1, 2) h = h + emb_out h = self.out_layers(h) return self.skip_connection(x) + h @@ -362,7 +360,7 @@ def apply_control(h, control, name): try: h += ctrl except: - print("warning control could not be applied", h.shape, ctrl.shape) + logging.warning("warning control could not be applied {} {}".format(h.shape, ctrl.shape)) return h class UNetModel(nn.Module): @@ -433,6 +431,7 @@ class UNetModel(nn.Module): video_kernel_size=None, disable_temporal_crossattention=False, max_ddpm_temb_period=10000, + attn_precision=None, device=None, operations=ops, ): @@ -487,7 +486,6 @@ class UNetModel(nn.Module): self.predict_codebook_ids = n_embed is not None self.default_num_video_frames = None - self.default_image_only_indicator = None time_embed_dim = model_channels * 4 self.time_embed = nn.Sequential( @@ -500,7 +498,7 @@ class UNetModel(nn.Module): if isinstance(self.num_classes, int): self.label_emb = nn.Embedding(num_classes, time_embed_dim, dtype=self.dtype, device=device) elif self.num_classes == "continuous": - print("setting up linear c_adm embedding layer") + logging.debug("setting up linear c_adm embedding layer") self.label_emb = nn.Linear(1, time_embed_dim) elif self.num_classes == "sequential": assert adm_in_channels is not None @@ -553,13 +551,14 @@ class UNetModel(nn.Module): disable_self_attn=disable_self_attn, disable_temporal_crossattention=disable_temporal_crossattention, max_time_embed_period=max_ddpm_temb_period, + attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations ) else: return SpatialTransformer( ch, num_heads, dim_head, depth=depth, context_dim=context_dim, disable_self_attn=disable_self_attn, use_linear=use_linear_in_transformer, - use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations + use_checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations ) def get_resblock( @@ -711,27 +710,30 @@ class UNetModel(nn.Module): device=device, operations=operations )] - if transformer_depth_middle >= 0: - mid_block += [get_attention_layer( # always uses a self-attn - ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim, - disable_self_attn=disable_middle_self_attn, use_checkpoint=use_checkpoint - ), - get_resblock( - merge_factor=merge_factor, - merge_strategy=merge_strategy, - video_kernel_size=video_kernel_size, - ch=ch, - time_embed_dim=time_embed_dim, - dropout=dropout, - out_channels=None, - dims=dims, - use_checkpoint=use_checkpoint, - use_scale_shift_norm=use_scale_shift_norm, - dtype=self.dtype, - device=device, - operations=operations - )] - self.middle_block = TimestepEmbedSequential(*mid_block) + + self.middle_block = None + if transformer_depth_middle >= -1: + if transformer_depth_middle >= 0: + mid_block += [get_attention_layer( # always uses a self-attn + ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim, + disable_self_attn=disable_middle_self_attn, use_checkpoint=use_checkpoint + ), + get_resblock( + merge_factor=merge_factor, + merge_strategy=merge_strategy, + video_kernel_size=video_kernel_size, + ch=ch, + time_embed_dim=time_embed_dim, + dropout=dropout, + out_channels=None, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + dtype=self.dtype, + device=device, + operations=operations + )] + self.middle_block = TimestepEmbedSequential(*mid_block) self._feature_size += ch self.output_blocks = nn.ModuleList([]) @@ -830,7 +832,7 @@ class UNetModel(nn.Module): transformer_patches = transformer_options.get("patches", {}) num_video_frames = kwargs.get("num_video_frames", self.default_num_video_frames) - image_only_indicator = kwargs.get("image_only_indicator", self.default_image_only_indicator) + image_only_indicator = kwargs.get("image_only_indicator", None) time_context = kwargs.get("time_context", None) assert (y is not None) == ( @@ -861,7 +863,8 @@ class UNetModel(nn.Module): h = p(h, transformer_options) transformer_options["block"] = ("middle", 0) - h = forward_timestep_embed(self.middle_block, h, emb, context, transformer_options, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator) + if self.middle_block is not None: + h = forward_timestep_embed(self.middle_block, h, emb, context, transformer_options, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator) h = apply_control(h, control, 'middle') diff --git a/comfy/ldm/modules/diffusionmodules/util.py b/comfy/ldm/modules/diffusionmodules/util.py index ac7e27173..ce14ad5e1 100644 --- a/comfy/ldm/modules/diffusionmodules/util.py +++ b/comfy/ldm/modules/diffusionmodules/util.py @@ -46,23 +46,25 @@ class AlphaBlender(nn.Module): else: raise ValueError(f"unknown merge strategy {self.merge_strategy}") - def get_alpha(self, image_only_indicator: torch.Tensor) -> torch.Tensor: + def get_alpha(self, image_only_indicator: torch.Tensor, device) -> torch.Tensor: # skip_time_mix = rearrange(repeat(skip_time_mix, 'b -> (b t) () () ()', t=t), '(b t) 1 ... -> b 1 t ...', t=t) if self.merge_strategy == "fixed": # make shape compatible # alpha = repeat(self.mix_factor, '1 -> b () t () ()', t=t, b=bs) - alpha = self.mix_factor.to(image_only_indicator.device) + alpha = self.mix_factor.to(device) elif self.merge_strategy == "learned": - alpha = torch.sigmoid(self.mix_factor.to(image_only_indicator.device)) + alpha = torch.sigmoid(self.mix_factor.to(device)) # make shape compatible # alpha = repeat(alpha, '1 -> s () ()', s = t * bs) elif self.merge_strategy == "learned_with_images": - assert image_only_indicator is not None, "need image_only_indicator ..." - alpha = torch.where( - image_only_indicator.bool(), - torch.ones(1, 1, device=image_only_indicator.device), - rearrange(torch.sigmoid(self.mix_factor.to(image_only_indicator.device)), "... -> ... 1"), - ) + if image_only_indicator is None: + alpha = rearrange(torch.sigmoid(self.mix_factor.to(device)), "... -> ... 1") + else: + alpha = torch.where( + image_only_indicator.bool(), + torch.ones(1, 1, device=image_only_indicator.device), + rearrange(torch.sigmoid(self.mix_factor.to(image_only_indicator.device)), "... -> ... 1"), + ) alpha = rearrange(alpha, self.rearrange_pattern) # make shape compatible # alpha = repeat(alpha, '1 -> s () ()', s = t * bs) @@ -76,7 +78,7 @@ class AlphaBlender(nn.Module): x_temporal, image_only_indicator=None, ) -> torch.Tensor: - alpha = self.get_alpha(image_only_indicator) + alpha = self.get_alpha(image_only_indicator, x_spatial.device) x = ( alpha.to(x_spatial.dtype) * x_spatial + (1.0 - alpha).to(x_spatial.dtype) * x_temporal @@ -98,7 +100,7 @@ def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, alphas = torch.cos(alphas).pow(2) alphas = alphas / alphas[0] betas = 1 - alphas[1:] / alphas[:-1] - betas = np.clip(betas, a_min=0, a_max=0.999) + betas = torch.clamp(betas, min=0, max=0.999) elif schedule == "squaredcos_cap_v2": # used for karlo prior # return early @@ -113,7 +115,7 @@ def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5 else: raise ValueError(f"schedule '{schedule}' unknown.") - return betas.numpy() + return betas def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True): diff --git a/comfy/ldm/modules/encoders/noise_aug_modules.py b/comfy/ldm/modules/encoders/noise_aug_modules.py index 66767b587..a5d866030 100644 --- a/comfy/ldm/modules/encoders/noise_aug_modules.py +++ b/comfy/ldm/modules/encoders/noise_aug_modules.py @@ -23,13 +23,13 @@ class CLIPEmbeddingNoiseAugmentation(ImageConcatWithNoiseAugmentation): x = (x * self.data_std.to(x.device)) + self.data_mean.to(x.device) return x - def forward(self, x, noise_level=None): + def forward(self, x, noise_level=None, seed=None): if noise_level is None: noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long() else: assert isinstance(noise_level, torch.Tensor) x = self.scale(x) - z = self.q_sample(x, noise_level) + z = self.q_sample(x, noise_level, seed=seed) z = self.unscale(z) noise_level = self.time_embed(noise_level) return z, noise_level diff --git a/comfy/ldm/modules/sub_quadratic_attention.py b/comfy/ldm/modules/sub_quadratic_attention.py index cb0896b0d..1bc4138c3 100644 --- a/comfy/ldm/modules/sub_quadratic_attention.py +++ b/comfy/ldm/modules/sub_quadratic_attention.py @@ -14,6 +14,7 @@ import torch from torch import Tensor from torch.utils.checkpoint import checkpoint import math +import logging try: from typing import Optional, NamedTuple, List, Protocol @@ -170,7 +171,7 @@ def _get_attention_scores_no_kv_chunking( attn_probs = attn_scores.softmax(dim=-1) del attn_scores except model_management.OOM_EXCEPTION: - print("ran out of memory while running softmax in _get_attention_scores_no_kv_chunking, trying slower in place softmax instead") + logging.warning("ran out of memory while running softmax in _get_attention_scores_no_kv_chunking, trying slower in place softmax instead") attn_scores -= attn_scores.max(dim=-1, keepdim=True).values torch.exp(attn_scores, out=attn_scores) summed = torch.sum(attn_scores, dim=-1, keepdim=True) diff --git a/comfy/lora.py b/comfy/lora.py index 5e4009b47..096285bba 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -1,4 +1,5 @@ import comfy.utils +import logging LORA_CLIP_MAP = { "mlp.fc1": "mlp_fc1", @@ -20,6 +21,12 @@ def load_lora(lora, to_load): alpha = lora[alpha_name].item() loaded_keys.add(alpha_name) + dora_scale_name = "{}.dora_scale".format(x) + dora_scale = None + if dora_scale_name in lora.keys(): + dora_scale = lora[dora_scale_name] + loaded_keys.add(dora_scale_name) + regular_lora = "{}.lora_up.weight".format(x) diffusers_lora = "{}_lora.up.weight".format(x) transformers_lora = "{}.lora_linear_layer.up.weight".format(x) @@ -43,7 +50,7 @@ def load_lora(lora, to_load): if mid_name is not None and mid_name in lora.keys(): mid = lora[mid_name] loaded_keys.add(mid_name) - patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid)) + patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid, dora_scale)) loaded_keys.add(A_name) loaded_keys.add(B_name) @@ -64,7 +71,7 @@ def load_lora(lora, to_load): loaded_keys.add(hada_t1_name) loaded_keys.add(hada_t2_name) - patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2)) + patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2, dora_scale)) loaded_keys.add(hada_w1_a_name) loaded_keys.add(hada_w1_b_name) loaded_keys.add(hada_w2_a_name) @@ -116,7 +123,7 @@ def load_lora(lora, to_load): loaded_keys.add(lokr_t2_name) if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None): - patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2)) + patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale)) #glora a1_name = "{}.a1.weight".format(x) @@ -124,7 +131,7 @@ def load_lora(lora, to_load): b1_name = "{}.b1.weight".format(x) b2_name = "{}.b2.weight".format(x) if a1_name in lora: - patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha)) + patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha, dora_scale)) loaded_keys.add(a1_name) loaded_keys.add(a2_name) loaded_keys.add(b1_name) @@ -156,7 +163,7 @@ def load_lora(lora, to_load): for x in lora.keys(): if x not in loaded_keys: - print("lora key not loaded", x) + logging.warning("lora key not loaded: {}".format(x)) return patch_dict def model_lora_keys_clip(model, key_map={}): @@ -197,6 +204,15 @@ def model_lora_keys_clip(model, key_map={}): key_map[lora_key] = k lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora key_map[lora_key] = k + lora_key = "lora_prior_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #cascade lora: TODO put lora key prefix in the model config + key_map[lora_key] = k + + + k = "clip_g.transformer.text_projection.weight" + if k in sdk: + key_map["lora_prior_te_text_projection"] = k #cascade lora? + # key_map["text_encoder.text_projection"] = k #TODO: check if other lora have the text_projection too + # key_map["lora_te_text_projection"] = k return key_map @@ -207,6 +223,7 @@ def model_lora_keys_unet(model, key_map={}): if k.startswith("diffusion_model.") and k.endswith(".weight"): key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_") key_map["lora_unet_{}".format(key_lora)] = k + key_map["lora_prior_unet_{}".format(key_lora)] = k #cascade lora: TODO put lora key prefix in the model config diffusers_keys = comfy.utils.unet_to_diffusers(model.model_config.unet_config) for k in diffusers_keys: diff --git a/comfy/model_base.py b/comfy/model_base.py index 52c87ede6..841598b73 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1,21 +1,25 @@ import torch +import logging from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep +from comfy.ldm.cascade.stage_c import StageC +from comfy.ldm.cascade.stage_b import StageB from comfy.ldm.modules.encoders.noise_aug_modules import CLIPEmbeddingNoiseAugmentation from comfy.ldm.modules.diffusionmodules.upscaling import ImageConcatWithNoiseAugmentation import comfy.model_management import comfy.conds import comfy.ops from enum import Enum -import contextlib from . import utils class ModelType(Enum): EPS = 1 V_PREDICTION = 2 V_PREDICTION_EDM = 3 + STABLE_CASCADE = 4 + EDM = 5 -from comfy.model_sampling import EPS, V_PREDICTION, ModelSamplingDiscrete, ModelSamplingContinuousEDM +from comfy.model_sampling import EPS, V_PREDICTION, EDM, ModelSamplingDiscrete, ModelSamplingContinuousEDM, StableCascadeSampling def model_sampling(model_config, model_type): @@ -28,6 +32,12 @@ def model_sampling(model_config, model_type): elif model_type == ModelType.V_PREDICTION_EDM: c = V_PREDICTION s = ModelSamplingContinuousEDM + elif model_type == ModelType.STABLE_CASCADE: + c = EPS + s = StableCascadeSampling + elif model_type == ModelType.EDM: + c = EDM + s = ModelSamplingContinuousEDM class ModelSampling(s, c): pass @@ -36,7 +46,7 @@ def model_sampling(model_config, model_type): class BaseModel(torch.nn.Module): - def __init__(self, model_config, model_type=ModelType.EPS, device=None): + def __init__(self, model_config, model_type=ModelType.EPS, device=None, unet_model=UNetModel): super().__init__() unet_config = model_config.unet_config @@ -49,16 +59,17 @@ class BaseModel(torch.nn.Module): operations = comfy.ops.manual_cast else: operations = comfy.ops.disable_weight_init - self.diffusion_model = UNetModel(**unet_config, device=device, operations=operations) + self.diffusion_model = unet_model(**unet_config, device=device, operations=operations) self.model_type = model_type self.model_sampling = model_sampling(model_config, model_type) self.adm_channels = unet_config.get("adm_in_channels", None) if self.adm_channels is None: self.adm_channels = 0 - self.inpaint_model = False - print("model_type", model_type.name) - print("adm", self.adm_channels) + + self.concat_keys = () + logging.info("model_type {}".format(model_type.name)) + logging.debug("adm {}".format(self.adm_channels)) def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs): sigma = t @@ -97,8 +108,7 @@ class BaseModel(torch.nn.Module): def extra_conds(self, **kwargs): out = {} - if self.inpaint_model: - concat_keys = ("mask", "masked_image") + if len(self.concat_keys) > 0: cond_concat = [] denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) concat_latent_image = kwargs.get("concat_latent_image", None) @@ -115,24 +125,16 @@ class BaseModel(torch.nn.Module): concat_latent_image = utils.resize_to_batch_size(concat_latent_image, noise.shape[0]) - if len(denoise_mask.shape) == len(noise.shape): - denoise_mask = denoise_mask[:,:1] + if denoise_mask is not None: + if len(denoise_mask.shape) == len(noise.shape): + denoise_mask = denoise_mask[:,:1] - denoise_mask = denoise_mask.reshape((-1, 1, denoise_mask.shape[-2], denoise_mask.shape[-1])) - if denoise_mask.shape[-2:] != noise.shape[-2:]: - denoise_mask = utils.common_upscale(denoise_mask, noise.shape[-1], noise.shape[-2], "bilinear", "center") - denoise_mask = utils.resize_to_batch_size(denoise_mask.round(), noise.shape[0]) + denoise_mask = denoise_mask.reshape((-1, 1, denoise_mask.shape[-2], denoise_mask.shape[-1])) + if denoise_mask.shape[-2:] != noise.shape[-2:]: + denoise_mask = utils.common_upscale(denoise_mask, noise.shape[-1], noise.shape[-2], "bilinear", "center") + denoise_mask = utils.resize_to_batch_size(denoise_mask.round(), noise.shape[0]) - def blank_inpaint_image_like(latent_image): - blank_image = torch.ones_like(latent_image) - # these are the values for "zero" in pixel space translated to latent space - blank_image[:,0] *= 0.8223 - blank_image[:,1] *= -0.6876 - blank_image[:,2] *= 0.6364 - blank_image[:,3] *= 0.1380 - return blank_image - - for ck in concat_keys: + for ck in self.concat_keys: if denoise_mask is not None: if ck == "mask": cond_concat.append(denoise_mask.to(device)) @@ -142,7 +144,7 @@ class BaseModel(torch.nn.Module): if ck == "mask": cond_concat.append(torch.ones_like(noise)[:,:1]) elif ck == "masked_image": - cond_concat.append(blank_inpaint_image_like(noise)) + cond_concat.append(self.blank_inpaint_image_like(noise)) data = torch.cat(cond_concat, dim=1) out['c_concat'] = comfy.conds.CONDNoiseShape(data) @@ -154,6 +156,14 @@ class BaseModel(torch.nn.Module): if cross_attn is not None: out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn) + cross_attn_cnet = kwargs.get("cross_attn_controlnet", None) + if cross_attn_cnet is not None: + out['crossattn_controlnet'] = comfy.conds.CONDCrossAttn(cross_attn_cnet) + + c_concat = kwargs.get("noise_concat", None) + if c_concat is not None: + out['c_concat'] = comfy.conds.CONDNoiseShape(c_concat) + return out def load_model_weights(self, sd, unet_prefix=""): @@ -166,10 +176,10 @@ class BaseModel(torch.nn.Module): to_load = self.model_config.process_unet_state_dict(to_load) m, u = self.diffusion_model.load_state_dict(to_load, strict=False) if len(m) > 0: - print("unet missing:", m) + logging.warning("unet missing: {}".format(m)) if len(u) > 0: - print("unet unexpected:", u) + logging.warning("unet unexpected: {}".format(u)) del to_load return self @@ -179,22 +189,40 @@ class BaseModel(torch.nn.Module): def process_latent_out(self, latent): return self.latent_format.process_out(latent) - def state_dict_for_saving(self, clip_state_dict, vae_state_dict): - clip_state_dict = self.model_config.process_clip_state_dict_for_saving(clip_state_dict) + def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None): + extra_sds = [] + if clip_state_dict is not None: + extra_sds.append(self.model_config.process_clip_state_dict_for_saving(clip_state_dict)) + if vae_state_dict is not None: + extra_sds.append(self.model_config.process_vae_state_dict_for_saving(vae_state_dict)) + if clip_vision_state_dict is not None: + extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict)) + unet_state_dict = self.diffusion_model.state_dict() unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict) - vae_state_dict = self.model_config.process_vae_state_dict_for_saving(vae_state_dict) + if self.get_dtype() == torch.float16: - clip_state_dict = utils.convert_sd_to(clip_state_dict, torch.float16) - vae_state_dict = utils.convert_sd_to(vae_state_dict, torch.float16) + extra_sds = map(lambda sd: utils.convert_sd_to(sd, torch.float16), extra_sds) if self.model_type == ModelType.V_PREDICTION: unet_state_dict["v_pred"] = torch.tensor([]) - return {**unet_state_dict, **vae_state_dict, **clip_state_dict} + for sd in extra_sds: + unet_state_dict.update(sd) + + return unet_state_dict def set_inpaint(self): - self.inpaint_model = True + self.concat_keys = ("mask", "masked_image") + def blank_inpaint_image_like(latent_image): + blank_image = torch.ones_like(latent_image) + # these are the values for "zero" in pixel space translated to latent space + blank_image[:,0] *= 0.8223 + blank_image[:,1] *= -0.6876 + blank_image[:,2] *= 0.6364 + blank_image[:,3] *= 0.1380 + return blank_image + self.blank_inpaint_image_like = blank_inpaint_image_like def memory_required(self, input_shape): if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention(): @@ -210,7 +238,7 @@ class BaseModel(torch.nn.Module): return (((area * 0.6) / 0.9) + 1024) * (1024 * 1024) -def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0): +def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0, seed=None): adm_inputs = [] weights = [] noise_aug = [] @@ -219,7 +247,7 @@ def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge weight = unclip_cond["strength"] noise_augment = unclip_cond["noise_augmentation"] noise_level = round((noise_augmentor.max_noise_level - 1) * noise_augment) - c_adm, noise_level_emb = noise_augmentor(adm_cond.to(device), noise_level=torch.tensor([noise_level], device=device)) + c_adm, noise_level_emb = noise_augmentor(adm_cond.to(device), noise_level=torch.tensor([noise_level], device=device), seed=seed) adm_out = torch.cat((c_adm, noise_level_emb), 1) * weight weights.append(weight) noise_aug.append(noise_augment) @@ -245,11 +273,11 @@ class SD21UNCLIP(BaseModel): if unclip_conditioning is None: return torch.zeros((1, self.adm_channels)) else: - return unclip_adm(unclip_conditioning, device, self.noise_augmentor, kwargs.get("unclip_noise_augment_merge", 0.05)) + return unclip_adm(unclip_conditioning, device, self.noise_augmentor, kwargs.get("unclip_noise_augment_merge", 0.05), kwargs.get("seed", 0) - 10) def sdxl_pooled(args, noise_augmentor): if "unclip_conditioning" in args: - return unclip_adm(args.get("unclip_conditioning", None), args["device"], noise_augmentor)[:,:1280] + return unclip_adm(args.get("unclip_conditioning", None), args["device"], noise_augmentor, seed=args.get("seed", 0) - 10)[:,:1280] else: return args["pooled_output"] @@ -350,10 +378,39 @@ class SVD_img2vid(BaseModel): if "time_conditioning" in kwargs: out["time_context"] = comfy.conds.CONDCrossAttn(kwargs["time_conditioning"]) - out['image_only_indicator'] = comfy.conds.CONDConstant(torch.zeros((1,), device=device)) out['num_video_frames'] = comfy.conds.CONDConstant(noise.shape[0]) return out +class SV3D_u(SVD_img2vid): + def encode_adm(self, **kwargs): + augmentation = kwargs.get("augmentation_level", 0) + + out = [] + out.append(self.embedder(torch.flatten(torch.Tensor([augmentation])))) + + flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0) + return flat + +class SV3D_p(SVD_img2vid): + def __init__(self, model_config, model_type=ModelType.V_PREDICTION_EDM, device=None): + super().__init__(model_config, model_type, device=device) + self.embedder_512 = Timestep(512) + + def encode_adm(self, **kwargs): + augmentation = kwargs.get("augmentation_level", 0) + elevation = kwargs.get("elevation", 0) #elevation and azimuth are in degrees here + azimuth = kwargs.get("azimuth", 0) + noise = kwargs.get("noise", None) + + out = [] + out.append(self.embedder(torch.flatten(torch.Tensor([augmentation])))) + out.append(self.embedder_512(torch.deg2rad(torch.fmod(torch.flatten(90 - torch.Tensor([elevation])), 360.0)))) + out.append(self.embedder_512(torch.deg2rad(torch.fmod(torch.flatten(torch.Tensor([azimuth])), 360.0)))) + + out = list(map(lambda a: utils.resize_to_batch_size(a, noise.shape[0]), out)) + return torch.cat(out, dim=1) + + class Stable_Zero123(BaseModel): def __init__(self, model_config, model_type=ModelType.EPS, device=None, cc_projection_weight=None, cc_projection_bias=None): super().__init__(model_config, model_type, device=device) @@ -415,3 +472,88 @@ class SD_X4Upscaler(BaseModel): out['c_concat'] = comfy.conds.CONDNoiseShape(image) out['y'] = comfy.conds.CONDRegular(noise_level) return out + +class IP2P: + def extra_conds(self, **kwargs): + out = {} + + image = kwargs.get("concat_latent_image", None) + noise = kwargs.get("noise", None) + device = kwargs["device"] + + if image is None: + image = torch.zeros_like(noise) + + if image.shape[1:] != noise.shape[1:]: + image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + + image = utils.resize_to_batch_size(image, noise.shape[0]) + + out['c_concat'] = comfy.conds.CONDNoiseShape(self.process_ip2p_image_in(image)) + adm = self.encode_adm(**kwargs) + if adm is not None: + out['y'] = comfy.conds.CONDRegular(adm) + return out + +class SD15_instructpix2pix(IP2P, BaseModel): + def __init__(self, model_config, model_type=ModelType.EPS, device=None): + super().__init__(model_config, model_type, device=device) + self.process_ip2p_image_in = lambda image: image + +class SDXL_instructpix2pix(IP2P, SDXL): + def __init__(self, model_config, model_type=ModelType.EPS, device=None): + super().__init__(model_config, model_type, device=device) + if model_type == ModelType.V_PREDICTION_EDM: + self.process_ip2p_image_in = lambda image: comfy.latent_formats.SDXL().process_in(image) #cosxl ip2p + else: + self.process_ip2p_image_in = lambda image: image #diffusers ip2p + + +class StableCascade_C(BaseModel): + def __init__(self, model_config, model_type=ModelType.STABLE_CASCADE, device=None): + super().__init__(model_config, model_type, device=device, unet_model=StageC) + self.diffusion_model.eval().requires_grad_(False) + + def extra_conds(self, **kwargs): + out = {} + clip_text_pooled = kwargs["pooled_output"] + if clip_text_pooled is not None: + out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled) + + if "unclip_conditioning" in kwargs: + embeds = [] + for unclip_cond in kwargs["unclip_conditioning"]: + weight = unclip_cond["strength"] + embeds.append(unclip_cond["clip_vision_output"].image_embeds.unsqueeze(0) * weight) + clip_img = torch.cat(embeds, dim=1) + else: + clip_img = torch.zeros((1, 1, 768)) + out["clip_img"] = comfy.conds.CONDRegular(clip_img) + out["sca"] = comfy.conds.CONDRegular(torch.zeros((1,))) + out["crp"] = comfy.conds.CONDRegular(torch.zeros((1,))) + + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['clip_text'] = comfy.conds.CONDCrossAttn(cross_attn) + return out + + +class StableCascade_B(BaseModel): + def __init__(self, model_config, model_type=ModelType.STABLE_CASCADE, device=None): + super().__init__(model_config, model_type, device=device, unet_model=StageB) + self.diffusion_model.eval().requires_grad_(False) + + def extra_conds(self, **kwargs): + out = {} + noise = kwargs.get("noise", None) + + clip_text_pooled = kwargs["pooled_output"] + if clip_text_pooled is not None: + out['clip'] = comfy.conds.CONDRegular(clip_text_pooled) + + #size of prior doesn't really matter if zeros because it gets resized but I still want it to get batched + prior = kwargs.get("stable_cascade_prior", torch.zeros((1, 16, (noise.shape[2] * 4) // 42, (noise.shape[3] * 4) // 42), dtype=noise.dtype, layout=noise.layout, device=noise.device)) + + out["effnet"] = comfy.conds.CONDRegular(prior) + out["sca"] = comfy.conds.CONDRegular(torch.zeros((1,))) + return out diff --git a/comfy/model_detection.py b/comfy/model_detection.py index ea824c44c..23358a2c0 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -1,5 +1,6 @@ import comfy.supported_models import comfy.supported_models_base +import logging def count_blocks(state_dict_keys, prefix_string): count = 0 @@ -28,9 +29,38 @@ def calculate_transformer_depth(prefix, state_dict_keys, state_dict): return last_transformer_depth, context_dim, use_linear_in_transformer, time_stack return None -def detect_unet_config(state_dict, key_prefix, dtype): +def detect_unet_config(state_dict, key_prefix): state_dict_keys = list(state_dict.keys()) + if '{}clf.1.weight'.format(key_prefix) in state_dict_keys: #stable cascade + unet_config = {} + text_mapper_name = '{}clip_txt_mapper.weight'.format(key_prefix) + if text_mapper_name in state_dict_keys: + unet_config['stable_cascade_stage'] = 'c' + w = state_dict[text_mapper_name] + if w.shape[0] == 1536: #stage c lite + unet_config['c_cond'] = 1536 + unet_config['c_hidden'] = [1536, 1536] + unet_config['nhead'] = [24, 24] + unet_config['blocks'] = [[4, 12], [12, 4]] + elif w.shape[0] == 2048: #stage c full + unet_config['c_cond'] = 2048 + elif '{}clip_mapper.weight'.format(key_prefix) in state_dict_keys: + unet_config['stable_cascade_stage'] = 'b' + w = state_dict['{}down_blocks.1.0.channelwise.0.weight'.format(key_prefix)] + if w.shape[-1] == 640: + unet_config['c_hidden'] = [320, 640, 1280, 1280] + unet_config['nhead'] = [-1, -1, 20, 20] + unet_config['blocks'] = [[2, 6, 28, 6], [6, 28, 6, 2]] + unet_config['block_repeat'] = [[1, 1, 1, 1], [3, 3, 2, 2]] + elif w.shape[-1] == 576: #stage b lite + unet_config['c_hidden'] = [320, 576, 1152, 1152] + unet_config['nhead'] = [-1, 9, 18, 18] + unet_config['blocks'] = [[2, 4, 14, 4], [4, 14, 4, 2]] + unet_config['block_repeat'] = [[1, 1, 1, 1], [2, 2, 2, 2]] + + return unet_config + unet_config = { "use_checkpoint": False, "image_size": 32, @@ -45,7 +75,6 @@ def detect_unet_config(state_dict, key_prefix, dtype): else: unet_config["adm_in_channels"] = None - unet_config["dtype"] = dtype model_channels = state_dict['{}input_blocks.0.0.weight'.format(key_prefix)].shape[0] in_channels = state_dict['{}input_blocks.0.0.weight'.format(key_prefix)].shape[1] @@ -123,8 +152,10 @@ def detect_unet_config(state_dict, key_prefix, dtype): channel_mult.append(last_channel_mult) if "{}middle_block.1.proj_in.weight".format(key_prefix) in state_dict_keys: transformer_depth_middle = count_blocks(state_dict_keys, '{}middle_block.1.transformer_blocks.'.format(key_prefix) + '{}') - else: + elif "{}middle_block.0.in_layers.0.weight".format(key_prefix) in state_dict_keys: transformer_depth_middle = -1 + else: + transformer_depth_middle = -2 unet_config["in_channels"] = in_channels unet_config["out_channels"] = out_channels @@ -151,17 +182,17 @@ def detect_unet_config(state_dict, key_prefix, dtype): return unet_config -def model_config_from_unet_config(unet_config): +def model_config_from_unet_config(unet_config, state_dict=None): for model_config in comfy.supported_models.models: - if model_config.matches(unet_config): + if model_config.matches(unet_config, state_dict): return model_config(unet_config) - print("no match", unet_config) + logging.error("no match {}".format(unet_config)) return None -def model_config_from_unet(state_dict, unet_key_prefix, dtype, use_base_if_no_match=False): - unet_config = detect_unet_config(state_dict, unet_key_prefix, dtype) - model_config = model_config_from_unet_config(unet_config) +def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=False): + unet_config = detect_unet_config(state_dict, unet_key_prefix) + model_config = model_config_from_unet_config(unet_config, state_dict) if model_config is None and use_base_if_no_match: return comfy.supported_models_base.BASE(unet_config) else: @@ -206,7 +237,7 @@ def convert_config(unet_config): return new_config -def unet_config_from_diffusers_unet(state_dict, dtype): +def unet_config_from_diffusers_unet(state_dict, dtype=None): match = {} transformer_depth = [] @@ -214,6 +245,7 @@ def unet_config_from_diffusers_unet(state_dict, dtype): down_blocks = count_blocks(state_dict, "down_blocks.{}") for i in range(down_blocks): attn_blocks = count_blocks(state_dict, "down_blocks.{}.attentions.".format(i) + '{}') + res_blocks = count_blocks(state_dict, "down_blocks.{}.resnets.".format(i) + '{}') for ab in range(attn_blocks): transformer_count = count_blocks(state_dict, "down_blocks.{}.attentions.{}.transformer_blocks.".format(i, ab) + '{}') transformer_depth.append(transformer_count) @@ -222,8 +254,8 @@ def unet_config_from_diffusers_unet(state_dict, dtype): attn_res *= 2 if attn_blocks == 0: - transformer_depth.append(0) - transformer_depth.append(0) + for i in range(res_blocks): + transformer_depth.append(0) match["transformer_depth"] = transformer_depth @@ -289,6 +321,12 @@ def unet_config_from_diffusers_unet(state_dict, dtype): 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10], 'use_temporal_attention': False, 'use_temporal_resblock': False} + SDXL_diffusers_ip2p = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, + 'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 8, 'model_channels': 320, + 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10, + 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10], + 'use_temporal_attention': False, 'use_temporal_resblock': False} + SSD_1B = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 4, 4], 'transformer_depth_output': [0, 0, 0, 1, 1, 2, 10, 4, 4], @@ -301,7 +339,32 @@ def unet_config_from_diffusers_unet(state_dict, dtype): 'channel_mult': [1, 2, 4], 'transformer_depth_middle': -1, 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'use_temporal_attention': False, 'use_temporal_resblock': False} - supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega] + KOALA_700M = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, + 'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, + 'num_res_blocks': [1, 1, 1], 'transformer_depth': [0, 2, 5], 'transformer_depth_output': [0, 0, 2, 2, 5, 5], + 'channel_mult': [1, 2, 4], 'transformer_depth_middle': -2, 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, + 'use_temporal_attention': False, 'use_temporal_resblock': False} + + KOALA_1B = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, + 'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, + 'num_res_blocks': [1, 1, 1], 'transformer_depth': [0, 2, 6], 'transformer_depth_output': [0, 0, 2, 2, 6, 6], + 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 6, 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, + 'use_temporal_attention': False, 'use_temporal_resblock': False} + + SD09_XS = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, + 'adm_in_channels': None, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [1, 1, 1], + 'transformer_depth': [1, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': -2, 'use_linear_in_transformer': True, + 'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1], + 'use_temporal_attention': False, 'use_temporal_resblock': False, 'disable_self_attentions': [True, False, False]} + + SD_XS = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, + 'adm_in_channels': None, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [1, 1, 1], + 'transformer_depth': [0, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': -2, 'use_linear_in_transformer': False, + 'context_dim': 768, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 1, 1, 1, 1], + 'use_temporal_attention': False, 'use_temporal_resblock': False} + + + supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p] for unet_config in supported_models: matches = True @@ -313,8 +376,8 @@ def unet_config_from_diffusers_unet(state_dict, dtype): return convert_config(unet_config) return None -def model_config_from_diffusers_unet(state_dict, dtype): - unet_config = unet_config_from_diffusers_unet(state_dict, dtype) +def model_config_from_diffusers_unet(state_dict): + unet_config = unet_config_from_diffusers_unet(state_dict) if unet_config is not None: return model_config_from_unet_config(unet_config) return None diff --git a/comfy/model_management.py b/comfy/model_management.py index fefd3c8c9..3b9fad362 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1,9 +1,10 @@ import psutil +import logging from enum import Enum from comfy.cli_args import args -import comfy.utils import torch import sys +import platform class VRAMState(Enum): DISABLED = 0 #No vram present: no need to move models to vram @@ -29,7 +30,7 @@ lowvram_available = True xpu_available = False if args.deterministic: - print("Using deterministic algorithms for pytorch") + logging.info("Using deterministic algorithms for pytorch") torch.use_deterministic_algorithms(True, warn_only=True) directml_enabled = False @@ -41,7 +42,7 @@ if args.directml is not None: directml_device = torch_directml.device() else: directml_device = torch_directml.device(device_index) - print("Using directml with device:", torch_directml.device_name(device_index)) + logging.info("Using directml with device: {}".format(torch_directml.device_name(device_index))) # torch_directml.disable_tiled_resources(True) lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default. @@ -82,7 +83,7 @@ def get_torch_device(): return torch.device("cpu") else: if is_intel_xpu(): - return torch.device("xpu") + return torch.device("xpu", torch.xpu.current_device()) else: return torch.device(torch.cuda.current_device()) @@ -101,8 +102,8 @@ def get_total_memory(dev=None, torch_total_too=False): elif is_intel_xpu(): stats = torch.xpu.memory_stats(dev) mem_reserved = stats['reserved_bytes.all.current'] - mem_total = torch.xpu.get_device_properties(dev).total_memory mem_total_torch = mem_reserved + mem_total = torch.xpu.get_device_properties(dev).total_memory else: stats = torch.cuda.memory_stats(dev) mem_reserved = stats['reserved_bytes.all.current'] @@ -117,11 +118,12 @@ def get_total_memory(dev=None, torch_total_too=False): total_vram = get_total_memory(get_torch_device()) / (1024 * 1024) total_ram = psutil.virtual_memory().total / (1024 * 1024) -print("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram)) -if not args.normalvram and not args.cpu: - if lowvram_available and total_vram <= 4096: - print("Trying to enable lowvram mode because your GPU seems to have 4GB or less. If you don't want this use: --normalvram") - set_vram_to = VRAMState.LOW_VRAM +logging.info("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram)) + +try: + logging.info("pytorch version: {}".format(torch.version.__version__)) +except: + pass try: OOM_EXCEPTION = torch.cuda.OutOfMemoryError @@ -143,12 +145,10 @@ else: pass try: XFORMERS_VERSION = xformers.version.__version__ - print("xformers version:", XFORMERS_VERSION) + logging.info("xformers version: {}".format(XFORMERS_VERSION)) if XFORMERS_VERSION.startswith("0.0.18"): - print() - print("WARNING: This version of xformers has a major bug where you will get black images when generating high resolution images.") - print("Please downgrade or upgrade xformers to a different version.") - print() + logging.warning("\nWARNING: This version of xformers has a major bug where you will get black images when generating high resolution images.") + logging.warning("Please downgrade or upgrade xformers to a different version.\n") XFORMERS_ENABLED_VAE = False except: pass @@ -175,7 +175,7 @@ try: if int(torch_version[0]) >= 2: if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False: ENABLE_PYTORCH_ATTENTION = True - if torch.cuda.is_bf16_supported(): + if torch.cuda.is_bf16_supported() and torch.cuda.get_device_properties(torch.cuda.current_device()).major >= 8: VAE_DTYPE = torch.bfloat16 if is_intel_xpu(): if args.use_split_cross_attention == False and args.use_quad_cross_attention == False: @@ -213,11 +213,11 @@ elif args.highvram or args.gpu_only: FORCE_FP32 = False FORCE_FP16 = False if args.force_fp32: - print("Forcing FP32, if this improves things please report it.") + logging.info("Forcing FP32, if this improves things please report it.") FORCE_FP32 = True if args.force_fp16: - print("Forcing FP16.") + logging.info("Forcing FP16.") FORCE_FP16 = True if lowvram_available: @@ -231,12 +231,12 @@ if cpu_state != CPUState.GPU: if cpu_state == CPUState.MPS: vram_state = VRAMState.SHARED -print(f"Set vram state to: {vram_state.name}") +logging.info(f"Set vram state to: {vram_state.name}") DISABLE_SMART_MEMORY = args.disable_smart_memory if DISABLE_SMART_MEMORY: - print("Disabling smart memory management") + logging.info("Disabling smart memory management") def get_torch_device_name(device): if hasattr(device, 'type'): @@ -254,11 +254,11 @@ def get_torch_device_name(device): return "CUDA {}: {}".format(device, torch.cuda.get_device_name(device)) try: - print("Device:", get_torch_device_name(get_torch_device())) + logging.info("Device: {}".format(get_torch_device_name(get_torch_device()))) except: - print("Could not pick default device.") + logging.warning("Could not pick default device.") -print("VAE dtype:", VAE_DTYPE) +logging.info("VAE dtype: {}".format(VAE_DTYPE)) current_loaded_models = [] @@ -273,8 +273,9 @@ def module_size(module): class LoadedModel: def __init__(self, model): self.model = model - self.model_accelerated = False self.device = model.load_device + self.weights_loaded = False + self.real_model = None def model_memory(self): return self.model.model_size() @@ -285,55 +286,40 @@ class LoadedModel: else: return self.model_memory() - def model_load(self, lowvram_model_memory=0): - patch_model_to = None - if lowvram_model_memory == 0: - patch_model_to = self.device + def model_load(self, lowvram_model_memory=0, force_patch_weights=False): + patch_model_to = self.device self.model.model_patches_to(self.device) self.model.model_patches_to(self.model.model_dtype()) + load_weights = not self.weights_loaded + try: - self.real_model = self.model.patch_model(device_to=patch_model_to) #TODO: do something with loras and offloading to CPU + if lowvram_model_memory > 0 and load_weights: + self.real_model = self.model.patch_model_lowvram(device_to=patch_model_to, lowvram_model_memory=lowvram_model_memory, force_patch_weights=force_patch_weights) + else: + self.real_model = self.model.patch_model(device_to=patch_model_to, patch_weights=load_weights) except Exception as e: self.model.unpatch_model(self.model.offload_device) self.model_unload() raise e - if lowvram_model_memory > 0: - print("loading in lowvram mode", lowvram_model_memory/(1024 * 1024)) - mem_counter = 0 - for m in self.real_model.modules(): - if hasattr(m, "comfy_cast_weights"): - m.prev_comfy_cast_weights = m.comfy_cast_weights - m.comfy_cast_weights = True - module_mem = module_size(m) - if mem_counter + module_mem < lowvram_model_memory: - m.to(self.device) - mem_counter += module_mem - elif hasattr(m, "weight"): #only modules with comfy_cast_weights can be set to lowvram mode - m.to(self.device) - mem_counter += module_size(m) - print("lowvram: loaded module regularly", m) - - self.model_accelerated = True - if is_intel_xpu() and not args.disable_ipex_optimize: - self.real_model = torch.xpu.optimize(self.real_model.eval(), inplace=True, auto_kernel_selection=True, graph_mode=True) + self.real_model = ipex.optimize(self.real_model.eval(), graph_mode=True, concat_linear=True) + self.weights_loaded = True return self.real_model - def model_unload(self): - if self.model_accelerated: - for m in self.real_model.modules(): - if hasattr(m, "prev_comfy_cast_weights"): - m.comfy_cast_weights = m.prev_comfy_cast_weights - del m.prev_comfy_cast_weights + def should_reload_model(self, force_patch_weights=False): + if force_patch_weights and self.model.lowvram_patch_counter > 0: + return True + return False - self.model_accelerated = False - - self.model.unpatch_model(self.model.offload_device) + def model_unload(self, unpatch_weights=True): + self.model.unpatch_model(self.model.offload_device, unpatch_weights=unpatch_weights) self.model.model_patches_to(self.model.offload_device) + self.weights_loaded = self.weights_loaded and not unpatch_weights + self.real_model = None def __eq__(self, other): return self.model is other.model @@ -341,31 +327,57 @@ class LoadedModel: def minimum_inference_memory(): return (1024 * 1024 * 1024) -def unload_model_clones(model): +def unload_model_clones(model, unload_weights_only=True, force_unload=True): to_unload = [] for i in range(len(current_loaded_models)): if model.is_clone(current_loaded_models[i].model): to_unload = [i] + to_unload + if len(to_unload) == 0: + return True + + same_weights = 0 for i in to_unload: - print("unload clone", i) - current_loaded_models.pop(i).model_unload() + if model.clone_has_same_weights(current_loaded_models[i].model): + same_weights += 1 + + if same_weights == len(to_unload): + unload_weight = False + else: + unload_weight = True + + if not force_unload: + if unload_weights_only and unload_weight == False: + return None + + for i in to_unload: + logging.debug("unload clone {} {}".format(i, unload_weight)) + current_loaded_models.pop(i).model_unload(unpatch_weights=unload_weight) + + return unload_weight def free_memory(memory_required, device, keep_loaded=[]): - unloaded_model = False + unloaded_model = [] + can_unload = [] + for i in range(len(current_loaded_models) -1, -1, -1): - if not DISABLE_SMART_MEMORY: - if get_free_memory(device) > memory_required: - break shift_model = current_loaded_models[i] if shift_model.device == device: if shift_model not in keep_loaded: - m = current_loaded_models.pop(i) - m.model_unload() - del m - unloaded_model = True + can_unload.append((sys.getrefcount(shift_model.model), shift_model.model_memory(), i)) - if unloaded_model: + for x in sorted(can_unload): + i = x[-1] + if not DISABLE_SMART_MEMORY: + if get_free_memory(device) > memory_required: + break + current_loaded_models[i].model_unload() + unloaded_model.append(i) + + for i in sorted(unloaded_model, reverse=True): + current_loaded_models.pop(i) + + if len(unloaded_model) > 0: soft_empty_cache() else: if vram_state != VRAMState.HIGH_VRAM: @@ -373,24 +385,36 @@ def free_memory(memory_required, device, keep_loaded=[]): if mem_free_torch > mem_free_total * 0.25: soft_empty_cache() -def load_models_gpu(models, memory_required=0): +def load_models_gpu(models, memory_required=0, force_patch_weights=False): global vram_state inference_memory = minimum_inference_memory() extra_mem = max(inference_memory, memory_required) + models = set(models) + models_to_load = [] models_already_loaded = [] for x in models: loaded_model = LoadedModel(x) + loaded = None - if loaded_model in current_loaded_models: - index = current_loaded_models.index(loaded_model) - current_loaded_models.insert(0, current_loaded_models.pop(index)) - models_already_loaded.append(loaded_model) - else: + try: + loaded_model_index = current_loaded_models.index(loaded_model) + except: + loaded_model_index = None + + if loaded_model_index is not None: + loaded = current_loaded_models[loaded_model_index] + if loaded.should_reload_model(force_patch_weights=force_patch_weights): #TODO: cleanup this model reload logic + current_loaded_models.pop(loaded_model_index).model_unload(unpatch_weights=True) + loaded = None + else: + models_already_loaded.append(loaded) + + if loaded is None: if hasattr(x, "model"): - print(f"Requested to load {x.model.__class__.__name__}") + logging.info(f"Requested to load {x.model.__class__.__name__}") models_to_load.append(loaded_model) if len(models_to_load) == 0: @@ -400,17 +424,22 @@ def load_models_gpu(models, memory_required=0): free_memory(extra_mem, d, models_already_loaded) return - print(f"Loading {len(models_to_load)} new model{'s' if len(models_to_load) > 1 else ''}") + logging.info(f"Loading {len(models_to_load)} new model{'s' if len(models_to_load) > 1 else ''}") total_memory_required = {} for loaded_model in models_to_load: - unload_model_clones(loaded_model.model) - total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device) + if unload_model_clones(loaded_model.model, unload_weights_only=True, force_unload=False) == True:#unload clones where the weights are different + total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device) for device in total_memory_required: if device != torch.device("cpu"): free_memory(total_memory_required[device] * 1.3 + extra_mem, device, models_already_loaded) + for loaded_model in models_to_load: + weights_unloaded = unload_model_clones(loaded_model.model, unload_weights_only=False, force_unload=False) #unload the rest of the clones where the weights can stay loaded + if weights_unloaded is not None: + loaded_model.weights_loaded = not weights_unloaded + for loaded_model in models_to_load: model = loaded_model.model torch_dev = model.load_device @@ -423,15 +452,13 @@ def load_models_gpu(models, memory_required=0): model_size = loaded_model.model_memory_required(torch_dev) current_free_mem = get_free_memory(torch_dev) lowvram_model_memory = int(max(64 * (1024 * 1024), (current_free_mem - 1024 * (1024 * 1024)) / 1.3 )) - if model_size > (current_free_mem - inference_memory): #only switch to lowvram if really necessary - vram_set_state = VRAMState.LOW_VRAM - else: + if model_size <= (current_free_mem - inference_memory): #only switch to lowvram if really necessary lowvram_model_memory = 0 if vram_set_state == VRAMState.NO_VRAM: lowvram_model_memory = 64 * 1024 * 1024 - cur_loaded_model = loaded_model.model_load(lowvram_model_memory) + cur_loaded_model = loaded_model.model_load(lowvram_model_memory, force_patch_weights=force_patch_weights) current_loaded_models.insert(0, loaded_model) return @@ -439,11 +466,15 @@ def load_models_gpu(models, memory_required=0): def load_model_gpu(model): return load_models_gpu([model]) -def cleanup_models(): +def cleanup_models(keep_clone_weights_loaded=False): to_delete = [] for i in range(len(current_loaded_models)): if sys.getrefcount(current_loaded_models[i].model) <= 2: - to_delete = [i] + to_delete + if not keep_clone_weights_loaded: + to_delete = [i] + to_delete + #TODO: find a less fragile way to do this. + elif sys.getrefcount(current_loaded_models[i].real_model) <= 3: #references from .real_model + the .model + to_delete = [i] + to_delete for i in to_delete: x = current_loaded_models.pop(i) @@ -487,7 +518,7 @@ def unet_inital_load_device(parameters, dtype): else: return cpu_dev -def unet_dtype(device=None, model_params=0): +def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, torch.bfloat16, torch.float32]): if args.bf16_unet: return torch.bfloat16 if args.fp16_unet: @@ -496,21 +527,32 @@ def unet_dtype(device=None, model_params=0): return torch.float8_e4m3fn if args.fp8_e5m2_unet: return torch.float8_e5m2 - if should_use_fp16(device=device, model_params=model_params): - return torch.float16 + if should_use_fp16(device=device, model_params=model_params, manual_cast=True): + if torch.float16 in supported_dtypes: + return torch.float16 + if should_use_bf16(device, model_params=model_params, manual_cast=True): + if torch.bfloat16 in supported_dtypes: + return torch.bfloat16 return torch.float32 # None means no manual cast -def unet_manual_cast(weight_dtype, inference_device): +def unet_manual_cast(weight_dtype, inference_device, supported_dtypes=[torch.float16, torch.bfloat16, torch.float32]): if weight_dtype == torch.float32: return None - fp16_supported = comfy.model_management.should_use_fp16(inference_device, prioritize_performance=False) + fp16_supported = should_use_fp16(inference_device, prioritize_performance=False) if fp16_supported and weight_dtype == torch.float16: return None - if fp16_supported: + bf16_supported = should_use_bf16(inference_device) + if bf16_supported and weight_dtype == torch.bfloat16: + return None + + if fp16_supported and torch.float16 in supported_dtypes: return torch.float16 + + elif bf16_supported and torch.bfloat16 in supported_dtypes: + return torch.bfloat16 else: return torch.float32 @@ -524,8 +566,6 @@ def text_encoder_device(): if args.gpu_only: return get_torch_device() elif vram_state == VRAMState.HIGH_VRAM or vram_state == VRAMState.NORMAL_VRAM: - if is_intel_xpu(): - return torch.device("cpu") if should_use_fp16(prioritize_performance=False): return get_torch_device() else: @@ -546,10 +586,8 @@ def text_encoder_dtype(device=None): if is_device_cpu(device): return torch.float16 - if should_use_fp16(device, prioritize_performance=False): - return torch.float16 - else: - return torch.float32 + return torch.float16 + def intermediate_device(): if args.gpu_only: @@ -591,8 +629,19 @@ def supports_dtype(device, dtype): #TODO def device_supports_non_blocking(device): if is_device_mps(device): return False #pytorch bug? mps doesn't support non blocking + if args.deterministic: #TODO: figure out why deterministic breaks non blocking from gpu to cpu (previews) + return False + if directml_enabled: + return False return True +def device_should_use_non_blocking(device): + if not device_supports_non_blocking(device): + return False + return False + # return True #TODO: figure out why this causes memory issues on Nvidia and possibly others + + def cast_to_device(tensor, device, dtype, copy=False): device_supports_cast = False if tensor.dtype == torch.float32 or tensor.dtype == torch.float16: @@ -603,7 +652,7 @@ def cast_to_device(tensor, device, dtype, copy=False): elif is_intel_xpu(): device_supports_cast = True - non_blocking = device_supports_non_blocking(device) + non_blocking = device_should_use_non_blocking(device) if device_supports_cast: if copy: @@ -646,6 +695,18 @@ def pytorch_attention_flash_attention(): return True return False +def force_upcast_attention_dtype(): + upcast = args.force_upcast_attention + try: + if platform.mac_ver()[0] in ['14.5']: #black image bug on OSX Sonoma 14.5 + upcast = True + except: + pass + if upcast: + return torch.float32 + else: + return None + def get_free_memory(dev=None, torch_free_too=False): global directml_enabled if dev is None: @@ -661,10 +722,10 @@ def get_free_memory(dev=None, torch_free_too=False): elif is_intel_xpu(): stats = torch.xpu.memory_stats(dev) mem_active = stats['active_bytes.all.current'] - mem_allocated = stats['allocated_bytes.all.current'] mem_reserved = stats['reserved_bytes.all.current'] mem_free_torch = mem_reserved - mem_active - mem_free_total = torch.xpu.get_device_properties(dev).total_memory - mem_allocated + mem_free_xpu = torch.xpu.get_device_properties(dev).total_memory - mem_reserved + mem_free_total = mem_free_xpu + mem_free_torch else: stats = torch.cuda.memory_stats(dev) mem_active = stats['active_bytes.all.current'] @@ -686,19 +747,22 @@ def mps_mode(): global cpu_state return cpu_state == CPUState.MPS -def is_device_cpu(device): +def is_device_type(device, type): if hasattr(device, 'type'): - if (device.type == 'cpu'): + if (device.type == type): return True return False +def is_device_cpu(device): + return is_device_type(device, 'cpu') + def is_device_mps(device): - if hasattr(device, 'type'): - if (device.type == 'mps'): - return True - return False + return is_device_type(device, 'mps') -def should_use_fp16(device=None, model_params=0, prioritize_performance=True): +def is_device_cuda(device): + return is_device_type(device, 'cuda') + +def should_use_fp16(device=None, model_params=0, prioritize_performance=True, manual_cast=False): global directml_enabled if device is not None: @@ -708,9 +772,9 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True): if FORCE_FP16: return True - if device is not None: #TODO + if device is not None: if is_device_mps(device): - return False + return True if FORCE_FP32: return False @@ -718,16 +782,22 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True): if directml_enabled: return False - if cpu_mode() or mps_mode(): - return False #TODO ? + if mps_mode(): + return True + + if cpu_mode(): + return False if is_intel_xpu(): return True - if torch.cuda.is_bf16_supported(): + if torch.version.hip: return True props = torch.cuda.get_device_properties("cuda") + if props.major >= 8: + return True + if props.major < 6: return False @@ -735,12 +805,12 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True): #FP16 is confirmed working on a 1080 (GP104) but it's a bit slower than FP32 so it should only be enabled #when the model doesn't actually fit on the card #TODO: actually test if GP106 and others have the same type of behavior - nvidia_10_series = ["1080", "1070", "titan x", "p3000", "p3200", "p4000", "p4200", "p5000", "p5200", "p6000", "1060", "1050"] + nvidia_10_series = ["1080", "1070", "titan x", "p3000", "p3200", "p4000", "p4200", "p5000", "p5200", "p6000", "1060", "1050", "p40", "p100", "p6", "p4"] for x in nvidia_10_series: if x in props.name.lower(): fp16_works = True - if fp16_works: + if fp16_works or manual_cast: free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory()) if (not prioritize_performance) or model_params * 4 > free_model_memory: return True @@ -756,6 +826,43 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True): return True +def should_use_bf16(device=None, model_params=0, prioritize_performance=True, manual_cast=False): + if device is not None: + if is_device_cpu(device): #TODO ? bf16 works on CPU but is extremely slow + return False + + if device is not None: #TODO not sure about mps bf16 support + if is_device_mps(device): + return False + + if FORCE_FP32: + return False + + if directml_enabled: + return False + + if cpu_mode() or mps_mode(): + return False + + if is_intel_xpu(): + return True + + if device is None: + device = torch.device("cuda") + + props = torch.cuda.get_device_properties(device) + if props.major >= 8: + return True + + bf16_works = torch.cuda.is_bf16_supported() + + if bf16_works or manual_cast: + free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory()) + if (not prioritize_performance) or model_params * 4 > free_model_memory: + return True + + return False + def soft_empty_cache(force=False): global cpu_state if cpu_state == CPUState.MPS: @@ -772,6 +879,7 @@ def unload_all_models(): def resolve_lowvram_weight(weight, model, key): #TODO: remove + print("WARNING: The comfy.model_management.resolve_lowvram_weight function will be removed soon, please stop using it.") return weight #TODO: might be cleaner to put this somewhere else diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index a88b737cc..84592f931 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -1,9 +1,55 @@ import torch import copy import inspect +import logging +import uuid import comfy.utils import comfy.model_management +from comfy.types import UnetWrapperFunction + + +def weight_decompose(dora_scale, weight, lora_diff, alpha, strength): + dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, torch.float32) + lora_diff *= alpha + weight_calc = weight + lora_diff.type(weight.dtype) + weight_norm = ( + weight_calc.transpose(0, 1) + .reshape(weight_calc.shape[1], -1) + .norm(dim=1, keepdim=True) + .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1)) + .transpose(0, 1) + ) + + weight_calc *= (dora_scale / weight_norm).type(weight.dtype) + if strength != 1.0: + weight_calc -= weight + weight += strength * (weight_calc) + else: + weight[:] = weight_calc + return weight + + +def set_model_options_patch_replace(model_options, patch, name, block_name, number, transformer_index=None): + to = model_options["transformer_options"].copy() + + if "patches_replace" not in to: + to["patches_replace"] = {} + else: + to["patches_replace"] = to["patches_replace"].copy() + + if name not in to["patches_replace"]: + to["patches_replace"][name] = {} + else: + to["patches_replace"][name] = to["patches_replace"][name].copy() + + if transformer_index is not None: + block = (block_name, number, transformer_index) + else: + block = (block_name, number) + to["patches_replace"][name][block] = patch + model_options["transformer_options"] = to + return model_options class ModelPatcher: def __init__(self, model, load_device, offload_device, size=0, current_device=None, weight_inplace_update=False): @@ -23,13 +69,14 @@ class ModelPatcher: self.current_device = current_device self.weight_inplace_update = weight_inplace_update + self.model_lowvram = False + self.lowvram_patch_counter = 0 + self.patches_uuid = uuid.uuid4() def model_size(self): if self.size > 0: return self.size - model_sd = self.model.state_dict() self.size = comfy.model_management.module_size(self.model) - self.model_keys = set(model_sd.keys()) return self.size def clone(self): @@ -37,10 +84,12 @@ class ModelPatcher: n.patches = {} for k in self.patches: n.patches[k] = self.patches[k][:] + n.patches_uuid = self.patches_uuid n.object_patches = self.object_patches.copy() n.model_options = copy.deepcopy(self.model_options) - n.model_keys = self.model_keys + n.backup = self.backup + n.object_patches_backup = self.object_patches_backup return n def is_clone(self, other): @@ -48,6 +97,19 @@ class ModelPatcher: return True return False + def clone_has_same_weights(self, clone): + if not self.is_clone(clone): + return False + + if len(self.patches) == 0 and len(clone.patches) == 0: + return True + + if self.patches_uuid == clone.patches_uuid: + if len(self.patches) != len(clone.patches): + logging.warning("WARNING: something went wrong, same patch uuid but different length of patches.") + else: + return True + def memory_required(self, input_shape): return self.model.memory_required(input_shape=input_shape) @@ -64,9 +126,12 @@ class ModelPatcher: if disable_cfg1_optimization: self.model_options["disable_cfg1_optimization"] = True - def set_model_unet_function_wrapper(self, unet_wrapper_function): + def set_model_unet_function_wrapper(self, unet_wrapper_function: UnetWrapperFunction): self.model_options["model_function_wrapper"] = unet_wrapper_function + def set_model_denoise_mask_function(self, denoise_mask_function): + self.model_options["denoise_mask_function"] = denoise_mask_function + def set_model_patch(self, patch, name): to = self.model_options["transformer_options"] if "patches" not in to: @@ -74,16 +139,7 @@ class ModelPatcher: to["patches"][name] = to["patches"].get(name, []) + [patch] def set_model_patch_replace(self, patch, name, block_name, number, transformer_index=None): - to = self.model_options["transformer_options"] - if "patches_replace" not in to: - to["patches_replace"] = {} - if name not in to["patches_replace"]: - to["patches_replace"][name] = {} - if transformer_index is not None: - block = (block_name, number, transformer_index) - else: - block = (block_name, number) - to["patches_replace"][name][block] = patch + self.model_options = set_model_options_patch_replace(self.model_options, patch, name, block_name, number, transformer_index=transformer_index) def set_model_attn1_patch(self, patch): self.set_model_patch(patch, "attn1_patch") @@ -115,6 +171,15 @@ class ModelPatcher: def add_object_patch(self, name, obj): self.object_patches[name] = obj + def get_model_object(self, name): + if name in self.object_patches: + return self.object_patches[name] + else: + if name in self.object_patches_backup: + return self.object_patches_backup[name] + else: + return comfy.utils.get_attr(self.model, name) + def model_patches_to(self, device): to = self.model_options["transformer_options"] if "patches" in to: @@ -142,13 +207,15 @@ class ModelPatcher: def add_patches(self, patches, strength_patch=1.0, strength_model=1.0): p = set() + model_sd = self.model.state_dict() for k in patches: - if k in self.model_keys: + if k in model_sd: p.add(k) current_patches = self.patches.get(k, []) current_patches.append((strength_patch, patches[k], strength_model)) self.patches[k] = current_patches + self.patches_uuid = uuid.uuid4() return list(p) def get_key_patches(self, filter_prefix=None): @@ -174,37 +241,41 @@ class ModelPatcher: sd.pop(k) return sd + def patch_weight_to_device(self, key, device_to=None): + if key not in self.patches: + return + + weight = comfy.utils.get_attr(self.model, key) + + inplace_update = self.weight_inplace_update + + if key not in self.backup: + self.backup[key] = weight.to(device=self.offload_device, copy=inplace_update) + + if device_to is not None: + temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True) + else: + temp_weight = weight.to(torch.float32, copy=True) + out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype) + if inplace_update: + comfy.utils.copy_to_param(self.model, key, out_weight) + else: + comfy.utils.set_attr_param(self.model, key, out_weight) + def patch_model(self, device_to=None, patch_weights=True): for k in self.object_patches: - old = getattr(self.model, k) + old = comfy.utils.set_attr(self.model, k, self.object_patches[k]) if k not in self.object_patches_backup: self.object_patches_backup[k] = old - setattr(self.model, k, self.object_patches[k]) if patch_weights: model_sd = self.model_state_dict() for key in self.patches: if key not in model_sd: - print("could not patch. key doesn't exist in model:", key) + logging.warning("could not patch. key doesn't exist in model: {}".format(key)) continue - weight = model_sd[key] - - inplace_update = self.weight_inplace_update - - if key not in self.backup: - self.backup[key] = weight.to(device=self.offload_device, copy=inplace_update) - - if device_to is not None: - temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True) - else: - temp_weight = weight.to(torch.float32, copy=True) - out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype) - if inplace_update: - comfy.utils.copy_to_param(self.model, key, out_weight) - else: - comfy.utils.set_attr(self.model, key, out_weight) - del temp_weight + self.patch_weight_to_device(key, device_to) if device_to is not None: self.model.to(device_to) @@ -212,9 +283,60 @@ class ModelPatcher: return self.model + def patch_model_lowvram(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False): + self.patch_model(device_to, patch_weights=False) + + logging.info("loading in lowvram mode {}".format(lowvram_model_memory/(1024 * 1024))) + class LowVramPatch: + def __init__(self, key, model_patcher): + self.key = key + self.model_patcher = model_patcher + def __call__(self, weight): + return self.model_patcher.calculate_weight(self.model_patcher.patches[self.key], weight, self.key) + + mem_counter = 0 + patch_counter = 0 + for n, m in self.model.named_modules(): + lowvram_weight = False + if hasattr(m, "comfy_cast_weights"): + module_mem = comfy.model_management.module_size(m) + if mem_counter + module_mem >= lowvram_model_memory: + lowvram_weight = True + + weight_key = "{}.weight".format(n) + bias_key = "{}.bias".format(n) + + if lowvram_weight: + if weight_key in self.patches: + if force_patch_weights: + self.patch_weight_to_device(weight_key) + else: + m.weight_function = LowVramPatch(weight_key, self) + patch_counter += 1 + if bias_key in self.patches: + if force_patch_weights: + self.patch_weight_to_device(bias_key) + else: + m.bias_function = LowVramPatch(bias_key, self) + patch_counter += 1 + + m.prev_comfy_cast_weights = m.comfy_cast_weights + m.comfy_cast_weights = True + else: + if hasattr(m, "weight"): + self.patch_weight_to_device(weight_key, device_to) + self.patch_weight_to_device(bias_key, device_to) + m.to(device_to) + mem_counter += comfy.model_management.module_size(m) + logging.debug("lowvram: loaded module regularly {}".format(m)) + + self.model_lowvram = True + self.lowvram_patch_counter = patch_counter + return self.model + def calculate_weight(self, patches, weight, key): for p in patches: - alpha = p[0] + strength = p[0] v = p[1] strength_model = p[2] @@ -232,25 +354,33 @@ class ModelPatcher: if patch_type == "diff": w1 = v[0] - if alpha != 0.0: + if strength != 0.0: if w1.shape != weight.shape: - print("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, w1.shape, weight.shape)) + logging.warning("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, w1.shape, weight.shape)) else: - weight += alpha * comfy.model_management.cast_to_device(w1, weight.device, weight.dtype) + weight += strength * comfy.model_management.cast_to_device(w1, weight.device, weight.dtype) elif patch_type == "lora": #lora/locon mat1 = comfy.model_management.cast_to_device(v[0], weight.device, torch.float32) mat2 = comfy.model_management.cast_to_device(v[1], weight.device, torch.float32) + dora_scale = v[4] if v[2] is not None: - alpha *= v[2] / mat2.shape[0] + alpha = v[2] / mat2.shape[0] + else: + alpha = 1.0 + if v[3] is not None: #locon mid weights, hopefully the math is fine because I didn't properly test it mat3 = comfy.model_management.cast_to_device(v[3], weight.device, torch.float32) final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]] mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1) try: - weight += (alpha * torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1))).reshape(weight.shape).type(weight.dtype) + lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape) + if dora_scale is not None: + weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength) + else: + weight += ((strength * alpha) * lora_diff).type(weight.dtype) except Exception as e: - print("ERROR", key, e) + logging.error("ERROR {} {} {}".format(patch_type, key, e)) elif patch_type == "lokr": w1 = v[0] w2 = v[1] @@ -259,6 +389,7 @@ class ModelPatcher: w2_a = v[5] w2_b = v[6] t2 = v[7] + dora_scale = v[8] dim = None if w1 is None: @@ -284,19 +415,29 @@ class ModelPatcher: if len(w2.shape) == 4: w1 = w1.unsqueeze(2).unsqueeze(2) if v[2] is not None and dim is not None: - alpha *= v[2] / dim + alpha = v[2] / dim + else: + alpha = 1.0 try: - weight += alpha * torch.kron(w1, w2).reshape(weight.shape).type(weight.dtype) + lora_diff = torch.kron(w1, w2).reshape(weight.shape) + if dora_scale is not None: + weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength) + else: + weight += ((strength * alpha) * lora_diff).type(weight.dtype) except Exception as e: - print("ERROR", key, e) + logging.error("ERROR {} {} {}".format(patch_type, key, e)) elif patch_type == "loha": w1a = v[0] w1b = v[1] if v[2] is not None: - alpha *= v[2] / w1b.shape[0] + alpha = v[2] / w1b.shape[0] + else: + alpha = 1.0 + w2a = v[3] w2b = v[4] + dora_scale = v[7] if v[5] is not None: #cp decomposition t1 = v[5] t2 = v[6] @@ -316,42 +457,69 @@ class ModelPatcher: comfy.model_management.cast_to_device(w2b, weight.device, torch.float32)) try: - weight += (alpha * m1 * m2).reshape(weight.shape).type(weight.dtype) + lora_diff = (m1 * m2).reshape(weight.shape) + if dora_scale is not None: + weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength) + else: + weight += ((strength * alpha) * lora_diff).type(weight.dtype) except Exception as e: - print("ERROR", key, e) + logging.error("ERROR {} {} {}".format(patch_type, key, e)) elif patch_type == "glora": if v[4] is not None: - alpha *= v[4] / v[0].shape[0] + alpha = v[4] / v[0].shape[0] + else: + alpha = 1.0 + + dora_scale = v[5] a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, torch.float32) a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, torch.float32) b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, torch.float32) b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, torch.float32) - weight += ((torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1), a2), a1)) * alpha).reshape(weight.shape).type(weight.dtype) + try: + lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1), a2), a1)).reshape(weight.shape) + if dora_scale is not None: + weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength) + else: + weight += ((strength * alpha) * lora_diff).type(weight.dtype) + except Exception as e: + logging.error("ERROR {} {} {}".format(patch_type, key, e)) else: - print("patch type not recognized", patch_type, key) + logging.warning("patch type not recognized {} {}".format(patch_type, key)) return weight - def unpatch_model(self, device_to=None): - keys = list(self.backup.keys()) + def unpatch_model(self, device_to=None, unpatch_weights=True): + if unpatch_weights: + if self.model_lowvram: + for m in self.model.modules(): + if hasattr(m, "prev_comfy_cast_weights"): + m.comfy_cast_weights = m.prev_comfy_cast_weights + del m.prev_comfy_cast_weights + m.weight_function = None + m.bias_function = None - if self.weight_inplace_update: - for k in keys: - comfy.utils.copy_to_param(self.model, k, self.backup[k]) - else: - for k in keys: - comfy.utils.set_attr(self.model, k, self.backup[k]) + self.model_lowvram = False + self.lowvram_patch_counter = 0 - self.backup = {} + keys = list(self.backup.keys()) - if device_to is not None: - self.model.to(device_to) - self.current_device = device_to + if self.weight_inplace_update: + for k in keys: + comfy.utils.copy_to_param(self.model, k, self.backup[k]) + else: + for k in keys: + comfy.utils.set_attr_param(self.model, k, self.backup[k]) + + self.backup.clear() + + if device_to is not None: + self.model.to(device_to) + self.current_device = device_to keys = list(self.object_patches_backup.keys()) for k in keys: - setattr(self.model, k, self.object_patches_backup[k]) + comfy.utils.set_attr(self.model, k, self.object_patches_backup[k]) - self.object_patches_backup = {} + self.object_patches_backup.clear() diff --git a/comfy/model_sampling.py b/comfy/model_sampling.py index cc8745c10..37976b326 100644 --- a/comfy/model_sampling.py +++ b/comfy/model_sampling.py @@ -1,5 +1,4 @@ import torch -import numpy as np from comfy.ldm.modules.diffusionmodules.util import make_beta_schedule import math @@ -12,12 +11,28 @@ class EPS: sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1)) return model_input - model_output * sigma + def noise_scaling(self, sigma, noise, latent_image, max_denoise=False): + if max_denoise: + noise = noise * torch.sqrt(1.0 + sigma ** 2.0) + else: + noise = noise * sigma + + noise += latent_image + return noise + + def inverse_noise_scaling(self, sigma, latent): + return latent class V_PREDICTION(EPS): def calculate_denoised(self, sigma, model_output, model_input): sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1)) return model_input * self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) - model_output * sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2) ** 0.5 +class EDM(V_PREDICTION): + def calculate_denoised(self, sigma, model_output, model_input): + sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1)) + return model_input * self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + model_output * sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2) ** 0.5 + class ModelSamplingDiscrete(torch.nn.Module): def __init__(self, model_config=None): @@ -42,8 +57,7 @@ class ModelSamplingDiscrete(torch.nn.Module): else: betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s) alphas = 1. - betas - alphas_cumprod = torch.tensor(np.cumprod(alphas, axis=0), dtype=torch.float32) - # alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + alphas_cumprod = torch.cumprod(alphas, dim=0) timesteps, = betas.shape self.num_timesteps = int(timesteps) @@ -58,8 +72,8 @@ class ModelSamplingDiscrete(torch.nn.Module): self.set_sigmas(sigmas) def set_sigmas(self, sigmas): - self.register_buffer('sigmas', sigmas) - self.register_buffer('log_sigmas', sigmas.log()) + self.register_buffer('sigmas', sigmas.float()) + self.register_buffer('log_sigmas', sigmas.log().float()) @property def sigma_min(self): @@ -94,8 +108,6 @@ class ModelSamplingDiscrete(torch.nn.Module): class ModelSamplingContinuousEDM(torch.nn.Module): def __init__(self, model_config=None): super().__init__() - self.sigma_data = 1.0 - if model_config is not None: sampling_settings = model_config.sampling_settings else: @@ -103,9 +115,11 @@ class ModelSamplingContinuousEDM(torch.nn.Module): sigma_min = sampling_settings.get("sigma_min", 0.002) sigma_max = sampling_settings.get("sigma_max", 120.0) - self.set_sigma_range(sigma_min, sigma_max) + sigma_data = sampling_settings.get("sigma_data", 1.0) + self.set_parameters(sigma_min, sigma_max, sigma_data) - def set_sigma_range(self, sigma_min, sigma_max): + def set_parameters(self, sigma_min, sigma_max, sigma_data): + self.sigma_data = sigma_data sigmas = torch.linspace(math.log(sigma_min), math.log(sigma_max), 1000).exp() self.register_buffer('sigmas', sigmas) #for compatibility with some schedulers @@ -134,3 +148,56 @@ class ModelSamplingContinuousEDM(torch.nn.Module): log_sigma_min = math.log(self.sigma_min) return math.exp((math.log(self.sigma_max) - log_sigma_min) * percent + log_sigma_min) + +class StableCascadeSampling(ModelSamplingDiscrete): + def __init__(self, model_config=None): + super().__init__() + + if model_config is not None: + sampling_settings = model_config.sampling_settings + else: + sampling_settings = {} + + self.set_parameters(sampling_settings.get("shift", 1.0)) + + def set_parameters(self, shift=1.0, cosine_s=8e-3): + self.shift = shift + self.cosine_s = torch.tensor(cosine_s) + self._init_alpha_cumprod = torch.cos(self.cosine_s / (1 + self.cosine_s) * torch.pi * 0.5) ** 2 + + #This part is just for compatibility with some schedulers in the codebase + self.num_timesteps = 10000 + sigmas = torch.empty((self.num_timesteps), dtype=torch.float32) + for x in range(self.num_timesteps): + t = (x + 1) / self.num_timesteps + sigmas[x] = self.sigma(t) + + self.set_sigmas(sigmas) + + def sigma(self, timestep): + alpha_cumprod = (torch.cos((timestep + self.cosine_s) / (1 + self.cosine_s) * torch.pi * 0.5) ** 2 / self._init_alpha_cumprod) + + if self.shift != 1.0: + var = alpha_cumprod + logSNR = (var/(1-var)).log() + logSNR += 2 * torch.log(1.0 / torch.tensor(self.shift)) + alpha_cumprod = logSNR.sigmoid() + + alpha_cumprod = alpha_cumprod.clamp(0.0001, 0.9999) + return ((1 - alpha_cumprod) / alpha_cumprod) ** 0.5 + + def timestep(self, sigma): + var = 1 / ((sigma * sigma) + 1) + var = var.clamp(0, 1.0) + s, min_var = self.cosine_s.to(var.device), self._init_alpha_cumprod.to(var.device) + t = (((var * min_var) ** 0.5).acos() / (torch.pi * 0.5)) * (1 + s) - s + return t + + def percent_to_sigma(self, percent): + if percent <= 0.0: + return 999999999.9 + if percent >= 1.0: + return 0.0 + + percent = 1.0 - percent + return self.sigma(torch.tensor(percent)) diff --git a/comfy/ops.py b/comfy/ops.py index f6f85de60..7ebb3dd2f 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -1,19 +1,43 @@ +""" + This file is part of ComfyUI. + Copyright (C) 2024 Stability AI + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +""" + import torch -from contextlib import contextmanager import comfy.model_management def cast_bias_weight(s, input): bias = None - non_blocking = comfy.model_management.device_supports_non_blocking(input.device) + non_blocking = comfy.model_management.device_should_use_non_blocking(input.device) if s.bias is not None: bias = s.bias.to(device=input.device, dtype=input.dtype, non_blocking=non_blocking) + if s.bias_function is not None: + bias = s.bias_function(bias) weight = s.weight.to(device=input.device, dtype=input.dtype, non_blocking=non_blocking) + if s.weight_function is not None: + weight = s.weight_function(weight) return weight, bias +class CastWeightBiasOp: + comfy_cast_weights = False + weight_function = None + bias_function = None class disable_weight_init: - class Linear(torch.nn.Linear): - comfy_cast_weights = False + class Linear(torch.nn.Linear, CastWeightBiasOp): def reset_parameters(self): return None @@ -27,8 +51,7 @@ class disable_weight_init: else: return super().forward(*args, **kwargs) - class Conv2d(torch.nn.Conv2d): - comfy_cast_weights = False + class Conv2d(torch.nn.Conv2d, CastWeightBiasOp): def reset_parameters(self): return None @@ -42,8 +65,7 @@ class disable_weight_init: else: return super().forward(*args, **kwargs) - class Conv3d(torch.nn.Conv3d): - comfy_cast_weights = False + class Conv3d(torch.nn.Conv3d, CastWeightBiasOp): def reset_parameters(self): return None @@ -57,8 +79,7 @@ class disable_weight_init: else: return super().forward(*args, **kwargs) - class GroupNorm(torch.nn.GroupNorm): - comfy_cast_weights = False + class GroupNorm(torch.nn.GroupNorm, CastWeightBiasOp): def reset_parameters(self): return None @@ -73,13 +94,16 @@ class disable_weight_init: return super().forward(*args, **kwargs) - class LayerNorm(torch.nn.LayerNorm): - comfy_cast_weights = False + class LayerNorm(torch.nn.LayerNorm, CastWeightBiasOp): def reset_parameters(self): return None def forward_comfy_cast_weights(self, input): - weight, bias = cast_bias_weight(self, input) + if self.weight is not None: + weight, bias = cast_bias_weight(self, input) + else: + weight = None + bias = None return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps) def forward(self, *args, **kwargs): @@ -88,6 +112,27 @@ class disable_weight_init: else: return super().forward(*args, **kwargs) + class ConvTranspose2d(torch.nn.ConvTranspose2d, CastWeightBiasOp): + def reset_parameters(self): + return None + + def forward_comfy_cast_weights(self, input, output_size=None): + num_spatial_dims = 2 + output_padding = self._output_padding( + input, output_size, self.stride, self.padding, self.kernel_size, + num_spatial_dims, self.dilation) + + weight, bias = cast_bias_weight(self, input) + return torch.nn.functional.conv_transpose2d( + input, weight, bias, self.stride, self.padding, + output_padding, self.groups, self.dilation) + + def forward(self, *args, **kwargs): + if self.comfy_cast_weights: + return self.forward_comfy_cast_weights(*args, **kwargs) + else: + return super().forward(*args, **kwargs) + @classmethod def conv_nd(s, dims, *args, **kwargs): if dims == 2: @@ -113,3 +158,6 @@ class manual_cast(disable_weight_init): class LayerNorm(disable_weight_init.LayerNorm): comfy_cast_weights = True + + class ConvTranspose2d(disable_weight_init.ConvTranspose2d): + comfy_cast_weights = True diff --git a/comfy/sample.py b/comfy/sample.py index 5c8a7d130..e51bd67d6 100644 --- a/comfy/sample.py +++ b/comfy/sample.py @@ -1,10 +1,9 @@ import torch import comfy.model_management import comfy.samplers -import comfy.conds import comfy.utils -import math import numpy as np +import logging def prepare_noise(latent_image, seed, noise_inds=None): """ @@ -25,94 +24,21 @@ def prepare_noise(latent_image, seed, noise_inds=None): noises = torch.cat(noises, axis=0) return noises -def prepare_mask(noise_mask, shape, device): - """ensures noise mask is of proper dimensions""" - noise_mask = torch.nn.functional.interpolate(noise_mask.reshape((-1, 1, noise_mask.shape[-2], noise_mask.shape[-1])), size=(shape[2], shape[3]), mode="bilinear") - noise_mask = torch.cat([noise_mask] * shape[1], dim=1) - noise_mask = comfy.utils.repeat_to_batch_size(noise_mask, shape[0]) - noise_mask = noise_mask.to(device) - return noise_mask - -def get_models_from_cond(cond, model_type): - models = [] - for c in cond: - if model_type in c: - models += [c[model_type]] - return models - -def convert_cond(cond): - out = [] - for c in cond: - temp = c[1].copy() - model_conds = temp.get("model_conds", {}) - if c[0] is not None: - model_conds["c_crossattn"] = comfy.conds.CONDCrossAttn(c[0]) #TODO: remove - temp["cross_attn"] = c[0] - temp["model_conds"] = model_conds - out.append(temp) - return out - -def get_additional_models(positive, negative, dtype): - """loads additional models in positive and negative conditioning""" - control_nets = set(get_models_from_cond(positive, "control") + get_models_from_cond(negative, "control")) - - inference_memory = 0 - control_models = [] - for m in control_nets: - control_models += m.get_models() - inference_memory += m.inference_memory_requirements(dtype) - - gligen = get_models_from_cond(positive, "gligen") + get_models_from_cond(negative, "gligen") - gligen = [x[1] for x in gligen] - models = control_models + gligen - return models, inference_memory +def prepare_sampling(model, noise_shape, positive, negative, noise_mask): + logging.warning("Warning: comfy.sample.prepare_sampling isn't used anymore and can be removed") + return model, positive, negative, noise_mask, [] def cleanup_additional_models(models): - """cleanup additional models that were loaded""" - for m in models: - if hasattr(m, 'cleanup'): - m.cleanup() - -def prepare_sampling(model, noise_shape, positive, negative, noise_mask): - device = model.load_device - positive = convert_cond(positive) - negative = convert_cond(negative) - - if noise_mask is not None: - noise_mask = prepare_mask(noise_mask, noise_shape, device) - - real_model = None - models, inference_memory = get_additional_models(positive, negative, model.model_dtype()) - comfy.model_management.load_models_gpu([model] + models, model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:])) + inference_memory) - real_model = model.model - - return real_model, positive, negative, noise_mask, models - + logging.warning("Warning: comfy.sample.cleanup_additional_models isn't used anymore and can be removed") def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=1.0, disable_noise=False, start_step=None, last_step=None, force_full_denoise=False, noise_mask=None, sigmas=None, callback=None, disable_pbar=False, seed=None): - real_model, positive_copy, negative_copy, noise_mask, models = prepare_sampling(model, noise.shape, positive, negative, noise_mask) + sampler = comfy.samplers.KSampler(model, steps=steps, device=model.load_device, sampler=sampler_name, scheduler=scheduler, denoise=denoise, model_options=model.model_options) - noise = noise.to(model.load_device) - latent_image = latent_image.to(model.load_device) - - sampler = comfy.samplers.KSampler(real_model, steps=steps, device=model.load_device, sampler=sampler_name, scheduler=scheduler, denoise=denoise, model_options=model.model_options) - - samples = sampler.sample(noise, positive_copy, negative_copy, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed) + samples = sampler.sample(noise, positive, negative, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed) samples = samples.to(comfy.model_management.intermediate_device()) - - cleanup_additional_models(models) - cleanup_additional_models(set(get_models_from_cond(positive_copy, "control") + get_models_from_cond(negative_copy, "control"))) return samples def sample_custom(model, noise, cfg, sampler, sigmas, positive, negative, latent_image, noise_mask=None, callback=None, disable_pbar=False, seed=None): - real_model, positive_copy, negative_copy, noise_mask, models = prepare_sampling(model, noise.shape, positive, negative, noise_mask) - noise = noise.to(model.load_device) - latent_image = latent_image.to(model.load_device) - sigmas = sigmas.to(model.load_device) - - samples = comfy.samplers.sample(real_model, noise, positive_copy, negative_copy, cfg, model.load_device, sampler, sigmas, model_options=model.model_options, latent_image=latent_image, denoise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed) + samples = comfy.samplers.sample(model, noise, positive, negative, cfg, model.load_device, sampler, sigmas, model_options=model.model_options, latent_image=latent_image, denoise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed) samples = samples.to(comfy.model_management.intermediate_device()) - cleanup_additional_models(models) - cleanup_additional_models(set(get_models_from_cond(positive_copy, "control") + get_models_from_cond(negative_copy, "control"))) return samples - diff --git a/comfy/sampler_helpers.py b/comfy/sampler_helpers.py new file mode 100644 index 000000000..a18abd9e9 --- /dev/null +++ b/comfy/sampler_helpers.py @@ -0,0 +1,76 @@ +import torch +import comfy.model_management +import comfy.conds + +def prepare_mask(noise_mask, shape, device): + """ensures noise mask is of proper dimensions""" + noise_mask = torch.nn.functional.interpolate(noise_mask.reshape((-1, 1, noise_mask.shape[-2], noise_mask.shape[-1])), size=(shape[2], shape[3]), mode="bilinear") + noise_mask = torch.cat([noise_mask] * shape[1], dim=1) + noise_mask = comfy.utils.repeat_to_batch_size(noise_mask, shape[0]) + noise_mask = noise_mask.to(device) + return noise_mask + +def get_models_from_cond(cond, model_type): + models = [] + for c in cond: + if model_type in c: + models += [c[model_type]] + return models + +def convert_cond(cond): + out = [] + for c in cond: + temp = c[1].copy() + model_conds = temp.get("model_conds", {}) + if c[0] is not None: + model_conds["c_crossattn"] = comfy.conds.CONDCrossAttn(c[0]) #TODO: remove + temp["cross_attn"] = c[0] + temp["model_conds"] = model_conds + out.append(temp) + return out + +def get_additional_models(conds, dtype): + """loads additional models in conditioning""" + cnets = [] + gligen = [] + + for k in conds: + cnets += get_models_from_cond(conds[k], "control") + gligen += get_models_from_cond(conds[k], "gligen") + + control_nets = set(cnets) + + inference_memory = 0 + control_models = [] + for m in control_nets: + control_models += m.get_models() + inference_memory += m.inference_memory_requirements(dtype) + + gligen = [x[1] for x in gligen] + models = control_models + gligen + return models, inference_memory + +def cleanup_additional_models(models): + """cleanup additional models that were loaded""" + for m in models: + if hasattr(m, 'cleanup'): + m.cleanup() + + +def prepare_sampling(model, noise_shape, conds): + device = model.load_device + real_model = None + models, inference_memory = get_additional_models(conds, model.model_dtype()) + comfy.model_management.load_models_gpu([model] + models, model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:])) + inference_memory) + real_model = model.model + + return real_model, conds, models + +def cleanup_models(conds, models): + cleanup_additional_models(models) + + control_cleanup = [] + for k in conds: + control_cleanup += get_models_from_cond(conds[k], "control") + + cleanup_additional_models(set(control_cleanup)) diff --git a/comfy/samplers.py b/comfy/samplers.py index 89d8d4f28..29962a916 100644 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -1,13 +1,11 @@ from .k_diffusion import sampling as k_diffusion_sampling from .extra_samplers import uni_pc import torch -import enum import collections from comfy import model_management import math -from comfy import model_base -import comfy.utils -import comfy.conds +import logging +import comfy.sampler_helpers def get_area_and_mult(conds, x_in, timestep_in): area = (x_in.shape[2], x_in.shape[3], 0, 0) @@ -36,7 +34,7 @@ def get_area_and_mult(conds, x_in, timestep_in): mask = conds['mask'] assert(mask.shape[1] == x_in.shape[2]) assert(mask.shape[2] == x_in.shape[3]) - mask = mask[:,area[2]:area[0] + area[2],area[3]:area[1] + area[3]] * mask_strength + mask = mask[:input_x.shape[0],area[2]:area[0] + area[2],area[3]:area[1] + area[3]] * mask_strength mask = mask.unsqueeze(1).repeat(input_x.shape[0] // mask.shape[0], input_x.shape[1], 1, 1) else: mask = torch.ones_like(input_x) @@ -130,30 +128,23 @@ def cond_cat(c_list): return out -def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): - out_cond = torch.zeros_like(x_in) - out_count = torch.ones_like(x_in) * 1e-37 - - out_uncond = torch.zeros_like(x_in) - out_uncond_count = torch.ones_like(x_in) * 1e-37 - - COND = 0 - UNCOND = 1 - +def calc_cond_batch(model, conds, x_in, timestep, model_options): + out_conds = [] + out_counts = [] to_run = [] - for x in cond: - p = get_area_and_mult(x, x_in, timestep) - if p is None: - continue - to_run += [(p, COND)] - if uncond is not None: - for x in uncond: - p = get_area_and_mult(x, x_in, timestep) - if p is None: - continue + for i in range(len(conds)): + out_conds.append(torch.zeros_like(x_in)) + out_counts.append(torch.ones_like(x_in) * 1e-37) - to_run += [(p, UNCOND)] + cond = conds[i] + if cond is not None: + for x in cond: + p = get_area_and_mult(x, x_in, timestep) + if p is None: + continue + + to_run += [(p, i)] while len(to_run) > 0: first = to_run[0] @@ -212,6 +203,7 @@ def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): cur_patches[p] = cur_patches[p] + patches[p] else: cur_patches[p] = patches[p] + transformer_options["patches"] = cur_patches else: transformer_options["patches"] = patches @@ -224,71 +216,66 @@ def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): output = model_options['model_function_wrapper'](model.apply_model, {"input": input_x, "timestep": timestep_, "c": c, "cond_or_uncond": cond_or_uncond}).chunk(batch_chunks) else: output = model.apply_model(input_x, timestep_, **c).chunk(batch_chunks) - del input_x for o in range(batch_chunks): - if cond_or_uncond[o] == COND: - out_cond[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += output[o] * mult[o] - out_count[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += mult[o] - else: - out_uncond[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += output[o] * mult[o] - out_uncond_count[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += mult[o] - del mult + cond_index = cond_or_uncond[o] + out_conds[cond_index][:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += output[o] * mult[o] + out_counts[cond_index][:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += mult[o] - out_cond /= out_count - del out_count - out_uncond /= out_uncond_count - del out_uncond_count - return out_cond, out_uncond + for i in range(len(out_conds)): + out_conds[i] /= out_counts[i] + + return out_conds + +def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): #TODO: remove + logging.warning("WARNING: The comfy.samplers.calc_cond_uncond_batch function is deprecated please use the calc_cond_batch one instead.") + return tuple(calc_cond_batch(model, [cond, uncond], x_in, timestep, model_options)) + +def cfg_function(model, cond_pred, uncond_pred, cond_scale, x, timestep, model_options={}, cond=None, uncond=None): + if "sampler_cfg_function" in model_options: + args = {"cond": x - cond_pred, "uncond": x - uncond_pred, "cond_scale": cond_scale, "timestep": timestep, "input": x, "sigma": timestep, + "cond_denoised": cond_pred, "uncond_denoised": uncond_pred, "model": model, "model_options": model_options} + cfg_result = x - model_options["sampler_cfg_function"](args) + else: + cfg_result = uncond_pred + (cond_pred - uncond_pred) * cond_scale + + for fn in model_options.get("sampler_post_cfg_function", []): + args = {"denoised": cfg_result, "cond": cond, "uncond": uncond, "model": model, "uncond_denoised": uncond_pred, "cond_denoised": cond_pred, + "sigma": timestep, "model_options": model_options, "input": x} + cfg_result = fn(args) + + return cfg_result #The main sampling function shared by all the samplers #Returns denoised def sampling_function(model, x, timestep, uncond, cond, cond_scale, model_options={}, seed=None): - if math.isclose(cond_scale, 1.0) and model_options.get("disable_cfg1_optimization", False) == False: - uncond_ = None - else: - uncond_ = uncond + if math.isclose(cond_scale, 1.0) and model_options.get("disable_cfg1_optimization", False) == False: + uncond_ = None + else: + uncond_ = uncond - cond_pred, uncond_pred = calc_cond_uncond_batch(model, cond, uncond_, x, timestep, model_options) - if "sampler_cfg_function" in model_options: - args = {"cond": x - cond_pred, "uncond": x - uncond_pred, "cond_scale": cond_scale, "timestep": timestep, "input": x, "sigma": timestep, - "cond_denoised": cond_pred, "uncond_denoised": uncond_pred, "model": model, "model_options": model_options} - cfg_result = x - model_options["sampler_cfg_function"](args) - else: - cfg_result = uncond_pred + (cond_pred - uncond_pred) * cond_scale + conds = [cond, uncond_] + out = calc_cond_batch(model, conds, x, timestep, model_options) + return cfg_function(model, out[0], out[1], cond_scale, x, timestep, model_options=model_options, cond=cond, uncond=uncond_) - for fn in model_options.get("sampler_post_cfg_function", []): - args = {"denoised": cfg_result, "cond": cond, "uncond": uncond, "model": model, "uncond_denoised": uncond_pred, "cond_denoised": cond_pred, - "sigma": timestep, "model_options": model_options, "input": x} - cfg_result = fn(args) - return cfg_result - -class CFGNoisePredictor(torch.nn.Module): - def __init__(self, model): - super().__init__() +class KSamplerX0Inpaint: + def __init__(self, model, sigmas): self.inner_model = model - def apply_model(self, x, timestep, cond, uncond, cond_scale, model_options={}, seed=None): - out = sampling_function(self.inner_model, x, timestep, uncond, cond, cond_scale, model_options=model_options, seed=seed) - return out - def forward(self, *args, **kwargs): - return self.apply_model(*args, **kwargs) - -class KSamplerX0Inpaint(torch.nn.Module): - def __init__(self, model): - super().__init__() - self.inner_model = model - def forward(self, x, sigma, uncond, cond, cond_scale, denoise_mask, model_options={}, seed=None): + self.sigmas = sigmas + def __call__(self, x, sigma, denoise_mask, model_options={}, seed=None): if denoise_mask is not None: + if "denoise_mask_function" in model_options: + denoise_mask = model_options["denoise_mask_function"](sigma, denoise_mask, extra_options={"model": self.inner_model, "sigmas": self.sigmas}) latent_mask = 1. - denoise_mask - x = x * denoise_mask + (self.latent_image + self.noise * sigma.reshape([sigma.shape[0]] + [1] * (len(self.noise.shape) - 1))) * latent_mask - out = self.inner_model(x, sigma, cond=cond, uncond=uncond, cond_scale=cond_scale, model_options=model_options, seed=seed) + x = x * denoise_mask + self.inner_model.inner_model.model_sampling.noise_scaling(sigma.reshape([sigma.shape[0]] + [1] * (len(self.noise.shape) - 1)), self.noise, self.latent_image) * latent_mask + out = self.inner_model(x, sigma, model_options=model_options, seed=seed) if denoise_mask is not None: out = out * denoise_mask + self.latent_image * latent_mask return out -def simple_scheduler(model, steps): - s = model.model_sampling +def simple_scheduler(model_sampling, steps): + s = model_sampling sigs = [] ss = len(s.sigmas) / steps for x in range(steps): @@ -296,10 +283,10 @@ def simple_scheduler(model, steps): sigs += [0.0] return torch.FloatTensor(sigs) -def ddim_scheduler(model, steps): - s = model.model_sampling +def ddim_scheduler(model_sampling, steps): + s = model_sampling sigs = [] - ss = len(s.sigmas) // steps + ss = max(len(s.sigmas) // steps, 1) x = 1 while x < len(s.sigmas): sigs += [float(s.sigmas[x])] @@ -308,8 +295,8 @@ def ddim_scheduler(model, steps): sigs += [0.0] return torch.FloatTensor(sigs) -def normal_scheduler(model, steps, sgm=False, floor=False): - s = model.model_sampling +def normal_scheduler(model_sampling, steps, sgm=False, floor=False): + s = model_sampling start = s.timestep(s.sigma_max) end = s.timestep(s.sigma_min) @@ -517,14 +504,6 @@ class Sampler: sigma = float(sigmas[0]) return math.isclose(max_sigma, sigma, rel_tol=1e-05) or sigma > max_sigma -class UNIPC(Sampler): - def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False): - return uni_pc.sample_unipc(model_wrap, noise, latent_image, sigmas, max_denoise=self.max_denoise(model_wrap, sigmas), extra_args=extra_args, noise_mask=denoise_mask, callback=callback, disable=disable_pbar) - -class UNIPCBH2(Sampler): - def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False): - return uni_pc.sample_unipc(model_wrap, noise, latent_image, sigmas, max_denoise=self.max_denoise(model_wrap, sigmas), extra_args=extra_args, noise_mask=denoise_mask, callback=callback, variant='bh2', disable=disable_pbar) - KSAMPLER_NAMES = ["euler", "euler_ancestral", "heun", "heunpp2","dpm_2", "dpm_2_ancestral", "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_sde", "dpmpp_sde_gpu", "dpmpp_2m", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm"] @@ -537,7 +516,7 @@ class KSAMPLER(Sampler): def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False): extra_args["denoise_mask"] = denoise_mask - model_k = KSamplerX0Inpaint(model_wrap) + model_k = KSamplerX0Inpaint(model_wrap, sigmas) model_k.latent_image = latent_image if self.inpaint_options.get("random", False): #TODO: Should this be the default? generator = torch.manual_seed(extra_args.get("seed", 41) + 1) @@ -545,26 +524,24 @@ class KSAMPLER(Sampler): else: model_k.noise = noise - if self.max_denoise(model_wrap, sigmas): - noise = noise * torch.sqrt(1.0 + sigmas[0] ** 2.0) - else: - noise = noise * sigmas[0] + noise = model_wrap.inner_model.model_sampling.noise_scaling(sigmas[0], noise, latent_image, self.max_denoise(model_wrap, sigmas)) k_callback = None total_steps = len(sigmas) - 1 if callback is not None: k_callback = lambda x: callback(x["i"], x["denoised"], x["x"], total_steps) - if latent_image is not None: - noise += latent_image - samples = self.sampler_function(model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar, **self.extra_options) + samples = model_wrap.inner_model.model_sampling.inverse_noise_scaling(sigmas[-1], samples) return samples def ksampler(sampler_name, extra_options={}, inpaint_options={}): if sampler_name == "dpm_fast": def dpm_fast_function(model, noise, sigmas, extra_args, callback, disable): + if len(sigmas) <= 1: + return noise + sigma_min = sigmas[-1] if sigma_min == 0: sigma_min = sigmas[-2] @@ -572,81 +549,145 @@ def ksampler(sampler_name, extra_options={}, inpaint_options={}): return k_diffusion_sampling.sample_dpm_fast(model, noise, sigma_min, sigmas[0], total_steps, extra_args=extra_args, callback=callback, disable=disable) sampler_function = dpm_fast_function elif sampler_name == "dpm_adaptive": - def dpm_adaptive_function(model, noise, sigmas, extra_args, callback, disable): + def dpm_adaptive_function(model, noise, sigmas, extra_args, callback, disable, **extra_options): + if len(sigmas) <= 1: + return noise + sigma_min = sigmas[-1] if sigma_min == 0: sigma_min = sigmas[-2] - return k_diffusion_sampling.sample_dpm_adaptive(model, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=callback, disable=disable) + return k_diffusion_sampling.sample_dpm_adaptive(model, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=callback, disable=disable, **extra_options) sampler_function = dpm_adaptive_function else: sampler_function = getattr(k_diffusion_sampling, "sample_{}".format(sampler_name)) return KSAMPLER(sampler_function, extra_options, inpaint_options) -def wrap_model(model): - model_denoise = CFGNoisePredictor(model) - return model_denoise -def sample(model, noise, positive, negative, cfg, device, sampler, sigmas, model_options={}, latent_image=None, denoise_mask=None, callback=None, disable_pbar=False, seed=None): - positive = positive[:] - negative = negative[:] +def process_conds(model, noise, conds, device, latent_image=None, denoise_mask=None, seed=None): + for k in conds: + conds[k] = conds[k][:] + resolve_areas_and_cond_masks(conds[k], noise.shape[2], noise.shape[3], device) - resolve_areas_and_cond_masks(positive, noise.shape[2], noise.shape[3], device) - resolve_areas_and_cond_masks(negative, noise.shape[2], noise.shape[3], device) - - model_wrap = wrap_model(model) - - calculate_start_end_timesteps(model, negative) - calculate_start_end_timesteps(model, positive) - - if latent_image is not None: - latent_image = model.process_latent_in(latent_image) + for k in conds: + calculate_start_end_timesteps(model, conds[k]) if hasattr(model, 'extra_conds'): - positive = encode_model_conds(model.extra_conds, positive, noise, device, "positive", latent_image=latent_image, denoise_mask=denoise_mask, seed=seed) - negative = encode_model_conds(model.extra_conds, negative, noise, device, "negative", latent_image=latent_image, denoise_mask=denoise_mask, seed=seed) + for k in conds: + conds[k] = encode_model_conds(model.extra_conds, conds[k], noise, device, k, latent_image=latent_image, denoise_mask=denoise_mask, seed=seed) #make sure each cond area has an opposite one with the same area - for c in positive: - create_cond_with_same_area_if_none(negative, c) - for c in negative: - create_cond_with_same_area_if_none(positive, c) + for k in conds: + for c in conds[k]: + for kk in conds: + if k != kk: + create_cond_with_same_area_if_none(conds[kk], c) - pre_run_control(model, negative + positive) + for k in conds: + pre_run_control(model, conds[k]) - apply_empty_x_to_equal_area(list(filter(lambda c: c.get('control_apply_to_uncond', False) == True, positive)), negative, 'control', lambda cond_cnets, x: cond_cnets[x]) - apply_empty_x_to_equal_area(positive, negative, 'gligen', lambda cond_cnets, x: cond_cnets[x]) + if "positive" in conds: + positive = conds["positive"] + for k in conds: + if k != "positive": + apply_empty_x_to_equal_area(list(filter(lambda c: c.get('control_apply_to_uncond', False) == True, positive)), conds[k], 'control', lambda cond_cnets, x: cond_cnets[x]) + apply_empty_x_to_equal_area(positive, conds[k], 'gligen', lambda cond_cnets, x: cond_cnets[x]) - extra_args = {"cond":positive, "uncond":negative, "cond_scale": cfg, "model_options": model_options, "seed":seed} + return conds + +class CFGGuider: + def __init__(self, model_patcher): + self.model_patcher = model_patcher + self.model_options = model_patcher.model_options + self.original_conds = {} + self.cfg = 1.0 + + def set_conds(self, positive, negative): + self.inner_set_conds({"positive": positive, "negative": negative}) + + def set_cfg(self, cfg): + self.cfg = cfg + + def inner_set_conds(self, conds): + for k in conds: + self.original_conds[k] = comfy.sampler_helpers.convert_cond(conds[k]) + + def __call__(self, *args, **kwargs): + return self.predict_noise(*args, **kwargs) + + def predict_noise(self, x, timestep, model_options={}, seed=None): + return sampling_function(self.inner_model, x, timestep, self.conds.get("negative", None), self.conds.get("positive", None), self.cfg, model_options=model_options, seed=seed) + + def inner_sample(self, noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed): + if latent_image is not None and torch.count_nonzero(latent_image) > 0: #Don't shift the empty latent image. + latent_image = self.inner_model.process_latent_in(latent_image) + + self.conds = process_conds(self.inner_model, noise, self.conds, device, latent_image, denoise_mask, seed) + + extra_args = {"model_options": self.model_options, "seed":seed} + + samples = sampler.sample(self, sigmas, extra_args, callback, noise, latent_image, denoise_mask, disable_pbar) + return self.inner_model.process_latent_out(samples.to(torch.float32)) + + def sample(self, noise, latent_image, sampler, sigmas, denoise_mask=None, callback=None, disable_pbar=False, seed=None): + if sigmas.shape[-1] == 0: + return latent_image + + self.conds = {} + for k in self.original_conds: + self.conds[k] = list(map(lambda a: a.copy(), self.original_conds[k])) + + self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds) + device = self.model_patcher.load_device + + if denoise_mask is not None: + denoise_mask = comfy.sampler_helpers.prepare_mask(denoise_mask, noise.shape, device) + + noise = noise.to(device) + latent_image = latent_image.to(device) + sigmas = sigmas.to(device) + + output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed) + + comfy.sampler_helpers.cleanup_models(self.conds, self.loaded_models) + del self.inner_model + del self.conds + del self.loaded_models + return output + + +def sample(model, noise, positive, negative, cfg, device, sampler, sigmas, model_options={}, latent_image=None, denoise_mask=None, callback=None, disable_pbar=False, seed=None): + cfg_guider = CFGGuider(model) + cfg_guider.set_conds(positive, negative) + cfg_guider.set_cfg(cfg) + return cfg_guider.sample(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed) - samples = sampler.sample(model_wrap, sigmas, extra_args, callback, noise, latent_image, denoise_mask, disable_pbar) - return model.process_latent_out(samples.to(torch.float32)) SCHEDULER_NAMES = ["normal", "karras", "exponential", "sgm_uniform", "simple", "ddim_uniform"] SAMPLER_NAMES = KSAMPLER_NAMES + ["ddim", "uni_pc", "uni_pc_bh2"] -def calculate_sigmas_scheduler(model, scheduler_name, steps): +def calculate_sigmas(model_sampling, scheduler_name, steps): if scheduler_name == "karras": - sigmas = k_diffusion_sampling.get_sigmas_karras(n=steps, sigma_min=float(model.model_sampling.sigma_min), sigma_max=float(model.model_sampling.sigma_max)) + sigmas = k_diffusion_sampling.get_sigmas_karras(n=steps, sigma_min=float(model_sampling.sigma_min), sigma_max=float(model_sampling.sigma_max)) elif scheduler_name == "exponential": - sigmas = k_diffusion_sampling.get_sigmas_exponential(n=steps, sigma_min=float(model.model_sampling.sigma_min), sigma_max=float(model.model_sampling.sigma_max)) + sigmas = k_diffusion_sampling.get_sigmas_exponential(n=steps, sigma_min=float(model_sampling.sigma_min), sigma_max=float(model_sampling.sigma_max)) elif scheduler_name == "normal": - sigmas = normal_scheduler(model, steps) + sigmas = normal_scheduler(model_sampling, steps) elif scheduler_name == "simple": - sigmas = simple_scheduler(model, steps) + sigmas = simple_scheduler(model_sampling, steps) elif scheduler_name == "ddim_uniform": - sigmas = ddim_scheduler(model, steps) + sigmas = ddim_scheduler(model_sampling, steps) elif scheduler_name == "sgm_uniform": - sigmas = normal_scheduler(model, steps, sgm=True) + sigmas = normal_scheduler(model_sampling, steps, sgm=True) else: - print("error invalid scheduler", self.scheduler) + logging.error("error invalid scheduler {}".format(scheduler_name)) return sigmas def sampler_object(name): if name == "uni_pc": - sampler = UNIPC() + sampler = KSAMPLER(uni_pc.sample_unipc) elif name == "uni_pc_bh2": - sampler = UNIPCBH2() + sampler = KSAMPLER(uni_pc.sample_unipc_bh2) elif name == "ddim": sampler = ksampler("euler", inpaint_options={"random": True}) else: @@ -656,6 +697,7 @@ def sampler_object(name): class KSampler: SCHEDULERS = SCHEDULER_NAMES SAMPLERS = SAMPLER_NAMES + DISCARD_PENULTIMATE_SIGMA_SAMPLERS = set(('dpm_2', 'dpm_2_ancestral', 'uni_pc', 'uni_pc_bh2')) def __init__(self, model, steps, device, sampler=None, scheduler=None, denoise=None, model_options={}): self.model = model @@ -674,11 +716,11 @@ class KSampler: sigmas = None discard_penultimate_sigma = False - if self.sampler in ['dpm_2', 'dpm_2_ancestral', 'uni_pc', 'uni_pc_bh2']: + if self.sampler in self.DISCARD_PENULTIMATE_SIGMA_SAMPLERS: steps += 1 discard_penultimate_sigma = True - sigmas = calculate_sigmas_scheduler(self.model, self.scheduler, steps) + sigmas = calculate_sigmas(self.model.get_model_object("model_sampling"), self.scheduler, steps) if discard_penultimate_sigma: sigmas = torch.cat([sigmas[:-2], sigmas[-1:]]) @@ -689,9 +731,12 @@ class KSampler: if denoise is None or denoise > 0.9999: self.sigmas = self.calculate_sigmas(steps).to(self.device) else: - new_steps = int(steps/denoise) - sigmas = self.calculate_sigmas(new_steps).to(self.device) - self.sigmas = sigmas[-(steps + 1):] + if denoise <= 0.0: + self.sigmas = torch.FloatTensor([]) + else: + new_steps = int(steps/denoise) + sigmas = self.calculate_sigmas(new_steps).to(self.device) + self.sigmas = sigmas[-(steps + 1):] def sample(self, noise, positive, negative, cfg, latent_image=None, start_step=None, last_step=None, force_full_denoise=False, denoise_mask=None, sigmas=None, callback=None, disable_pbar=False, seed=None): if sigmas is None: diff --git a/comfy/sd.py b/comfy/sd.py index 1ff25bec6..343d2a02c 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1,10 +1,12 @@ import torch -import contextlib -import math +from enum import Enum +import logging from comfy import model_management -from .ldm.util import instantiate_from_config from .ldm.models.autoencoder import AutoencoderKL, AutoencodingEngine +from .ldm.cascade.stage_a import StageA +from .ldm.cascade.stage_c_coder import StageC_coder + import yaml import comfy.utils @@ -12,7 +14,6 @@ import comfy.utils from . import clip_vision from . import gligen from . import diffusers_convert -from . import model_base from . import model_detection from . import sd1_clip @@ -36,7 +37,7 @@ def load_model_weights(model, sd): w = sd.pop(x) del w if len(m) > 0: - print("missing", m) + logging.warning("missing {}".format(m)) return model def load_clip_weights(model, sd): @@ -51,7 +52,7 @@ def load_clip_weights(model, sd): if ids.dtype == torch.float32: sd['cond_stage_model.transformer.text_model.embeddings.position_ids'] = ids.round() - sd = comfy.utils.transformers_convert(sd, "cond_stage_model.model.", "cond_stage_model.transformer.text_model.", 24) + sd = comfy.utils.clip_text_transformers_convert(sd, "cond_stage_model.model.", "cond_stage_model.transformer.") return load_model_weights(model, sd) @@ -80,7 +81,7 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip): k1 = set(k1) for x in loaded: if (x not in k) and (x not in k1): - print("NOT LOADED", x) + logging.warning("NOT LOADED {}".format(x)) return (new_modelpatcher, new_clip) @@ -122,10 +123,13 @@ class CLIP: return self.tokenizer.tokenize_with_weights(text, return_word_ids) def encode_from_tokens(self, tokens, return_pooled=False): + self.cond_stage_model.reset_clip_options() + if self.layer_idx is not None: - self.cond_stage_model.clip_layer(self.layer_idx) - else: - self.cond_stage_model.reset_clip_layer() + self.cond_stage_model.set_clip_options({"layer": self.layer_idx}) + + if return_pooled == "unprojected": + self.cond_stage_model.set_clip_options({"projected_pooled": False}) self.load_model() cond, pooled = self.cond_stage_model.encode_token_weights(tokens) @@ -137,8 +141,11 @@ class CLIP: tokens = self.tokenize(text) return self.encode_from_tokens(tokens) - def load_sd(self, sd): - return self.cond_stage_model.load_sd(sd) + def load_sd(self, sd, full_model=False): + if full_model: + return self.cond_stage_model.load_state_dict(sd, strict=False) + else: + return self.cond_stage_model.load_sd(sd) def get_sd(self): return self.cond_stage_model.state_dict() @@ -158,7 +165,10 @@ class VAE: self.memory_used_encode = lambda shape, dtype: (1767 * shape[2] * shape[3]) * model_management.dtype_size(dtype) #These are for AutoencoderKL and need tweaking (should be lower) self.memory_used_decode = lambda shape, dtype: (2178 * shape[2] * shape[3] * 64) * model_management.dtype_size(dtype) self.downscale_ratio = 8 + self.upscale_ratio = 8 self.latent_channels = 4 + self.process_input = lambda image: image * 2.0 - 1.0 + self.process_output = lambda image: torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0) if config is None: if "decoder.mid.block_1.mix_factor" in sd: @@ -171,25 +181,64 @@ class VAE: decoder_config={'target': "comfy.ldm.modules.temporal_ae.VideoDecoder", 'params': decoder_config}) elif "taesd_decoder.1.weight" in sd: self.first_stage_model = comfy.taesd.taesd.TAESD() - else: + elif "vquantizer.codebook.weight" in sd: #VQGan: stage a of stable cascade + self.first_stage_model = StageA() + self.downscale_ratio = 4 + self.upscale_ratio = 4 + #TODO + #self.memory_used_encode + #self.memory_used_decode + self.process_input = lambda image: image + self.process_output = lambda image: image + elif "backbone.1.0.block.0.1.num_batches_tracked" in sd: #effnet: encoder for stage c latent of stable cascade + self.first_stage_model = StageC_coder() + self.downscale_ratio = 32 + self.latent_channels = 16 + new_sd = {} + for k in sd: + new_sd["encoder.{}".format(k)] = sd[k] + sd = new_sd + elif "blocks.11.num_batches_tracked" in sd: #previewer: decoder for stage c latent of stable cascade + self.first_stage_model = StageC_coder() + self.latent_channels = 16 + new_sd = {} + for k in sd: + new_sd["previewer.{}".format(k)] = sd[k] + sd = new_sd + elif "encoder.backbone.1.0.block.0.1.num_batches_tracked" in sd: #combined effnet and previewer for stable cascade + self.first_stage_model = StageC_coder() + self.downscale_ratio = 32 + self.latent_channels = 16 + elif "decoder.conv_in.weight" in sd: #default SD1.x/SD2.x VAE parameters ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0} - if 'encoder.down.2.downsample.conv.weight' not in sd: #Stable diffusion x4 upscaler VAE + if 'encoder.down.2.downsample.conv.weight' not in sd and 'decoder.up.3.upsample.conv.weight' not in sd: #Stable diffusion x4 upscaler VAE ddconfig['ch_mult'] = [1, 2, 4] self.downscale_ratio = 4 + self.upscale_ratio = 4 - self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=4) + self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] + if 'quant_conv.weight' in sd: + self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=4) + else: + self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"}, + encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig}, + decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig}) + else: + logging.warning("WARNING: No VAE weights detected, VAE not initalized.") + self.first_stage_model = None + return else: self.first_stage_model = AutoencoderKL(**(config['params'])) self.first_stage_model = self.first_stage_model.eval() m, u = self.first_stage_model.load_state_dict(sd, strict=False) if len(m) > 0: - print("Missing VAE keys", m) + logging.warning("Missing VAE keys {}".format(m)) if len(u) > 0: - print("Leftover VAE keys", u) + logging.debug("Leftover VAE keys {}".format(u)) if device is None: device = model_management.vae_device() @@ -203,18 +252,27 @@ class VAE: self.patcher = comfy.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device) + def vae_encode_crop_pixels(self, pixels): + x = (pixels.shape[1] // self.downscale_ratio) * self.downscale_ratio + y = (pixels.shape[2] // self.downscale_ratio) * self.downscale_ratio + if pixels.shape[1] != x or pixels.shape[2] != y: + x_offset = (pixels.shape[1] % self.downscale_ratio) // 2 + y_offset = (pixels.shape[2] % self.downscale_ratio) // 2 + pixels = pixels[:, x_offset:x + x_offset, y_offset:y + y_offset, :] + return pixels + def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16): steps = samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x, tile_y, overlap) steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x // 2, tile_y * 2, overlap) steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x * 2, tile_y // 2, overlap) pbar = comfy.utils.ProgressBar(steps) - decode_fn = lambda a: (self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)) + 1.0).float() - output = torch.clamp(( - (comfy.utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = self.downscale_ratio, output_device=self.output_device, pbar = pbar) + - comfy.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = self.downscale_ratio, output_device=self.output_device, pbar = pbar) + - comfy.utils.tiled_scale(samples, decode_fn, tile_x, tile_y, overlap, upscale_amount = self.downscale_ratio, output_device=self.output_device, pbar = pbar)) - / 3.0) / 2.0, min=0.0, max=1.0) + decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float() + output = self.process_output( + (comfy.utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar) + + comfy.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar) + + comfy.utils.tiled_scale(samples, decode_fn, tile_x, tile_y, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar)) + / 3.0) return output def encode_tiled_(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64): @@ -223,7 +281,7 @@ class VAE: steps += pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x * 2, tile_y // 2, overlap) pbar = comfy.utils.ProgressBar(steps) - encode_fn = lambda a: self.first_stage_model.encode((2. * a - 1.).to(self.vae_dtype).to(self.device)).float() + encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float() samples = comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar) samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar) samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar) @@ -238,12 +296,12 @@ class VAE: batch_number = int(free_memory / memory_used) batch_number = max(1, batch_number) - pixel_samples = torch.empty((samples_in.shape[0], 3, round(samples_in.shape[2] * self.downscale_ratio), round(samples_in.shape[3] * self.downscale_ratio)), device=self.output_device) + pixel_samples = torch.empty((samples_in.shape[0], 3, round(samples_in.shape[2] * self.upscale_ratio), round(samples_in.shape[3] * self.upscale_ratio)), device=self.output_device) for x in range(0, samples_in.shape[0], batch_number): samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device) - pixel_samples[x:x+batch_number] = torch.clamp((self.first_stage_model.decode(samples).to(self.output_device).float() + 1.0) / 2.0, min=0.0, max=1.0) + pixel_samples[x:x+batch_number] = self.process_output(self.first_stage_model.decode(samples).to(self.output_device).float()) except model_management.OOM_EXCEPTION as e: - print("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.") + logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.") pixel_samples = self.decode_tiled_(samples_in) pixel_samples = pixel_samples.to(self.output_device).movedim(1,-1) @@ -255,6 +313,7 @@ class VAE: return output.movedim(1,-1) def encode(self, pixel_samples): + pixel_samples = self.vae_encode_crop_pixels(pixel_samples) pixel_samples = pixel_samples.movedim(-1,1) try: memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype) @@ -264,16 +323,17 @@ class VAE: batch_number = max(1, batch_number) samples = torch.empty((pixel_samples.shape[0], self.latent_channels, round(pixel_samples.shape[2] // self.downscale_ratio), round(pixel_samples.shape[3] // self.downscale_ratio)), device=self.output_device) for x in range(0, pixel_samples.shape[0], batch_number): - pixels_in = (2. * pixel_samples[x:x+batch_number] - 1.).to(self.vae_dtype).to(self.device) + pixels_in = self.process_input(pixel_samples[x:x+batch_number]).to(self.vae_dtype).to(self.device) samples[x:x+batch_number] = self.first_stage_model.encode(pixels_in).to(self.output_device).float() except model_management.OOM_EXCEPTION as e: - print("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.") + logging.warning("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.") samples = self.encode_tiled_(pixel_samples) return samples def encode_tiled(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64): + pixel_samples = self.vae_encode_crop_pixels(pixel_samples) model_management.load_model_gpu(self.patcher) pixel_samples = pixel_samples.movedim(-1,1) samples = self.encode_tiled_(pixel_samples, tile_x=tile_x, tile_y=tile_y, overlap=overlap) @@ -300,8 +360,11 @@ def load_style_model(ckpt_path): model.load_state_dict(model_data) return StyleModel(model) +class CLIPType(Enum): + STABLE_DIFFUSION = 1 + STABLE_CASCADE = 2 -def load_clip(ckpt_paths, embedding_directory=None): +def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION): clip_data = [] for p in ckpt_paths: clip_data.append(comfy.utils.load_torch_file(p, safe_load=True)) @@ -311,14 +374,21 @@ def load_clip(ckpt_paths, embedding_directory=None): for i in range(len(clip_data)): if "transformer.resblocks.0.ln_1.weight" in clip_data[i]: - clip_data[i] = comfy.utils.transformers_convert(clip_data[i], "", "text_model.", 32) + clip_data[i] = comfy.utils.clip_text_transformers_convert(clip_data[i], "", "") + else: + if "text_projection" in clip_data[i]: + clip_data[i]["text_projection.weight"] = clip_data[i]["text_projection"].transpose(0, 1) #old models saved with the CLIPSave node clip_target = EmptyClass() clip_target.params = {} if len(clip_data) == 1: if "text_model.encoder.layers.30.mlp.fc1.weight" in clip_data[0]: - clip_target.clip = sdxl_clip.SDXLRefinerClipModel - clip_target.tokenizer = sdxl_clip.SDXLTokenizer + if clip_type == CLIPType.STABLE_CASCADE: + clip_target.clip = sdxl_clip.StableCascadeClipModel + clip_target.tokenizer = sdxl_clip.StableCascadeTokenizer + else: + clip_target.clip = sdxl_clip.SDXLRefinerClipModel + clip_target.tokenizer = sdxl_clip.SDXLTokenizer elif "text_model.encoder.layers.22.mlp.fc1.weight" in clip_data[0]: clip_target.clip = sd2_clip.SD2ClipModel clip_target.tokenizer = sd2_clip.SD2Tokenizer @@ -333,10 +403,10 @@ def load_clip(ckpt_paths, embedding_directory=None): for c in clip_data: m, u = clip.load_sd(c) if len(m) > 0: - print("clip missing:", m) + logging.warning("clip missing: {}".format(m)) if len(u) > 0: - print("clip unexpected:", u) + logging.debug("clip unexpected: {}".format(u)) return clip def load_gligen(ckpt_path): @@ -347,6 +417,8 @@ def load_gligen(ckpt_path): return comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device()) def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_clip=True, embedding_directory=None, state_dict=None, config=None): + logging.warning("Warning: The load checkpoint with config function is deprecated and will eventually be removed, please use the other one.") + model, clip, vae, _ = load_checkpoint_guess_config(ckpt_path, output_vae=output_vae, output_clip=output_clip, output_clipvision=False, embedding_directory=embedding_directory, output_model=True) #TODO: this function is a mess and should be removed eventually if config is None: with open(config_path, 'r') as stream: @@ -354,81 +426,20 @@ def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_cl model_config_params = config['model']['params'] clip_config = model_config_params['cond_stage_config'] scale_factor = model_config_params['scale_factor'] - vae_config = model_config_params['first_stage_config'] - - fp16 = False - if "unet_config" in model_config_params: - if "params" in model_config_params["unet_config"]: - unet_config = model_config_params["unet_config"]["params"] - if "use_fp16" in unet_config: - fp16 = unet_config.pop("use_fp16") - if fp16: - unet_config["dtype"] = torch.float16 - - noise_aug_config = None - if "noise_aug_config" in model_config_params: - noise_aug_config = model_config_params["noise_aug_config"] - - model_type = model_base.ModelType.EPS if "parameterization" in model_config_params: if model_config_params["parameterization"] == "v": - model_type = model_base.ModelType.V_PREDICTION + m = model.clone() + class ModelSamplingAdvanced(comfy.model_sampling.ModelSamplingDiscrete, comfy.model_sampling.V_PREDICTION): + pass + m.add_object_patch("model_sampling", ModelSamplingAdvanced(model.model.model_config)) + model = m - clip = None - vae = None + layer_idx = clip_config.get("params", {}).get("layer_idx", None) + if layer_idx is not None: + clip.clip_layer(layer_idx) - class WeightsLoader(torch.nn.Module): - pass - - if state_dict is None: - state_dict = comfy.utils.load_torch_file(ckpt_path) - - class EmptyClass: - pass - - model_config = comfy.supported_models_base.BASE({}) - - from . import latent_formats - model_config.latent_format = latent_formats.SD15(scale_factor=scale_factor) - model_config.unet_config = model_detection.convert_config(unet_config) - - if config['model']["target"].endswith("ImageEmbeddingConditionedLatentDiffusion"): - model = model_base.SD21UNCLIP(model_config, noise_aug_config["params"], model_type=model_type) - else: - model = model_base.BaseModel(model_config, model_type=model_type) - - if config['model']["target"].endswith("LatentInpaintDiffusion"): - model.set_inpaint() - - if fp16: - model = model.half() - - offload_device = model_management.unet_offload_device() - model = model.to(offload_device) - model.load_model_weights(state_dict, "model.diffusion_model.") - - if output_vae: - vae_sd = comfy.utils.state_dict_prefix_replace(state_dict, {"first_stage_model.": ""}, filter_keys=True) - vae = VAE(sd=vae_sd, config=vae_config) - - if output_clip: - w = WeightsLoader() - clip_target = EmptyClass() - clip_target.params = clip_config.get("params", {}) - if clip_config["target"].endswith("FrozenOpenCLIPEmbedder"): - clip_target.clip = sd2_clip.SD2ClipModel - clip_target.tokenizer = sd2_clip.SD2Tokenizer - clip = CLIP(clip_target, embedding_directory=embedding_directory) - w.cond_stage_model = clip.cond_stage_model.clip_h - elif clip_config["target"].endswith("FrozenCLIPEmbedder"): - clip_target.clip = sd1_clip.SD1ClipModel - clip_target.tokenizer = sd1_clip.SD1Tokenizer - clip = CLIP(clip_target, embedding_directory=embedding_directory) - w.cond_stage_model = clip.cond_stage_model.clip_l - load_clip_weights(w, state_dict) - - return (comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device), clip, vae) + return (model, clip, vae) def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True): sd = comfy.utils.load_torch_file(ckpt_path) @@ -441,15 +452,12 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o clip_target = None parameters = comfy.utils.calculate_parameters(sd, "model.diffusion_model.") - unet_dtype = model_management.unet_dtype(model_params=parameters) load_device = model_management.get_torch_device() - manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device) - class WeightsLoader(torch.nn.Module): - pass - - model_config = model_detection.model_config_from_unet(sd, "model.diffusion_model.", unet_dtype) - model_config.set_manual_cast(manual_cast_dtype) + model_config = model_detection.model_config_from_unet(sd, "model.diffusion_model.") + unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=model_config.supported_inference_dtypes) + manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) + model_config.set_inference_dtype(unet_dtype, manual_cast_dtype) if model_config is None: raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path)) @@ -465,27 +473,37 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o model.load_model_weights(sd, "model.diffusion_model.") if output_vae: - vae_sd = comfy.utils.state_dict_prefix_replace(sd, {"first_stage_model.": ""}, filter_keys=True) + vae_sd = comfy.utils.state_dict_prefix_replace(sd, {k: "" for k in model_config.vae_key_prefix}, filter_keys=True) vae_sd = model_config.process_vae_state_dict(vae_sd) vae = VAE(sd=vae_sd) if output_clip: - w = WeightsLoader() clip_target = model_config.clip_target() if clip_target is not None: - clip = CLIP(clip_target, embedding_directory=embedding_directory) - w.cond_stage_model = clip.cond_stage_model - sd = model_config.process_clip_state_dict(sd) - load_model_weights(w, sd) + clip_sd = model_config.process_clip_state_dict(sd) + if len(clip_sd) > 0: + clip = CLIP(clip_target, embedding_directory=embedding_directory) + m, u = clip.load_sd(clip_sd, full_model=True) + if len(m) > 0: + m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m)) + if len(m_filter) > 0: + logging.warning("clip missing: {}".format(m)) + else: + logging.debug("clip missing: {}".format(m)) + + if len(u) > 0: + logging.debug("clip unexpected {}:".format(u)) + else: + logging.warning("no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.") left_over = sd.keys() if len(left_over) > 0: - print("left over keys:", left_over) + logging.debug("left over keys: {}".format(left_over)) if output_model: model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device(), current_device=inital_load_device) if inital_load_device != torch.device("cpu"): - print("loaded straight to GPU") + logging.info("loaded straight to GPU") model_management.load_model_gpu(model_patcher) return (model_patcher, clip, vae, clipvision) @@ -495,16 +513,15 @@ def load_unet_state_dict(sd): #load unet in diffusers format parameters = comfy.utils.calculate_parameters(sd) unet_dtype = model_management.unet_dtype(model_params=parameters) load_device = model_management.get_torch_device() - manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device) - if "input_blocks.0.0.weight" in sd: #ldm - model_config = model_detection.model_config_from_unet(sd, "", unet_dtype) + if "input_blocks.0.0.weight" in sd or 'clf.1.weight' in sd: #ldm or stable cascade + model_config = model_detection.model_config_from_unet(sd, "") if model_config is None: return None new_sd = sd else: #diffusers - model_config = model_detection.model_config_from_diffusers_unet(sd, unet_dtype) + model_config = model_detection.model_config_from_diffusers_unet(sd) if model_config is None: return None @@ -515,26 +532,39 @@ def load_unet_state_dict(sd): #load unet in diffusers format if k in sd: new_sd[diffusers_keys[k]] = sd.pop(k) else: - print(diffusers_keys[k], k) + logging.warning("{} {}".format(diffusers_keys[k], k)) + offload_device = model_management.unet_offload_device() - model_config.set_manual_cast(manual_cast_dtype) + unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=model_config.supported_inference_dtypes) + manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) + model_config.set_inference_dtype(unet_dtype, manual_cast_dtype) model = model_config.get_model(new_sd, "") model = model.to(offload_device) model.load_model_weights(new_sd, "") left_over = sd.keys() if len(left_over) > 0: - print("left over keys in unet:", left_over) + logging.info("left over keys in unet: {}".format(left_over)) return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device) def load_unet(unet_path): sd = comfy.utils.load_torch_file(unet_path) model = load_unet_state_dict(sd) if model is None: - print("ERROR UNSUPPORTED UNET", unet_path) + logging.error("ERROR UNSUPPORTED UNET {}".format(unet_path)) raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path)) return model -def save_checkpoint(output_path, model, clip, vae, metadata=None): - model_management.load_models_gpu([model, clip.load_model()]) - sd = model.model.state_dict_for_saving(clip.get_sd(), vae.get_sd()) +def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, metadata=None, extra_keys={}): + clip_sd = None + load_models = [model] + if clip is not None: + load_models.append(clip.load_model()) + clip_sd = clip.get_sd() + + model_management.load_models_gpu(load_models, force_patch_weights=True) + clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None + sd = model.model.state_dict_for_saving(clip_sd, vae.get_sd(), clip_vision_sd) + for k in extra_keys: + sd[k] = extra_keys[k] + comfy.utils.save_torch_file(sd, output_path, metadata=metadata) diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index 6ffef515e..ff6db0d20 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -6,9 +6,9 @@ import torch import traceback import zipfile from . import model_management -import contextlib import comfy.clip_model import json +import logging def gen_empty_tokens(special_tokens, length): start_token = special_tokens.get("start", None) @@ -68,7 +68,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): ] def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_length=77, freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, dtype=None, model_class=comfy.clip_model.CLIPTextModel, - special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True): # clip-vit-base-patch32 + special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False, return_projected_pooled=True): # clip-vit-base-patch32 super().__init__() assert layer in self.LAYERS @@ -87,16 +87,18 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): self.layer = layer self.layer_idx = None self.special_tokens = special_tokens - self.text_projection = torch.nn.Parameter(torch.eye(self.transformer.get_input_embeddings().weight.shape[1])) + self.logit_scale = torch.nn.Parameter(torch.tensor(4.6055)) - self.enable_attention_masks = False + self.enable_attention_masks = enable_attention_masks self.layer_norm_hidden_state = layer_norm_hidden_state + self.return_projected_pooled = return_projected_pooled + if layer == "hidden": assert layer_idx is not None assert abs(layer_idx) < self.num_layers - self.clip_layer(layer_idx) - self.layer_default = (self.layer, self.layer_idx) + self.set_clip_options({"layer": layer_idx}) + self.options_default = (self.layer, self.layer_idx, self.return_projected_pooled) def freeze(self): self.transformer = self.transformer.eval() @@ -104,16 +106,19 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): for param in self.parameters(): param.requires_grad = False - def clip_layer(self, layer_idx): - if abs(layer_idx) > self.num_layers: + def set_clip_options(self, options): + layer_idx = options.get("layer", self.layer_idx) + self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled) + if layer_idx is None or abs(layer_idx) > self.num_layers: self.layer = "last" else: self.layer = "hidden" self.layer_idx = layer_idx - def reset_clip_layer(self): - self.layer = self.layer_default[0] - self.layer_idx = self.layer_default[1] + def reset_clip_options(self): + self.layer = self.options_default[0] + self.layer_idx = self.options_default[1] + self.return_projected_pooled = self.options_default[2] def set_up_textual_embeddings(self, tokens, current_embeds): out_tokens = [] @@ -133,7 +138,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): tokens_temp += [next_new_token] next_new_token += 1 else: - print("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored", y.shape[0], current_embeds.weight.shape[1]) + logging.warning("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored {} != {}".format(y.shape[0], current_embeds.weight.shape[1])) while len(tokens_temp) < len(x): tokens_temp += [self.special_tokens["pad"]] out_tokens += [tokens_temp] @@ -178,23 +183,19 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): else: z = outputs[1] - if outputs[2] is not None: - pooled_output = outputs[2].float() - else: - pooled_output = None + pooled_output = None + if len(outputs) >= 3: + if not self.return_projected_pooled and len(outputs) >= 4 and outputs[3] is not None: + pooled_output = outputs[3].float() + elif outputs[2] is not None: + pooled_output = outputs[2].float() - if self.text_projection is not None and pooled_output is not None: - pooled_output = pooled_output.float().to(self.text_projection.device) @ self.text_projection.float() return z.float(), pooled_output def encode(self, tokens): return self(tokens) def load_sd(self, sd): - if "text_projection" in sd: - self.text_projection[:] = sd.pop("text_projection") - if "text_projection.weight" in sd: - self.text_projection[:] = sd.pop("text_projection.weight").transpose(0, 1) return self.transformer.load_state_dict(sd, strict=False) def parse_parentheses(string): @@ -329,9 +330,7 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No else: embed = torch.load(embed_path, map_location="cpu") except Exception as e: - print(traceback.format_exc()) - print() - print("error loading embedding, skipping loading:", embedding_name) + logging.warning("{}\n\nerror loading embedding, skipping loading: {}".format(traceback.format_exc(), embedding_name)) return None if embed_out is None: @@ -355,11 +354,12 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No return embed_out class SDTokenizer: - def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, pad_to_max_length=True): + def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, pad_to_max_length=True, min_length=None): if tokenizer_path is None: tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer") self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path) self.max_length = max_length + self.min_length = min_length empty = self.tokenizer('')["input_ids"] if has_start_token: @@ -421,7 +421,7 @@ class SDTokenizer: embedding_name = word[len(self.embedding_identifier):].strip('\n') embed, leftover = self._try_get_embedding(embedding_name) if embed is None: - print(f"warning, embedding:{embedding_name} does not exist, ignoring") + logging.warning(f"warning, embedding:{embedding_name} does not exist, ignoring") else: if len(embed.shape) == 1: tokens.append([(embed, weight)]) @@ -471,6 +471,8 @@ class SDTokenizer: batch.append((self.end_token, 1.0, 0)) if self.pad_to_max_length: batch.extend([(pad_token, 1.0, 0)] * (self.max_length - len(batch))) + if self.min_length is not None and len(batch) < self.min_length: + batch.extend([(pad_token, 1.0, 0)] * (self.min_length - len(batch))) if not return_word_ids: batched_tokens = [[(t, w) for t, w,_ in x] for x in batched_tokens] @@ -504,11 +506,11 @@ class SD1ClipModel(torch.nn.Module): self.clip = "clip_{}".format(self.clip_name) setattr(self, self.clip, clip_model(device=device, dtype=dtype, **kwargs)) - def clip_layer(self, layer_idx): - getattr(self, self.clip).clip_layer(layer_idx) + def set_clip_options(self, options): + getattr(self, self.clip).set_clip_options(options) - def reset_clip_layer(self): - getattr(self, self.clip).reset_clip_layer() + def reset_clip_options(self): + getattr(self, self.clip).reset_clip_options() def encode_token_weights(self, token_weight_pairs): token_weight_pairs = token_weight_pairs[self.clip_name] diff --git a/comfy/sd2_clip.py b/comfy/sd2_clip.py index 9c878d54a..d14b44544 100644 --- a/comfy/sd2_clip.py +++ b/comfy/sd2_clip.py @@ -1,5 +1,4 @@ from comfy import sd1_clip -import torch import os class SD2ClipHModel(sd1_clip.SDClipModel): diff --git a/comfy/sdxl_clip.py b/comfy/sdxl_clip.py index b35056bb9..e62d1ed86 100644 --- a/comfy/sdxl_clip.py +++ b/comfy/sdxl_clip.py @@ -40,13 +40,13 @@ class SDXLClipModel(torch.nn.Module): self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False) self.clip_g = SDXLClipG(device=device, dtype=dtype) - def clip_layer(self, layer_idx): - self.clip_l.clip_layer(layer_idx) - self.clip_g.clip_layer(layer_idx) + def set_clip_options(self, options): + self.clip_l.set_clip_options(options) + self.clip_g.set_clip_options(options) - def reset_clip_layer(self): - self.clip_g.reset_clip_layer() - self.clip_l.reset_clip_layer() + def reset_clip_options(self): + self.clip_g.reset_clip_options() + self.clip_l.reset_clip_options() def encode_token_weights(self, token_weight_pairs): token_weight_pairs_g = token_weight_pairs["g"] @@ -64,3 +64,25 @@ class SDXLClipModel(torch.nn.Module): class SDXLRefinerClipModel(sd1_clip.SD1ClipModel): def __init__(self, device="cpu", dtype=None): super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=SDXLClipG) + + +class StableCascadeClipGTokenizer(sd1_clip.SDTokenizer): + def __init__(self, tokenizer_path=None, embedding_directory=None): + super().__init__(tokenizer_path, pad_with_end=True, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g') + +class StableCascadeTokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None): + super().__init__(embedding_directory=embedding_directory, clip_name="g", tokenizer=StableCascadeClipGTokenizer) + +class StableCascadeClipG(sd1_clip.SDClipModel): + def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None): + textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json") + super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, + special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True) + + def load_sd(self, sd): + return super().load_sd(sd) + +class StableCascadeClipModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None): + super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=StableCascadeClipG) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 1d442d4dd..6ca32e8ee 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -40,11 +40,16 @@ class SD15(supported_models_base.BASE): state_dict['cond_stage_model.transformer.text_model.embeddings.position_ids'] = ids.round() replace_prefix = {} - replace_prefix["cond_stage_model."] = "cond_stage_model.clip_l." - state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix) + replace_prefix["cond_stage_model."] = "clip_l." + state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=True) return state_dict def process_clip_state_dict_for_saving(self, state_dict): + pop_keys = ["clip_l.transformer.text_projection.weight", "clip_l.logit_scale"] + for p in pop_keys: + if p in state_dict: + state_dict.pop(p) + replace_prefix = {"clip_l.": "cond_stage_model."} return utils.state_dict_prefix_replace(state_dict, replace_prefix) @@ -60,22 +65,28 @@ class SD20(supported_models_base.BASE): "use_temporal_attention": False, } + unet_extra_config = { + "num_heads": -1, + "num_head_channels": 64, + "attn_precision": torch.float32, + } + latent_format = latent_formats.SD15 def model_type(self, state_dict, prefix=""): if self.unet_config["in_channels"] == 4: #SD2.0 inpainting models are not v prediction k = "{}output_blocks.11.1.transformer_blocks.0.norm1.bias".format(prefix) - out = state_dict[k] - if torch.std(out, unbiased=False) > 0.09: # not sure how well this will actually work. I guess we will find out. + out = state_dict.get(k, None) + if out is not None and torch.std(out, unbiased=False) > 0.09: # not sure how well this will actually work. I guess we will find out. return model_base.ModelType.V_PREDICTION return model_base.ModelType.EPS def process_clip_state_dict(self, state_dict): replace_prefix = {} - replace_prefix["conditioner.embedders.0.model."] = "cond_stage_model.model." #SD2 in sgm format - state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix) - - state_dict = utils.transformers_convert(state_dict, "cond_stage_model.model.", "cond_stage_model.clip_h.transformer.text_model.", 24) + replace_prefix["conditioner.embedders.0.model."] = "clip_h." #SD2 in sgm format + replace_prefix["cond_stage_model.model."] = "clip_h." + state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=True) + state_dict = utils.clip_text_transformers_convert(state_dict, "clip_h.", "clip_h.transformer.") return state_dict def process_clip_state_dict_for_saving(self, state_dict): @@ -131,11 +142,10 @@ class SDXLRefiner(supported_models_base.BASE): def process_clip_state_dict(self, state_dict): keys_to_replace = {} replace_prefix = {} + replace_prefix["conditioner.embedders.0.model."] = "clip_g." + state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=True) - state_dict = utils.transformers_convert(state_dict, "conditioner.embedders.0.model.", "cond_stage_model.clip_g.transformer.text_model.", 32) - keys_to_replace["conditioner.embedders.0.model.text_projection"] = "cond_stage_model.clip_g.text_projection" - keys_to_replace["conditioner.embedders.0.model.logit_scale"] = "cond_stage_model.clip_g.logit_scale" - + state_dict = utils.clip_text_transformers_convert(state_dict, "clip_g.", "clip_g.transformer.") state_dict = utils.state_dict_key_replace(state_dict, keys_to_replace) return state_dict @@ -164,7 +174,18 @@ class SDXL(supported_models_base.BASE): latent_format = latent_formats.SDXL def model_type(self, state_dict, prefix=""): - if "v_pred" in state_dict: + if 'edm_mean' in state_dict and 'edm_std' in state_dict: #Playground V2.5 + self.latent_format = latent_formats.SDXL_Playground_2_5() + self.sampling_settings["sigma_data"] = 0.5 + self.sampling_settings["sigma_max"] = 80.0 + self.sampling_settings["sigma_min"] = 0.002 + return model_base.ModelType.EDM + elif "edm_vpred.sigma_max" in state_dict: + self.sampling_settings["sigma_max"] = float(state_dict["edm_vpred.sigma_max"].item()) + if "edm_vpred.sigma_min" in state_dict: + self.sampling_settings["sigma_min"] = float(state_dict["edm_vpred.sigma_min"].item()) + return model_base.ModelType.V_PREDICTION_EDM + elif "v_pred" in state_dict: return model_base.ModelType.V_PREDICTION else: return model_base.ModelType.EPS @@ -179,26 +200,28 @@ class SDXL(supported_models_base.BASE): keys_to_replace = {} replace_prefix = {} - replace_prefix["conditioner.embedders.0.transformer.text_model"] = "cond_stage_model.clip_l.transformer.text_model" - state_dict = utils.transformers_convert(state_dict, "conditioner.embedders.1.model.", "cond_stage_model.clip_g.transformer.text_model.", 32) - keys_to_replace["conditioner.embedders.1.model.text_projection"] = "cond_stage_model.clip_g.text_projection" - keys_to_replace["conditioner.embedders.1.model.text_projection.weight"] = "cond_stage_model.clip_g.text_projection" - keys_to_replace["conditioner.embedders.1.model.logit_scale"] = "cond_stage_model.clip_g.logit_scale" + replace_prefix["conditioner.embedders.0.transformer.text_model"] = "clip_l.transformer.text_model" + replace_prefix["conditioner.embedders.1.model."] = "clip_g." + state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=True) - state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix) state_dict = utils.state_dict_key_replace(state_dict, keys_to_replace) + state_dict = utils.clip_text_transformers_convert(state_dict, "clip_g.", "clip_g.transformer.") return state_dict def process_clip_state_dict_for_saving(self, state_dict): replace_prefix = {} keys_to_replace = {} state_dict_g = diffusers_convert.convert_text_enc_state_dict_v20(state_dict, "clip_g") - if "clip_g.transformer.text_model.embeddings.position_ids" in state_dict_g: - state_dict_g.pop("clip_g.transformer.text_model.embeddings.position_ids") for k in state_dict: if k.startswith("clip_l"): state_dict_g[k] = state_dict[k] + state_dict_g["clip_l.transformer.text_model.embeddings.position_ids"] = torch.arange(77).expand((1, -1)) + pop_keys = ["clip_l.transformer.text_projection.weight", "clip_l.logit_scale"] + for p in pop_keys: + if p in state_dict_g: + state_dict_g.pop(p) + replace_prefix["clip_g"] = "conditioner.embedders.1.model" replace_prefix["clip_l"] = "conditioner.embedders.0" state_dict_g = utils.state_dict_prefix_replace(state_dict_g, replace_prefix) @@ -227,6 +250,26 @@ class Segmind_Vega(SDXL): "use_temporal_attention": False, } +class KOALA_700M(SDXL): + unet_config = { + "model_channels": 320, + "use_linear_in_transformer": True, + "transformer_depth": [0, 2, 5], + "context_dim": 2048, + "adm_in_channels": 2816, + "use_temporal_attention": False, + } + +class KOALA_1B(SDXL): + unet_config = { + "model_channels": 320, + "use_linear_in_transformer": True, + "transformer_depth": [0, 2, 6], + "context_dim": 2048, + "adm_in_channels": 2816, + "use_temporal_attention": False, + } + class SVD_img2vid(supported_models_base.BASE): unet_config = { "model_channels": 320, @@ -239,6 +282,12 @@ class SVD_img2vid(supported_models_base.BASE): "use_temporal_resblock": True } + unet_extra_config = { + "num_heads": -1, + "num_head_channels": 64, + "attn_precision": torch.float32, + } + clip_vision_prefix = "conditioner.embedders.0.open_clip.model.visual." latent_format = latent_formats.SD15 @@ -252,6 +301,41 @@ class SVD_img2vid(supported_models_base.BASE): def clip_target(self): return None +class SV3D_u(SVD_img2vid): + unet_config = { + "model_channels": 320, + "in_channels": 8, + "use_linear_in_transformer": True, + "transformer_depth": [1, 1, 1, 1, 1, 1, 0, 0], + "context_dim": 1024, + "adm_in_channels": 256, + "use_temporal_attention": True, + "use_temporal_resblock": True + } + + vae_key_prefix = ["conditioner.embedders.1.encoder."] + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.SV3D_u(self, device=device) + return out + +class SV3D_p(SV3D_u): + unet_config = { + "model_channels": 320, + "in_channels": 8, + "use_linear_in_transformer": True, + "transformer_depth": [1, 1, 1, 1, 1, 1, 0, 0], + "context_dim": 1024, + "adm_in_channels": 1280, + "use_temporal_attention": True, + "use_temporal_resblock": True + } + + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.SV3D_p(self, device=device) + return out + class Stable_Zero123(supported_models_base.BASE): unet_config = { "context_dim": 768, @@ -267,6 +351,11 @@ class Stable_Zero123(supported_models_base.BASE): "num_head_channels": -1, } + required_keys = { + "cc_projection.weight": None, + "cc_projection.bias": None, + } + clip_vision_prefix = "cond_stage_model.model.visual." latent_format = latent_formats.SD15 @@ -306,5 +395,99 @@ class SD_X4Upscaler(SD20): out = model_base.SD_X4Upscaler(self, device=device) return out -models = [Stable_Zero123, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXLRefiner, SDXL, SSD1B, Segmind_Vega, SD_X4Upscaler] +class Stable_Cascade_C(supported_models_base.BASE): + unet_config = { + "stable_cascade_stage": 'c', + } + + unet_extra_config = {} + + latent_format = latent_formats.SC_Prior + supported_inference_dtypes = [torch.bfloat16, torch.float32] + + sampling_settings = { + "shift": 2.0, + } + + vae_key_prefix = ["vae."] + text_encoder_key_prefix = ["text_encoder."] + clip_vision_prefix = "clip_l_vision." + + def process_unet_state_dict(self, state_dict): + key_list = list(state_dict.keys()) + for y in ["weight", "bias"]: + suffix = "in_proj_{}".format(y) + keys = filter(lambda a: a.endswith(suffix), key_list) + for k_from in keys: + weights = state_dict.pop(k_from) + prefix = k_from[:-(len(suffix) + 1)] + shape_from = weights.shape[0] // 3 + for x in range(3): + p = ["to_q", "to_k", "to_v"] + k_to = "{}.{}.{}".format(prefix, p[x], y) + state_dict[k_to] = weights[shape_from*x:shape_from*(x + 1)] + return state_dict + + def process_clip_state_dict(self, state_dict): + state_dict = utils.state_dict_prefix_replace(state_dict, {k: "" for k in self.text_encoder_key_prefix}, filter_keys=True) + if "clip_g.text_projection" in state_dict: + state_dict["clip_g.transformer.text_projection.weight"] = state_dict.pop("clip_g.text_projection").transpose(0, 1) + return state_dict + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.StableCascade_C(self, device=device) + return out + + def clip_target(self): + return supported_models_base.ClipTarget(sdxl_clip.StableCascadeTokenizer, sdxl_clip.StableCascadeClipModel) + +class Stable_Cascade_B(Stable_Cascade_C): + unet_config = { + "stable_cascade_stage": 'b', + } + + unet_extra_config = {} + + latent_format = latent_formats.SC_B + supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32] + + sampling_settings = { + "shift": 1.0, + } + + clip_vision_prefix = None + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.StableCascade_B(self, device=device) + return out + +class SD15_instructpix2pix(SD15): + unet_config = { + "context_dim": 768, + "model_channels": 320, + "use_linear_in_transformer": False, + "adm_in_channels": None, + "use_temporal_attention": False, + "in_channels": 8, + } + + def get_model(self, state_dict, prefix="", device=None): + return model_base.SD15_instructpix2pix(self, device=device) + +class SDXL_instructpix2pix(SDXL): + unet_config = { + "model_channels": 320, + "use_linear_in_transformer": True, + "transformer_depth": [0, 0, 2, 2, 10, 10], + "context_dim": 2048, + "adm_in_channels": 2816, + "use_temporal_attention": False, + "in_channels": 8, + } + + def get_model(self, state_dict, prefix="", device=None): + return model_base.SDXL_instructpix2pix(self, model_type=self.model_type(state_dict, prefix), device=device) + +models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p] + models += [SVD_img2vid] diff --git a/comfy/supported_models_base.py b/comfy/supported_models_base.py index 49087d23e..cf7cdff34 100644 --- a/comfy/supported_models_base.py +++ b/comfy/supported_models_base.py @@ -16,19 +16,28 @@ class BASE: "num_head_channels": 64, } + required_keys = {} + clip_prefix = [] clip_vision_prefix = None noise_aug_config = None sampling_settings = {} latent_format = latent_formats.LatentFormat + vae_key_prefix = ["first_stage_model."] + text_encoder_key_prefix = ["cond_stage_model."] + supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32] manual_cast_dtype = None @classmethod - def matches(s, unet_config): + def matches(s, unet_config, state_dict=None): for k in s.unet_config: - if s.unet_config[k] != unet_config[k]: + if k not in unet_config or s.unet_config[k] != unet_config[k]: return False + if state_dict is not None: + for k in s.required_keys: + if k not in state_dict: + return False return True def model_type(self, state_dict, prefix=""): @@ -38,7 +47,8 @@ class BASE: return self.unet_config["in_channels"] > 4 def __init__(self, unet_config): - self.unet_config = unet_config + self.unet_config = unet_config.copy() + self.sampling_settings = self.sampling_settings.copy() self.latent_format = self.latent_format() for x in self.unet_extra_config: self.unet_config[x] = self.unet_extra_config[x] @@ -53,6 +63,7 @@ class BASE: return out def process_clip_state_dict(self, state_dict): + state_dict = utils.state_dict_prefix_replace(state_dict, {k: "" for k in self.text_encoder_key_prefix}, filter_keys=True) return state_dict def process_unet_state_dict(self, state_dict): @@ -62,7 +73,13 @@ class BASE: return state_dict def process_clip_state_dict_for_saving(self, state_dict): - replace_prefix = {"": "cond_stage_model."} + replace_prefix = {"": self.text_encoder_key_prefix[0]} + return utils.state_dict_prefix_replace(state_dict, replace_prefix) + + def process_clip_vision_state_dict_for_saving(self, state_dict): + replace_prefix = {} + if self.clip_vision_prefix is not None: + replace_prefix[""] = self.clip_vision_prefix return utils.state_dict_prefix_replace(state_dict, replace_prefix) def process_unet_state_dict_for_saving(self, state_dict): @@ -70,8 +87,9 @@ class BASE: return utils.state_dict_prefix_replace(state_dict, replace_prefix) def process_vae_state_dict_for_saving(self, state_dict): - replace_prefix = {"": "first_stage_model."} + replace_prefix = {"": self.vae_key_prefix[0]} return utils.state_dict_prefix_replace(state_dict, replace_prefix) - def set_manual_cast(self, manual_cast_dtype): + def set_inference_dtype(self, dtype, manual_cast_dtype): + self.unet_config['dtype'] = dtype self.manual_cast_dtype = manual_cast_dtype diff --git a/comfy/types.py b/comfy/types.py new file mode 100644 index 000000000..70cf4b158 --- /dev/null +++ b/comfy/types.py @@ -0,0 +1,32 @@ +import torch +from typing import Callable, Protocol, TypedDict, Optional, List + + +class UnetApplyFunction(Protocol): + """Function signature protocol on comfy.model_base.BaseModel.apply_model""" + + def __call__(self, x: torch.Tensor, t: torch.Tensor, **kwargs) -> torch.Tensor: + pass + + +class UnetApplyConds(TypedDict): + """Optional conditions for unet apply function.""" + + c_concat: Optional[torch.Tensor] + c_crossattn: Optional[torch.Tensor] + control: Optional[torch.Tensor] + transformer_options: Optional[dict] + + +class UnetParams(TypedDict): + # Tensor of shape [B, C, H, W] + input: torch.Tensor + # Tensor of shape [B] + timestep: torch.Tensor + c: UnetApplyConds + # List of [0, 1], [0], [1], ... + # 0 means conditional, 1 means conditional unconditional + cond_or_uncond: List[int] + + +UnetWrapperFunction = Callable[[UnetApplyFunction, UnetParams], torch.Tensor] diff --git a/comfy/utils.py b/comfy/utils.py index f8026ddab..ab47b8f28 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -5,6 +5,7 @@ import comfy.checkpoint_pickle import safetensors.torch import numpy as np from PIL import Image +import logging def load_torch_file(ckpt, safe_load=False, device=None): if device is None: @@ -14,14 +15,14 @@ def load_torch_file(ckpt, safe_load=False, device=None): else: if safe_load: if not 'weights_only' in torch.load.__code__.co_varnames: - print("Warning torch.load doesn't support weights_only on this pytorch version, loading unsafely.") + logging.warning("Warning torch.load doesn't support weights_only on this pytorch version, loading unsafely.") safe_load = False if safe_load: pl_sd = torch.load(ckpt, map_location=device, weights_only=True) else: pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle) if "global_step" in pl_sd: - print(f"Global Step: {pl_sd['global_step']}") + logging.debug(f"Global Step: {pl_sd['global_step']}") if "state_dict" in pl_sd: sd = pl_sd["state_dict"] else: @@ -98,8 +99,22 @@ def transformers_convert(sd, prefix_from, prefix_to, number): p = ["self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj"] k_to = "{}encoder.layers.{}.{}.{}".format(prefix_to, resblock, p[x], y) sd[k_to] = weights[shape_from*x:shape_from*(x + 1)] + return sd +def clip_text_transformers_convert(sd, prefix_from, prefix_to): + sd = transformers_convert(sd, prefix_from, "{}text_model.".format(prefix_to), 32) + + tp = "{}text_projection.weight".format(prefix_from) + if tp in sd: + sd["{}text_projection.weight".format(prefix_to)] = sd.pop(tp) + + tp = "{}text_projection".format(prefix_from) + if tp in sd: + sd["{}text_projection.weight".format(prefix_to)] = sd.pop(tp).transpose(0, 1).contiguous() + return sd + + UNET_MAP_ATTENTIONS = { "proj_in.weight", "proj_in.bias", @@ -169,6 +184,8 @@ UNET_MAP_BASIC = { } def unet_to_diffusers(unet_config): + if "num_res_blocks" not in unet_config: + return {} num_res_blocks = unet_config["num_res_blocks"] channel_mult = unet_config["channel_mult"] transformer_depth = unet_config["transformer_depth"][:] @@ -278,8 +295,11 @@ def set_attr(obj, attr, value): for name in attrs[:-1]: obj = getattr(obj, name) prev = getattr(obj, attrs[-1]) - setattr(obj, attrs[-1], torch.nn.Parameter(value, requires_grad=False)) - del prev + setattr(obj, attrs[-1], value) + return prev + +def set_attr_param(obj, attr, value): + return set_attr(obj, attr, torch.nn.Parameter(value, requires_grad=False)) def copy_to_param(obj, attr, value): # inplace update tensor instead of replacing it @@ -413,6 +433,8 @@ def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_am out_div = torch.zeros((s.shape[0], out_channels, round(s.shape[2] * upscale_amount), round(s.shape[3] * upscale_amount)), device=output_device) for y in range(0, s.shape[2], tile_y - overlap): for x in range(0, s.shape[3], tile_x - overlap): + x = max(0, min(s.shape[-1] - overlap, x)) + y = max(0, min(s.shape[-2] - overlap, y)) s_in = s[:,:,y:y+tile_y,x:x+tile_x] ps = function(s_in).to(output_device) diff --git a/comfy_extras/chainner_models/architecture/DAT.py b/comfy_extras/chainner_models/architecture/DAT.py deleted file mode 100644 index 0bcc26ef4..000000000 --- a/comfy_extras/chainner_models/architecture/DAT.py +++ /dev/null @@ -1,1182 +0,0 @@ -# pylint: skip-file -import math -import re - -import numpy as np -import torch -import torch.nn as nn -import torch.utils.checkpoint as checkpoint -from einops import rearrange -from einops.layers.torch import Rearrange -from torch import Tensor -from torch.nn import functional as F - -from .timm.drop import DropPath -from .timm.weight_init import trunc_normal_ - - -def img2windows(img, H_sp, W_sp): - """ - Input: Image (B, C, H, W) - Output: Window Partition (B', N, C) - """ - B, C, H, W = img.shape - img_reshape = img.view(B, C, H // H_sp, H_sp, W // W_sp, W_sp) - img_perm = ( - img_reshape.permute(0, 2, 4, 3, 5, 1).contiguous().reshape(-1, H_sp * W_sp, C) - ) - return img_perm - - -def windows2img(img_splits_hw, H_sp, W_sp, H, W): - """ - Input: Window Partition (B', N, C) - Output: Image (B, H, W, C) - """ - B = int(img_splits_hw.shape[0] / (H * W / H_sp / W_sp)) - - img = img_splits_hw.view(B, H // H_sp, W // W_sp, H_sp, W_sp, -1) - img = img.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) - return img - - -class SpatialGate(nn.Module): - """Spatial-Gate. - Args: - dim (int): Half of input channels. - """ - - def __init__(self, dim): - super().__init__() - self.norm = nn.LayerNorm(dim) - self.conv = nn.Conv2d( - dim, dim, kernel_size=3, stride=1, padding=1, groups=dim - ) # DW Conv - - def forward(self, x, H, W): - # Split - x1, x2 = x.chunk(2, dim=-1) - B, N, C = x.shape - x2 = ( - self.conv(self.norm(x2).transpose(1, 2).contiguous().view(B, C // 2, H, W)) - .flatten(2) - .transpose(-1, -2) - .contiguous() - ) - - return x1 * x2 - - -class SGFN(nn.Module): - """Spatial-Gate Feed-Forward Network. - Args: - in_features (int): Number of input channels. - hidden_features (int | None): Number of hidden channels. Default: None - out_features (int | None): Number of output channels. Default: None - act_layer (nn.Module): Activation layer. Default: nn.GELU - drop (float): Dropout rate. Default: 0.0 - """ - - def __init__( - self, - in_features, - hidden_features=None, - out_features=None, - act_layer=nn.GELU, - drop=0.0, - ): - super().__init__() - out_features = out_features or in_features - hidden_features = hidden_features or in_features - self.fc1 = nn.Linear(in_features, hidden_features) - self.act = act_layer() - self.sg = SpatialGate(hidden_features // 2) - self.fc2 = nn.Linear(hidden_features // 2, out_features) - self.drop = nn.Dropout(drop) - - def forward(self, x, H, W): - """ - Input: x: (B, H*W, C), H, W - Output: x: (B, H*W, C) - """ - x = self.fc1(x) - x = self.act(x) - x = self.drop(x) - - x = self.sg(x, H, W) - x = self.drop(x) - - x = self.fc2(x) - x = self.drop(x) - return x - - -class DynamicPosBias(nn.Module): - # The implementation builds on Crossformer code https://github.com/cheerss/CrossFormer/blob/main/models/crossformer.py - """Dynamic Relative Position Bias. - Args: - dim (int): Number of input channels. - num_heads (int): Number of attention heads. - residual (bool): If True, use residual strage to connect conv. - """ - - def __init__(self, dim, num_heads, residual): - super().__init__() - self.residual = residual - self.num_heads = num_heads - self.pos_dim = dim // 4 - self.pos_proj = nn.Linear(2, self.pos_dim) - self.pos1 = nn.Sequential( - nn.LayerNorm(self.pos_dim), - nn.ReLU(inplace=True), - nn.Linear(self.pos_dim, self.pos_dim), - ) - self.pos2 = nn.Sequential( - nn.LayerNorm(self.pos_dim), - nn.ReLU(inplace=True), - nn.Linear(self.pos_dim, self.pos_dim), - ) - self.pos3 = nn.Sequential( - nn.LayerNorm(self.pos_dim), - nn.ReLU(inplace=True), - nn.Linear(self.pos_dim, self.num_heads), - ) - - def forward(self, biases): - if self.residual: - pos = self.pos_proj(biases) # 2Gh-1 * 2Gw-1, heads - pos = pos + self.pos1(pos) - pos = pos + self.pos2(pos) - pos = self.pos3(pos) - else: - pos = self.pos3(self.pos2(self.pos1(self.pos_proj(biases)))) - return pos - - -class Spatial_Attention(nn.Module): - """Spatial Window Self-Attention. - It supports rectangle window (containing square window). - Args: - dim (int): Number of input channels. - idx (int): The indentix of window. (0/1) - split_size (tuple(int)): Height and Width of spatial window. - dim_out (int | None): The dimension of the attention output. Default: None - num_heads (int): Number of attention heads. Default: 6 - attn_drop (float): Dropout ratio of attention weight. Default: 0.0 - proj_drop (float): Dropout ratio of output. Default: 0.0 - qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set - position_bias (bool): The dynamic relative position bias. Default: True - """ - - def __init__( - self, - dim, - idx, - split_size=[8, 8], - dim_out=None, - num_heads=6, - attn_drop=0.0, - proj_drop=0.0, - qk_scale=None, - position_bias=True, - ): - super().__init__() - self.dim = dim - self.dim_out = dim_out or dim - self.split_size = split_size - self.num_heads = num_heads - self.idx = idx - self.position_bias = position_bias - - head_dim = dim // num_heads - self.scale = qk_scale or head_dim**-0.5 - - if idx == 0: - H_sp, W_sp = self.split_size[0], self.split_size[1] - elif idx == 1: - W_sp, H_sp = self.split_size[0], self.split_size[1] - else: - print("ERROR MODE", idx) - exit(0) - self.H_sp = H_sp - self.W_sp = W_sp - - if self.position_bias: - self.pos = DynamicPosBias(self.dim // 4, self.num_heads, residual=False) - # generate mother-set - position_bias_h = torch.arange(1 - self.H_sp, self.H_sp) - position_bias_w = torch.arange(1 - self.W_sp, self.W_sp) - biases = torch.stack(torch.meshgrid([position_bias_h, position_bias_w])) - biases = biases.flatten(1).transpose(0, 1).contiguous().float() - self.register_buffer("rpe_biases", biases) - - # get pair-wise relative position index for each token inside the window - coords_h = torch.arange(self.H_sp) - coords_w = torch.arange(self.W_sp) - coords = torch.stack(torch.meshgrid([coords_h, coords_w])) - coords_flatten = torch.flatten(coords, 1) - relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] - relative_coords = relative_coords.permute(1, 2, 0).contiguous() - relative_coords[:, :, 0] += self.H_sp - 1 - relative_coords[:, :, 1] += self.W_sp - 1 - relative_coords[:, :, 0] *= 2 * self.W_sp - 1 - relative_position_index = relative_coords.sum(-1) - self.register_buffer("relative_position_index", relative_position_index) - - self.attn_drop = nn.Dropout(attn_drop) - - def im2win(self, x, H, W): - B, N, C = x.shape - x = x.transpose(-2, -1).contiguous().view(B, C, H, W) - x = img2windows(x, self.H_sp, self.W_sp) - x = ( - x.reshape(-1, self.H_sp * self.W_sp, self.num_heads, C // self.num_heads) - .permute(0, 2, 1, 3) - .contiguous() - ) - return x - - def forward(self, qkv, H, W, mask=None): - """ - Input: qkv: (B, 3*L, C), H, W, mask: (B, N, N), N is the window size - Output: x (B, H, W, C) - """ - q, k, v = qkv[0], qkv[1], qkv[2] - - B, L, C = q.shape - assert L == H * W, "flatten img_tokens has wrong size" - - # partition the q,k,v, image to window - q = self.im2win(q, H, W) - k = self.im2win(k, H, W) - v = self.im2win(v, H, W) - - q = q * self.scale - attn = q @ k.transpose(-2, -1) # B head N C @ B head C N --> B head N N - - # calculate drpe - if self.position_bias: - pos = self.pos(self.rpe_biases) - # select position bias - relative_position_bias = pos[self.relative_position_index.view(-1)].view( - self.H_sp * self.W_sp, self.H_sp * self.W_sp, -1 - ) - relative_position_bias = relative_position_bias.permute( - 2, 0, 1 - ).contiguous() - attn = attn + relative_position_bias.unsqueeze(0) - - N = attn.shape[3] - - # use mask for shift window - if mask is not None: - nW = mask.shape[0] - attn = attn.view(B, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze( - 0 - ) - attn = attn.view(-1, self.num_heads, N, N) - - attn = nn.functional.softmax(attn, dim=-1, dtype=attn.dtype) - attn = self.attn_drop(attn) - - x = attn @ v - x = x.transpose(1, 2).reshape( - -1, self.H_sp * self.W_sp, C - ) # B head N N @ B head N C - - # merge the window, window to image - x = windows2img(x, self.H_sp, self.W_sp, H, W) # B H' W' C - - return x - - -class Adaptive_Spatial_Attention(nn.Module): - # The implementation builds on CAT code https://github.com/Zhengchen1999/CAT - """Adaptive Spatial Self-Attention - Args: - dim (int): Number of input channels. - num_heads (int): Number of attention heads. Default: 6 - split_size (tuple(int)): Height and Width of spatial window. - shift_size (tuple(int)): Shift size for spatial window. - qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set. - drop (float): Dropout rate. Default: 0.0 - attn_drop (float): Attention dropout rate. Default: 0.0 - rg_idx (int): The indentix of Residual Group (RG) - b_idx (int): The indentix of Block in each RG - """ - - def __init__( - self, - dim, - num_heads, - reso=64, - split_size=[8, 8], - shift_size=[1, 2], - qkv_bias=False, - qk_scale=None, - drop=0.0, - attn_drop=0.0, - rg_idx=0, - b_idx=0, - ): - super().__init__() - self.dim = dim - self.num_heads = num_heads - self.split_size = split_size - self.shift_size = shift_size - self.b_idx = b_idx - self.rg_idx = rg_idx - self.patches_resolution = reso - self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) - - assert ( - 0 <= self.shift_size[0] < self.split_size[0] - ), "shift_size must in 0-split_size0" - assert ( - 0 <= self.shift_size[1] < self.split_size[1] - ), "shift_size must in 0-split_size1" - - self.branch_num = 2 - - self.proj = nn.Linear(dim, dim) - self.proj_drop = nn.Dropout(drop) - - self.attns = nn.ModuleList( - [ - Spatial_Attention( - dim // 2, - idx=i, - split_size=split_size, - num_heads=num_heads // 2, - dim_out=dim // 2, - qk_scale=qk_scale, - attn_drop=attn_drop, - proj_drop=drop, - position_bias=True, - ) - for i in range(self.branch_num) - ] - ) - - if (self.rg_idx % 2 == 0 and self.b_idx > 0 and (self.b_idx - 2) % 4 == 0) or ( - self.rg_idx % 2 != 0 and self.b_idx % 4 == 0 - ): - attn_mask = self.calculate_mask( - self.patches_resolution, self.patches_resolution - ) - self.register_buffer("attn_mask_0", attn_mask[0]) - self.register_buffer("attn_mask_1", attn_mask[1]) - else: - attn_mask = None - self.register_buffer("attn_mask_0", None) - self.register_buffer("attn_mask_1", None) - - self.dwconv = nn.Sequential( - nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, groups=dim), - nn.BatchNorm2d(dim), - nn.GELU(), - ) - self.channel_interaction = nn.Sequential( - nn.AdaptiveAvgPool2d(1), - nn.Conv2d(dim, dim // 8, kernel_size=1), - nn.BatchNorm2d(dim // 8), - nn.GELU(), - nn.Conv2d(dim // 8, dim, kernel_size=1), - ) - self.spatial_interaction = nn.Sequential( - nn.Conv2d(dim, dim // 16, kernel_size=1), - nn.BatchNorm2d(dim // 16), - nn.GELU(), - nn.Conv2d(dim // 16, 1, kernel_size=1), - ) - - def calculate_mask(self, H, W): - # The implementation builds on Swin Transformer code https://github.com/microsoft/Swin-Transformer/blob/main/models/swin_transformer.py - # calculate attention mask for shift window - img_mask_0 = torch.zeros((1, H, W, 1)) # 1 H W 1 idx=0 - img_mask_1 = torch.zeros((1, H, W, 1)) # 1 H W 1 idx=1 - h_slices_0 = ( - slice(0, -self.split_size[0]), - slice(-self.split_size[0], -self.shift_size[0]), - slice(-self.shift_size[0], None), - ) - w_slices_0 = ( - slice(0, -self.split_size[1]), - slice(-self.split_size[1], -self.shift_size[1]), - slice(-self.shift_size[1], None), - ) - - h_slices_1 = ( - slice(0, -self.split_size[1]), - slice(-self.split_size[1], -self.shift_size[1]), - slice(-self.shift_size[1], None), - ) - w_slices_1 = ( - slice(0, -self.split_size[0]), - slice(-self.split_size[0], -self.shift_size[0]), - slice(-self.shift_size[0], None), - ) - cnt = 0 - for h in h_slices_0: - for w in w_slices_0: - img_mask_0[:, h, w, :] = cnt - cnt += 1 - cnt = 0 - for h in h_slices_1: - for w in w_slices_1: - img_mask_1[:, h, w, :] = cnt - cnt += 1 - - # calculate mask for window-0 - img_mask_0 = img_mask_0.view( - 1, - H // self.split_size[0], - self.split_size[0], - W // self.split_size[1], - self.split_size[1], - 1, - ) - img_mask_0 = ( - img_mask_0.permute(0, 1, 3, 2, 4, 5) - .contiguous() - .view(-1, self.split_size[0], self.split_size[1], 1) - ) # nW, sw[0], sw[1], 1 - mask_windows_0 = img_mask_0.view(-1, self.split_size[0] * self.split_size[1]) - attn_mask_0 = mask_windows_0.unsqueeze(1) - mask_windows_0.unsqueeze(2) - attn_mask_0 = attn_mask_0.masked_fill( - attn_mask_0 != 0, float(-100.0) - ).masked_fill(attn_mask_0 == 0, float(0.0)) - - # calculate mask for window-1 - img_mask_1 = img_mask_1.view( - 1, - H // self.split_size[1], - self.split_size[1], - W // self.split_size[0], - self.split_size[0], - 1, - ) - img_mask_1 = ( - img_mask_1.permute(0, 1, 3, 2, 4, 5) - .contiguous() - .view(-1, self.split_size[1], self.split_size[0], 1) - ) # nW, sw[1], sw[0], 1 - mask_windows_1 = img_mask_1.view(-1, self.split_size[1] * self.split_size[0]) - attn_mask_1 = mask_windows_1.unsqueeze(1) - mask_windows_1.unsqueeze(2) - attn_mask_1 = attn_mask_1.masked_fill( - attn_mask_1 != 0, float(-100.0) - ).masked_fill(attn_mask_1 == 0, float(0.0)) - - return attn_mask_0, attn_mask_1 - - def forward(self, x, H, W): - """ - Input: x: (B, H*W, C), H, W - Output: x: (B, H*W, C) - """ - B, L, C = x.shape - assert L == H * W, "flatten img_tokens has wrong size" - - qkv = self.qkv(x).reshape(B, -1, 3, C).permute(2, 0, 1, 3) # 3, B, HW, C - # V without partition - v = qkv[2].transpose(-2, -1).contiguous().view(B, C, H, W) - - # image padding - max_split_size = max(self.split_size[0], self.split_size[1]) - pad_l = pad_t = 0 - pad_r = (max_split_size - W % max_split_size) % max_split_size - pad_b = (max_split_size - H % max_split_size) % max_split_size - - qkv = qkv.reshape(3 * B, H, W, C).permute(0, 3, 1, 2) # 3B C H W - qkv = ( - F.pad(qkv, (pad_l, pad_r, pad_t, pad_b)) - .reshape(3, B, C, -1) - .transpose(-2, -1) - ) # l r t b - _H = pad_b + H - _W = pad_r + W - _L = _H * _W - - # window-0 and window-1 on split channels [C/2, C/2]; for square windows (e.g., 8x8), window-0 and window-1 can be merged - # shift in block: (0, 4, 8, ...), (2, 6, 10, ...), (0, 4, 8, ...), (2, 6, 10, ...), ... - if (self.rg_idx % 2 == 0 and self.b_idx > 0 and (self.b_idx - 2) % 4 == 0) or ( - self.rg_idx % 2 != 0 and self.b_idx % 4 == 0 - ): - qkv = qkv.view(3, B, _H, _W, C) - qkv_0 = torch.roll( - qkv[:, :, :, :, : C // 2], - shifts=(-self.shift_size[0], -self.shift_size[1]), - dims=(2, 3), - ) - qkv_0 = qkv_0.view(3, B, _L, C // 2) - qkv_1 = torch.roll( - qkv[:, :, :, :, C // 2 :], - shifts=(-self.shift_size[1], -self.shift_size[0]), - dims=(2, 3), - ) - qkv_1 = qkv_1.view(3, B, _L, C // 2) - - if self.patches_resolution != _H or self.patches_resolution != _W: - mask_tmp = self.calculate_mask(_H, _W) - x1_shift = self.attns[0](qkv_0, _H, _W, mask=mask_tmp[0].to(x.device)) - x2_shift = self.attns[1](qkv_1, _H, _W, mask=mask_tmp[1].to(x.device)) - else: - x1_shift = self.attns[0](qkv_0, _H, _W, mask=self.attn_mask_0) - x2_shift = self.attns[1](qkv_1, _H, _W, mask=self.attn_mask_1) - - x1 = torch.roll( - x1_shift, shifts=(self.shift_size[0], self.shift_size[1]), dims=(1, 2) - ) - x2 = torch.roll( - x2_shift, shifts=(self.shift_size[1], self.shift_size[0]), dims=(1, 2) - ) - x1 = x1[:, :H, :W, :].reshape(B, L, C // 2) - x2 = x2[:, :H, :W, :].reshape(B, L, C // 2) - # attention output - attened_x = torch.cat([x1, x2], dim=2) - - else: - x1 = self.attns[0](qkv[:, :, :, : C // 2], _H, _W)[:, :H, :W, :].reshape( - B, L, C // 2 - ) - x2 = self.attns[1](qkv[:, :, :, C // 2 :], _H, _W)[:, :H, :W, :].reshape( - B, L, C // 2 - ) - # attention output - attened_x = torch.cat([x1, x2], dim=2) - - # convolution output - conv_x = self.dwconv(v) - - # Adaptive Interaction Module (AIM) - # C-Map (before sigmoid) - channel_map = ( - self.channel_interaction(conv_x) - .permute(0, 2, 3, 1) - .contiguous() - .view(B, 1, C) - ) - # S-Map (before sigmoid) - attention_reshape = attened_x.transpose(-2, -1).contiguous().view(B, C, H, W) - spatial_map = self.spatial_interaction(attention_reshape) - - # C-I - attened_x = attened_x * torch.sigmoid(channel_map) - # S-I - conv_x = torch.sigmoid(spatial_map) * conv_x - conv_x = conv_x.permute(0, 2, 3, 1).contiguous().view(B, L, C) - - x = attened_x + conv_x - - x = self.proj(x) - x = self.proj_drop(x) - - return x - - -class Adaptive_Channel_Attention(nn.Module): - # The implementation builds on XCiT code https://github.com/facebookresearch/xcit - """Adaptive Channel Self-Attention - Args: - dim (int): Number of input channels. - num_heads (int): Number of attention heads. Default: 6 - qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set. - attn_drop (float): Attention dropout rate. Default: 0.0 - drop_path (float): Stochastic depth rate. Default: 0.0 - """ - - def __init__( - self, - dim, - num_heads=8, - qkv_bias=False, - qk_scale=None, - attn_drop=0.0, - proj_drop=0.0, - ): - super().__init__() - self.num_heads = num_heads - self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1)) - - self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) - self.attn_drop = nn.Dropout(attn_drop) - self.proj = nn.Linear(dim, dim) - self.proj_drop = nn.Dropout(proj_drop) - - self.dwconv = nn.Sequential( - nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, groups=dim), - nn.BatchNorm2d(dim), - nn.GELU(), - ) - self.channel_interaction = nn.Sequential( - nn.AdaptiveAvgPool2d(1), - nn.Conv2d(dim, dim // 8, kernel_size=1), - nn.BatchNorm2d(dim // 8), - nn.GELU(), - nn.Conv2d(dim // 8, dim, kernel_size=1), - ) - self.spatial_interaction = nn.Sequential( - nn.Conv2d(dim, dim // 16, kernel_size=1), - nn.BatchNorm2d(dim // 16), - nn.GELU(), - nn.Conv2d(dim // 16, 1, kernel_size=1), - ) - - def forward(self, x, H, W): - """ - Input: x: (B, H*W, C), H, W - Output: x: (B, H*W, C) - """ - B, N, C = x.shape - qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads) - qkv = qkv.permute(2, 0, 3, 1, 4) - q, k, v = qkv[0], qkv[1], qkv[2] - - q = q.transpose(-2, -1) - k = k.transpose(-2, -1) - v = v.transpose(-2, -1) - - v_ = v.reshape(B, C, N).contiguous().view(B, C, H, W) - - q = torch.nn.functional.normalize(q, dim=-1) - k = torch.nn.functional.normalize(k, dim=-1) - - attn = (q @ k.transpose(-2, -1)) * self.temperature - attn = attn.softmax(dim=-1) - attn = self.attn_drop(attn) - - # attention output - attened_x = (attn @ v).permute(0, 3, 1, 2).reshape(B, N, C) - - # convolution output - conv_x = self.dwconv(v_) - - # Adaptive Interaction Module (AIM) - # C-Map (before sigmoid) - attention_reshape = attened_x.transpose(-2, -1).contiguous().view(B, C, H, W) - channel_map = self.channel_interaction(attention_reshape) - # S-Map (before sigmoid) - spatial_map = ( - self.spatial_interaction(conv_x) - .permute(0, 2, 3, 1) - .contiguous() - .view(B, N, 1) - ) - - # S-I - attened_x = attened_x * torch.sigmoid(spatial_map) - # C-I - conv_x = conv_x * torch.sigmoid(channel_map) - conv_x = conv_x.permute(0, 2, 3, 1).contiguous().view(B, N, C) - - x = attened_x + conv_x - - x = self.proj(x) - x = self.proj_drop(x) - - return x - - -class DATB(nn.Module): - def __init__( - self, - dim, - num_heads, - reso=64, - split_size=[2, 4], - shift_size=[1, 2], - expansion_factor=4.0, - qkv_bias=False, - qk_scale=None, - drop=0.0, - attn_drop=0.0, - drop_path=0.0, - act_layer=nn.GELU, - norm_layer=nn.LayerNorm, - rg_idx=0, - b_idx=0, - ): - super().__init__() - - self.norm1 = norm_layer(dim) - - if b_idx % 2 == 0: - # DSTB - self.attn = Adaptive_Spatial_Attention( - dim, - num_heads=num_heads, - reso=reso, - split_size=split_size, - shift_size=shift_size, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop, - attn_drop=attn_drop, - rg_idx=rg_idx, - b_idx=b_idx, - ) - else: - # DCTB - self.attn = Adaptive_Channel_Attention( - dim, - num_heads=num_heads, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - attn_drop=attn_drop, - proj_drop=drop, - ) - self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() - - ffn_hidden_dim = int(dim * expansion_factor) - self.ffn = SGFN( - in_features=dim, - hidden_features=ffn_hidden_dim, - out_features=dim, - act_layer=act_layer, - ) - self.norm2 = norm_layer(dim) - - def forward(self, x, x_size): - """ - Input: x: (B, H*W, C), x_size: (H, W) - Output: x: (B, H*W, C) - """ - H, W = x_size - x = x + self.drop_path(self.attn(self.norm1(x), H, W)) - x = x + self.drop_path(self.ffn(self.norm2(x), H, W)) - - return x - - -class ResidualGroup(nn.Module): - """ResidualGroup - Args: - dim (int): Number of input channels. - reso (int): Input resolution. - num_heads (int): Number of attention heads. - split_size (tuple(int)): Height and Width of spatial window. - expansion_factor (float): Ratio of ffn hidden dim to embedding dim. - qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set. Default: None - drop (float): Dropout rate. Default: 0 - attn_drop(float): Attention dropout rate. Default: 0 - drop_paths (float | None): Stochastic depth rate. - act_layer (nn.Module): Activation layer. Default: nn.GELU - norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm - depth (int): Number of dual aggregation Transformer blocks in residual group. - use_chk (bool): Whether to use checkpointing to save memory. - resi_connection: The convolutional block before residual connection. '1conv'/'3conv' - """ - - def __init__( - self, - dim, - reso, - num_heads, - split_size=[2, 4], - expansion_factor=4.0, - qkv_bias=False, - qk_scale=None, - drop=0.0, - attn_drop=0.0, - drop_paths=None, - act_layer=nn.GELU, - norm_layer=nn.LayerNorm, - depth=2, - use_chk=False, - resi_connection="1conv", - rg_idx=0, - ): - super().__init__() - self.use_chk = use_chk - self.reso = reso - - self.blocks = nn.ModuleList( - [ - DATB( - dim=dim, - num_heads=num_heads, - reso=reso, - split_size=split_size, - shift_size=[split_size[0] // 2, split_size[1] // 2], - expansion_factor=expansion_factor, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop, - attn_drop=attn_drop, - drop_path=drop_paths[i], - act_layer=act_layer, - norm_layer=norm_layer, - rg_idx=rg_idx, - b_idx=i, - ) - for i in range(depth) - ] - ) - - if resi_connection == "1conv": - self.conv = nn.Conv2d(dim, dim, 3, 1, 1) - elif resi_connection == "3conv": - self.conv = nn.Sequential( - nn.Conv2d(dim, dim // 4, 3, 1, 1), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(dim // 4, dim // 4, 1, 1, 0), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(dim // 4, dim, 3, 1, 1), - ) - - def forward(self, x, x_size): - """ - Input: x: (B, H*W, C), x_size: (H, W) - Output: x: (B, H*W, C) - """ - H, W = x_size - res = x - for blk in self.blocks: - if self.use_chk: - x = checkpoint.checkpoint(blk, x, x_size) - else: - x = blk(x, x_size) - x = rearrange(x, "b (h w) c -> b c h w", h=H, w=W) - x = self.conv(x) - x = rearrange(x, "b c h w -> b (h w) c") - x = res + x - - return x - - -class Upsample(nn.Sequential): - """Upsample module. - Args: - scale (int): Scale factor. Supported scales: 2^n and 3. - num_feat (int): Channel number of intermediate features. - """ - - def __init__(self, scale, num_feat): - m = [] - if (scale & (scale - 1)) == 0: # scale = 2^n - for _ in range(int(math.log(scale, 2))): - m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(2)) - elif scale == 3: - m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(3)) - else: - raise ValueError( - f"scale {scale} is not supported. " "Supported scales: 2^n and 3." - ) - super(Upsample, self).__init__(*m) - - -class UpsampleOneStep(nn.Sequential): - """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle) - Used in lightweight SR to save parameters. - - Args: - scale (int): Scale factor. Supported scales: 2^n and 3. - num_feat (int): Channel number of intermediate features. - - """ - - def __init__(self, scale, num_feat, num_out_ch, input_resolution=None): - self.num_feat = num_feat - self.input_resolution = input_resolution - m = [] - m.append(nn.Conv2d(num_feat, (scale**2) * num_out_ch, 3, 1, 1)) - m.append(nn.PixelShuffle(scale)) - super(UpsampleOneStep, self).__init__(*m) - - def flops(self): - h, w = self.input_resolution - flops = h * w * self.num_feat * 3 * 9 - return flops - - -class DAT(nn.Module): - """Dual Aggregation Transformer - Args: - img_size (int): Input image size. Default: 64 - in_chans (int): Number of input image channels. Default: 3 - embed_dim (int): Patch embedding dimension. Default: 180 - depths (tuple(int)): Depth of each residual group (number of DATB in each RG). - split_size (tuple(int)): Height and Width of spatial window. - num_heads (tuple(int)): Number of attention heads in different residual groups. - expansion_factor (float): Ratio of ffn hidden dim to embedding dim. Default: 4 - qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set. Default: None - drop_rate (float): Dropout rate. Default: 0 - attn_drop_rate (float): Attention dropout rate. Default: 0 - drop_path_rate (float): Stochastic depth rate. Default: 0.1 - act_layer (nn.Module): Activation layer. Default: nn.GELU - norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm - use_chk (bool): Whether to use checkpointing to save memory. - upscale: Upscale factor. 2/3/4 for image SR - img_range: Image range. 1. or 255. - resi_connection: The convolutional block before residual connection. '1conv'/'3conv' - """ - - def __init__(self, state_dict): - super().__init__() - - # defaults - img_size = 64 - in_chans = 3 - embed_dim = 180 - split_size = [2, 4] - depth = [2, 2, 2, 2] - num_heads = [2, 2, 2, 2] - expansion_factor = 4.0 - qkv_bias = True - qk_scale = None - drop_rate = 0.0 - attn_drop_rate = 0.0 - drop_path_rate = 0.1 - act_layer = nn.GELU - norm_layer = nn.LayerNorm - use_chk = False - upscale = 2 - img_range = 1.0 - resi_connection = "1conv" - upsampler = "pixelshuffle" - - self.model_arch = "DAT" - self.sub_type = "SR" - self.state = state_dict - - state_keys = state_dict.keys() - if "conv_before_upsample.0.weight" in state_keys: - if "conv_up1.weight" in state_keys: - upsampler = "nearest+conv" - else: - upsampler = "pixelshuffle" - supports_fp16 = False - elif "upsample.0.weight" in state_keys: - upsampler = "pixelshuffledirect" - else: - upsampler = "" - - num_feat = ( - state_dict.get("conv_before_upsample.0.weight", None).shape[1] - if state_dict.get("conv_before_upsample.weight", None) - else 64 - ) - - num_in_ch = state_dict["conv_first.weight"].shape[1] - in_chans = num_in_ch - if "conv_last.weight" in state_keys: - num_out_ch = state_dict["conv_last.weight"].shape[0] - else: - num_out_ch = num_in_ch - - upscale = 1 - if upsampler == "nearest+conv": - upsample_keys = [ - x for x in state_keys if "conv_up" in x and "bias" not in x - ] - - for upsample_key in upsample_keys: - upscale *= 2 - elif upsampler == "pixelshuffle": - upsample_keys = [ - x - for x in state_keys - if "upsample" in x and "conv" not in x and "bias" not in x - ] - for upsample_key in upsample_keys: - shape = state_dict[upsample_key].shape[0] - upscale *= math.sqrt(shape // num_feat) - upscale = int(upscale) - elif upsampler == "pixelshuffledirect": - upscale = int( - math.sqrt(state_dict["upsample.0.bias"].shape[0] // num_out_ch) - ) - - max_layer_num = 0 - max_block_num = 0 - for key in state_keys: - result = re.match(r"layers.(\d*).blocks.(\d*).norm1.weight", key) - if result: - layer_num, block_num = result.groups() - max_layer_num = max(max_layer_num, int(layer_num)) - max_block_num = max(max_block_num, int(block_num)) - - depth = [max_block_num + 1 for _ in range(max_layer_num + 1)] - - if "layers.0.blocks.1.attn.temperature" in state_keys: - num_heads_num = state_dict["layers.0.blocks.1.attn.temperature"].shape[0] - num_heads = [num_heads_num for _ in range(max_layer_num + 1)] - else: - num_heads = depth - - embed_dim = state_dict["conv_first.weight"].shape[0] - expansion_factor = float( - state_dict["layers.0.blocks.0.ffn.fc1.weight"].shape[0] / embed_dim - ) - - # TODO: could actually count the layers, but this should do - if "layers.0.conv.4.weight" in state_keys: - resi_connection = "3conv" - else: - resi_connection = "1conv" - - if "layers.0.blocks.2.attn.attn_mask_0" in state_keys: - attn_mask_0_x, attn_mask_0_y, attn_mask_0_z = state_dict[ - "layers.0.blocks.2.attn.attn_mask_0" - ].shape - - img_size = int(math.sqrt(attn_mask_0_x * attn_mask_0_y)) - - if "layers.0.blocks.0.attn.attns.0.rpe_biases" in state_keys: - split_sizes = ( - state_dict["layers.0.blocks.0.attn.attns.0.rpe_biases"][-1] + 1 - ) - split_size = [int(x) for x in split_sizes] - - self.in_nc = num_in_ch - self.out_nc = num_out_ch - self.num_feat = num_feat - self.embed_dim = embed_dim - self.num_heads = num_heads - self.depth = depth - self.scale = upscale - self.upsampler = upsampler - self.img_size = img_size - self.img_range = img_range - self.expansion_factor = expansion_factor - self.resi_connection = resi_connection - self.split_size = split_size - - self.supports_fp16 = False # Too much weirdness to support this at the moment - self.supports_bfp16 = True - self.min_size_restriction = 16 - - num_in_ch = in_chans - num_out_ch = in_chans - num_feat = 64 - self.img_range = img_range - if in_chans == 3: - rgb_mean = (0.4488, 0.4371, 0.4040) - self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1) - else: - self.mean = torch.zeros(1, 1, 1, 1) - self.upscale = upscale - self.upsampler = upsampler - - # ------------------------- 1, Shallow Feature Extraction ------------------------- # - self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1) - - # ------------------------- 2, Deep Feature Extraction ------------------------- # - self.num_layers = len(depth) - self.use_chk = use_chk - self.num_features = ( - self.embed_dim - ) = embed_dim # num_features for consistency with other models - heads = num_heads - - self.before_RG = nn.Sequential( - Rearrange("b c h w -> b (h w) c"), nn.LayerNorm(embed_dim) - ) - - curr_dim = embed_dim - dpr = [ - x.item() for x in torch.linspace(0, drop_path_rate, np.sum(depth)) - ] # stochastic depth decay rule - - self.layers = nn.ModuleList() - for i in range(self.num_layers): - layer = ResidualGroup( - dim=embed_dim, - num_heads=heads[i], - reso=img_size, - split_size=split_size, - expansion_factor=expansion_factor, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop_rate, - attn_drop=attn_drop_rate, - drop_paths=dpr[sum(depth[:i]) : sum(depth[: i + 1])], - act_layer=act_layer, - norm_layer=norm_layer, - depth=depth[i], - use_chk=use_chk, - resi_connection=resi_connection, - rg_idx=i, - ) - self.layers.append(layer) - - self.norm = norm_layer(curr_dim) - # build the last conv layer in deep feature extraction - if resi_connection == "1conv": - self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1) - elif resi_connection == "3conv": - # to save parameters and memory - self.conv_after_body = nn.Sequential( - nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1), - ) - - # ------------------------- 3, Reconstruction ------------------------- # - if self.upsampler == "pixelshuffle": - # for classical SR - self.conv_before_upsample = nn.Sequential( - nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.upsample = Upsample(upscale, num_feat) - self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - elif self.upsampler == "pixelshuffledirect": - # for lightweight SR (to save parameters) - self.upsample = UpsampleOneStep( - upscale, embed_dim, num_out_ch, (img_size, img_size) - ) - - self.apply(self._init_weights) - self.load_state_dict(state_dict, strict=True) - - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=0.02) - if isinstance(m, nn.Linear) and m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance( - m, (nn.LayerNorm, nn.BatchNorm2d, nn.GroupNorm, nn.InstanceNorm2d) - ): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) - - def forward_features(self, x): - _, _, H, W = x.shape - x_size = [H, W] - x = self.before_RG(x) - for layer in self.layers: - x = layer(x, x_size) - x = self.norm(x) - x = rearrange(x, "b (h w) c -> b c h w", h=H, w=W) - - return x - - def forward(self, x): - """ - Input: x: (B, C, H, W) - """ - self.mean = self.mean.type_as(x) - x = (x - self.mean) * self.img_range - - if self.upsampler == "pixelshuffle": - # for image SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.conv_before_upsample(x) - x = self.conv_last(self.upsample(x)) - elif self.upsampler == "pixelshuffledirect": - # for lightweight SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.upsample(x) - - x = x / self.img_range + self.mean - return x diff --git a/comfy_extras/chainner_models/architecture/HAT.py b/comfy_extras/chainner_models/architecture/HAT.py deleted file mode 100644 index 669474219..000000000 --- a/comfy_extras/chainner_models/architecture/HAT.py +++ /dev/null @@ -1,1277 +0,0 @@ -# pylint: skip-file -# HAT from https://github.com/XPixelGroup/HAT/blob/main/hat/archs/hat_arch.py -import math -import re - -import torch -import torch.nn as nn -import torch.nn.functional as F -from einops import rearrange - -from .timm.helpers import to_2tuple -from .timm.weight_init import trunc_normal_ - - -def drop_path(x, drop_prob: float = 0.0, training: bool = False): - """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). - From: https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py - """ - if drop_prob == 0.0 or not training: - return x - keep_prob = 1 - drop_prob - shape = (x.shape[0],) + (1,) * ( - x.ndim - 1 - ) # work with diff dim tensors, not just 2D ConvNets - random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device) - random_tensor.floor_() # binarize - output = x.div(keep_prob) * random_tensor - return output - - -class DropPath(nn.Module): - """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). - From: https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py - """ - - def __init__(self, drop_prob=None): - super(DropPath, self).__init__() - self.drop_prob = drop_prob - - def forward(self, x): - return drop_path(x, self.drop_prob, self.training) # type: ignore - - -class ChannelAttention(nn.Module): - """Channel attention used in RCAN. - Args: - num_feat (int): Channel number of intermediate features. - squeeze_factor (int): Channel squeeze factor. Default: 16. - """ - - def __init__(self, num_feat, squeeze_factor=16): - super(ChannelAttention, self).__init__() - self.attention = nn.Sequential( - nn.AdaptiveAvgPool2d(1), - nn.Conv2d(num_feat, num_feat // squeeze_factor, 1, padding=0), - nn.ReLU(inplace=True), - nn.Conv2d(num_feat // squeeze_factor, num_feat, 1, padding=0), - nn.Sigmoid(), - ) - - def forward(self, x): - y = self.attention(x) - return x * y - - -class CAB(nn.Module): - def __init__(self, num_feat, compress_ratio=3, squeeze_factor=30): - super(CAB, self).__init__() - - self.cab = nn.Sequential( - nn.Conv2d(num_feat, num_feat // compress_ratio, 3, 1, 1), - nn.GELU(), - nn.Conv2d(num_feat // compress_ratio, num_feat, 3, 1, 1), - ChannelAttention(num_feat, squeeze_factor), - ) - - def forward(self, x): - return self.cab(x) - - -class Mlp(nn.Module): - def __init__( - self, - in_features, - hidden_features=None, - out_features=None, - act_layer=nn.GELU, - drop=0.0, - ): - super().__init__() - out_features = out_features or in_features - hidden_features = hidden_features or in_features - self.fc1 = nn.Linear(in_features, hidden_features) - self.act = act_layer() - self.fc2 = nn.Linear(hidden_features, out_features) - self.drop = nn.Dropout(drop) - - def forward(self, x): - x = self.fc1(x) - x = self.act(x) - x = self.drop(x) - x = self.fc2(x) - x = self.drop(x) - return x - - -def window_partition(x, window_size): - """ - Args: - x: (b, h, w, c) - window_size (int): window size - Returns: - windows: (num_windows*b, window_size, window_size, c) - """ - b, h, w, c = x.shape - x = x.view(b, h // window_size, window_size, w // window_size, window_size, c) - windows = ( - x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, c) - ) - return windows - - -def window_reverse(windows, window_size, h, w): - """ - Args: - windows: (num_windows*b, window_size, window_size, c) - window_size (int): Window size - h (int): Height of image - w (int): Width of image - Returns: - x: (b, h, w, c) - """ - b = int(windows.shape[0] / (h * w / window_size / window_size)) - x = windows.view( - b, h // window_size, w // window_size, window_size, window_size, -1 - ) - x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(b, h, w, -1) - return x - - -class WindowAttention(nn.Module): - r"""Window based multi-head self attention (W-MSA) module with relative position bias. - It supports both of shifted and non-shifted window. - Args: - dim (int): Number of input channels. - window_size (tuple[int]): The height and width of the window. - num_heads (int): Number of attention heads. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set - attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 - proj_drop (float, optional): Dropout ratio of output. Default: 0.0 - """ - - def __init__( - self, - dim, - window_size, - num_heads, - qkv_bias=True, - qk_scale=None, - attn_drop=0.0, - proj_drop=0.0, - ): - super().__init__() - self.dim = dim - self.window_size = window_size # Wh, Ww - self.num_heads = num_heads - head_dim = dim // num_heads - self.scale = qk_scale or head_dim**-0.5 - - # define a parameter table of relative position bias - self.relative_position_bias_table = nn.Parameter( # type: ignore - torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads) - ) # 2*Wh-1 * 2*Ww-1, nH - - self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) - self.attn_drop = nn.Dropout(attn_drop) - self.proj = nn.Linear(dim, dim) - - self.proj_drop = nn.Dropout(proj_drop) - - trunc_normal_(self.relative_position_bias_table, std=0.02) - self.softmax = nn.Softmax(dim=-1) - - def forward(self, x, rpi, mask=None): - """ - Args: - x: input features with shape of (num_windows*b, n, c) - mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None - """ - b_, n, c = x.shape - qkv = ( - self.qkv(x) - .reshape(b_, n, 3, self.num_heads, c // self.num_heads) - .permute(2, 0, 3, 1, 4) - ) - q, k, v = ( - qkv[0], - qkv[1], - qkv[2], - ) # make torchscript happy (cannot use tensor as tuple) - - q = q * self.scale - attn = q @ k.transpose(-2, -1) - - relative_position_bias = self.relative_position_bias_table[rpi.view(-1)].view( - self.window_size[0] * self.window_size[1], - self.window_size[0] * self.window_size[1], - -1, - ) # Wh*Ww,Wh*Ww,nH - relative_position_bias = relative_position_bias.permute( - 2, 0, 1 - ).contiguous() # nH, Wh*Ww, Wh*Ww - attn = attn + relative_position_bias.unsqueeze(0) - - if mask is not None: - nw = mask.shape[0] - attn = attn.view(b_ // nw, nw, self.num_heads, n, n) + mask.unsqueeze( - 1 - ).unsqueeze(0) - attn = attn.view(-1, self.num_heads, n, n) - attn = self.softmax(attn) - else: - attn = self.softmax(attn) - - attn = self.attn_drop(attn) - - x = (attn @ v).transpose(1, 2).reshape(b_, n, c) - x = self.proj(x) - x = self.proj_drop(x) - return x - - -class HAB(nn.Module): - r"""Hybrid Attention Block. - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resolution. - num_heads (int): Number of attention heads. - window_size (int): Window size. - shift_size (int): Shift size for SW-MSA. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float, optional): Stochastic depth rate. Default: 0.0 - act_layer (nn.Module, optional): Activation layer. Default: nn.GELU - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - """ - - def __init__( - self, - dim, - input_resolution, - num_heads, - window_size=7, - shift_size=0, - compress_ratio=3, - squeeze_factor=30, - conv_scale=0.01, - mlp_ratio=4.0, - qkv_bias=True, - qk_scale=None, - drop=0.0, - attn_drop=0.0, - drop_path=0.0, - act_layer=nn.GELU, - norm_layer=nn.LayerNorm, - ): - super().__init__() - self.dim = dim - self.input_resolution = input_resolution - self.num_heads = num_heads - self.window_size = window_size - self.shift_size = shift_size - self.mlp_ratio = mlp_ratio - if min(self.input_resolution) <= self.window_size: - # if window size is larger than input resolution, we don't partition windows - self.shift_size = 0 - self.window_size = min(self.input_resolution) - assert ( - 0 <= self.shift_size < self.window_size - ), "shift_size must in 0-window_size" - - self.norm1 = norm_layer(dim) - self.attn = WindowAttention( - dim, - window_size=to_2tuple(self.window_size), - num_heads=num_heads, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - attn_drop=attn_drop, - proj_drop=drop, - ) - - self.conv_scale = conv_scale - self.conv_block = CAB( - num_feat=dim, compress_ratio=compress_ratio, squeeze_factor=squeeze_factor - ) - - self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() - self.norm2 = norm_layer(dim) - mlp_hidden_dim = int(dim * mlp_ratio) - self.mlp = Mlp( - in_features=dim, - hidden_features=mlp_hidden_dim, - act_layer=act_layer, - drop=drop, - ) - - def forward(self, x, x_size, rpi_sa, attn_mask): - h, w = x_size - b, _, c = x.shape - # assert seq_len == h * w, "input feature has wrong size" - - shortcut = x - x = self.norm1(x) - x = x.view(b, h, w, c) - - # Conv_X - conv_x = self.conv_block(x.permute(0, 3, 1, 2)) - conv_x = conv_x.permute(0, 2, 3, 1).contiguous().view(b, h * w, c) - - # cyclic shift - if self.shift_size > 0: - shifted_x = torch.roll( - x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2) - ) - attn_mask = attn_mask - else: - shifted_x = x - attn_mask = None - - # partition windows - x_windows = window_partition( - shifted_x, self.window_size - ) # nw*b, window_size, window_size, c - x_windows = x_windows.view( - -1, self.window_size * self.window_size, c - ) # nw*b, window_size*window_size, c - - # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size - attn_windows = self.attn(x_windows, rpi=rpi_sa, mask=attn_mask) - - # merge windows - attn_windows = attn_windows.view(-1, self.window_size, self.window_size, c) - shifted_x = window_reverse(attn_windows, self.window_size, h, w) # b h' w' c - - # reverse cyclic shift - if self.shift_size > 0: - attn_x = torch.roll( - shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2) - ) - else: - attn_x = shifted_x - attn_x = attn_x.view(b, h * w, c) - - # FFN - x = shortcut + self.drop_path(attn_x) + conv_x * self.conv_scale - x = x + self.drop_path(self.mlp(self.norm2(x))) - - return x - - -class PatchMerging(nn.Module): - r"""Patch Merging Layer. - Args: - input_resolution (tuple[int]): Resolution of input feature. - dim (int): Number of input channels. - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - """ - - def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm): - super().__init__() - self.input_resolution = input_resolution - self.dim = dim - self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) - self.norm = norm_layer(4 * dim) - - def forward(self, x): - """ - x: b, h*w, c - """ - h, w = self.input_resolution - b, seq_len, c = x.shape - assert seq_len == h * w, "input feature has wrong size" - assert h % 2 == 0 and w % 2 == 0, f"x size ({h}*{w}) are not even." - - x = x.view(b, h, w, c) - - x0 = x[:, 0::2, 0::2, :] # b h/2 w/2 c - x1 = x[:, 1::2, 0::2, :] # b h/2 w/2 c - x2 = x[:, 0::2, 1::2, :] # b h/2 w/2 c - x3 = x[:, 1::2, 1::2, :] # b h/2 w/2 c - x = torch.cat([x0, x1, x2, x3], -1) # b h/2 w/2 4*c - x = x.view(b, -1, 4 * c) # b h/2*w/2 4*c - - x = self.norm(x) - x = self.reduction(x) - - return x - - -class OCAB(nn.Module): - # overlapping cross-attention block - - def __init__( - self, - dim, - input_resolution, - window_size, - overlap_ratio, - num_heads, - qkv_bias=True, - qk_scale=None, - mlp_ratio=2, - norm_layer=nn.LayerNorm, - ): - super().__init__() - self.dim = dim - self.input_resolution = input_resolution - self.window_size = window_size - self.num_heads = num_heads - head_dim = dim // num_heads - self.scale = qk_scale or head_dim**-0.5 - self.overlap_win_size = int(window_size * overlap_ratio) + window_size - - self.norm1 = norm_layer(dim) - self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) - self.unfold = nn.Unfold( - kernel_size=(self.overlap_win_size, self.overlap_win_size), - stride=window_size, - padding=(self.overlap_win_size - window_size) // 2, - ) - - # define a parameter table of relative position bias - self.relative_position_bias_table = nn.Parameter( # type: ignore - torch.zeros( - (window_size + self.overlap_win_size - 1) - * (window_size + self.overlap_win_size - 1), - num_heads, - ) - ) # 2*Wh-1 * 2*Ww-1, nH - - trunc_normal_(self.relative_position_bias_table, std=0.02) - self.softmax = nn.Softmax(dim=-1) - - self.proj = nn.Linear(dim, dim) - - self.norm2 = norm_layer(dim) - mlp_hidden_dim = int(dim * mlp_ratio) - self.mlp = Mlp( - in_features=dim, hidden_features=mlp_hidden_dim, act_layer=nn.GELU - ) - - def forward(self, x, x_size, rpi): - h, w = x_size - b, _, c = x.shape - - shortcut = x - x = self.norm1(x) - x = x.view(b, h, w, c) - - qkv = self.qkv(x).reshape(b, h, w, 3, c).permute(3, 0, 4, 1, 2) # 3, b, c, h, w - q = qkv[0].permute(0, 2, 3, 1) # b, h, w, c - kv = torch.cat((qkv[1], qkv[2]), dim=1) # b, 2*c, h, w - - # partition windows - q_windows = window_partition( - q, self.window_size - ) # nw*b, window_size, window_size, c - q_windows = q_windows.view( - -1, self.window_size * self.window_size, c - ) # nw*b, window_size*window_size, c - - kv_windows = self.unfold(kv) # b, c*w*w, nw - kv_windows = rearrange( - kv_windows, - "b (nc ch owh oww) nw -> nc (b nw) (owh oww) ch", - nc=2, - ch=c, - owh=self.overlap_win_size, - oww=self.overlap_win_size, - ).contiguous() # 2, nw*b, ow*ow, c - # Do the above rearrangement without the rearrange function - # kv_windows = kv_windows.view( - # 2, b, self.overlap_win_size, self.overlap_win_size, c, -1 - # ) - # kv_windows = kv_windows.permute(0, 5, 1, 2, 3, 4).contiguous() - # kv_windows = kv_windows.view( - # 2, -1, self.overlap_win_size * self.overlap_win_size, c - # ) - - k_windows, v_windows = kv_windows[0], kv_windows[1] # nw*b, ow*ow, c - - b_, nq, _ = q_windows.shape - _, n, _ = k_windows.shape - d = self.dim // self.num_heads - q = q_windows.reshape(b_, nq, self.num_heads, d).permute( - 0, 2, 1, 3 - ) # nw*b, nH, nq, d - k = k_windows.reshape(b_, n, self.num_heads, d).permute( - 0, 2, 1, 3 - ) # nw*b, nH, n, d - v = v_windows.reshape(b_, n, self.num_heads, d).permute( - 0, 2, 1, 3 - ) # nw*b, nH, n, d - - q = q * self.scale - attn = q @ k.transpose(-2, -1) - - relative_position_bias = self.relative_position_bias_table[rpi.view(-1)].view( - self.window_size * self.window_size, - self.overlap_win_size * self.overlap_win_size, - -1, - ) # ws*ws, wse*wse, nH - relative_position_bias = relative_position_bias.permute( - 2, 0, 1 - ).contiguous() # nH, ws*ws, wse*wse - attn = attn + relative_position_bias.unsqueeze(0) - - attn = self.softmax(attn) - attn_windows = (attn @ v).transpose(1, 2).reshape(b_, nq, self.dim) - - # merge windows - attn_windows = attn_windows.view( - -1, self.window_size, self.window_size, self.dim - ) - x = window_reverse(attn_windows, self.window_size, h, w) # b h w c - x = x.view(b, h * w, self.dim) - - x = self.proj(x) + shortcut - - x = x + self.mlp(self.norm2(x)) - return x - - -class AttenBlocks(nn.Module): - """A series of attention blocks for one RHAG. - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resolution. - depth (int): Number of blocks. - num_heads (int): Number of attention heads. - window_size (int): Local window size. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. - """ - - def __init__( - self, - dim, - input_resolution, - depth, - num_heads, - window_size, - compress_ratio, - squeeze_factor, - conv_scale, - overlap_ratio, - mlp_ratio=4.0, - qkv_bias=True, - qk_scale=None, - drop=0.0, - attn_drop=0.0, - drop_path=0.0, - norm_layer=nn.LayerNorm, - downsample=None, - use_checkpoint=False, - ): - super().__init__() - self.dim = dim - self.input_resolution = input_resolution - self.depth = depth - self.use_checkpoint = use_checkpoint - - # build blocks - self.blocks = nn.ModuleList( - [ - HAB( - dim=dim, - input_resolution=input_resolution, - num_heads=num_heads, - window_size=window_size, - shift_size=0 if (i % 2 == 0) else window_size // 2, - compress_ratio=compress_ratio, - squeeze_factor=squeeze_factor, - conv_scale=conv_scale, - mlp_ratio=mlp_ratio, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop, - attn_drop=attn_drop, - drop_path=drop_path[i] - if isinstance(drop_path, list) - else drop_path, - norm_layer=norm_layer, - ) - for i in range(depth) - ] - ) - - # OCAB - self.overlap_attn = OCAB( - dim=dim, - input_resolution=input_resolution, - window_size=window_size, - overlap_ratio=overlap_ratio, - num_heads=num_heads, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - mlp_ratio=mlp_ratio, # type: ignore - norm_layer=norm_layer, - ) - - # patch merging layer - if downsample is not None: - self.downsample = downsample( - input_resolution, dim=dim, norm_layer=norm_layer - ) - else: - self.downsample = None - - def forward(self, x, x_size, params): - for blk in self.blocks: - x = blk(x, x_size, params["rpi_sa"], params["attn_mask"]) - - x = self.overlap_attn(x, x_size, params["rpi_oca"]) - - if self.downsample is not None: - x = self.downsample(x) - return x - - -class RHAG(nn.Module): - """Residual Hybrid Attention Group (RHAG). - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resolution. - depth (int): Number of blocks. - num_heads (int): Number of attention heads. - window_size (int): Local window size. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. - img_size: Input image size. - patch_size: Patch size. - resi_connection: The convolutional block before residual connection. - """ - - def __init__( - self, - dim, - input_resolution, - depth, - num_heads, - window_size, - compress_ratio, - squeeze_factor, - conv_scale, - overlap_ratio, - mlp_ratio=4.0, - qkv_bias=True, - qk_scale=None, - drop=0.0, - attn_drop=0.0, - drop_path=0.0, - norm_layer=nn.LayerNorm, - downsample=None, - use_checkpoint=False, - img_size=224, - patch_size=4, - resi_connection="1conv", - ): - super(RHAG, self).__init__() - - self.dim = dim - self.input_resolution = input_resolution - - self.residual_group = AttenBlocks( - dim=dim, - input_resolution=input_resolution, - depth=depth, - num_heads=num_heads, - window_size=window_size, - compress_ratio=compress_ratio, - squeeze_factor=squeeze_factor, - conv_scale=conv_scale, - overlap_ratio=overlap_ratio, - mlp_ratio=mlp_ratio, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop, - attn_drop=attn_drop, - drop_path=drop_path, - norm_layer=norm_layer, - downsample=downsample, - use_checkpoint=use_checkpoint, - ) - - if resi_connection == "1conv": - self.conv = nn.Conv2d(dim, dim, 3, 1, 1) - elif resi_connection == "identity": - self.conv = nn.Identity() - - self.patch_embed = PatchEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=0, - embed_dim=dim, - norm_layer=None, - ) - - self.patch_unembed = PatchUnEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=0, - embed_dim=dim, - norm_layer=None, - ) - - def forward(self, x, x_size, params): - return ( - self.patch_embed( - self.conv( - self.patch_unembed(self.residual_group(x, x_size, params), x_size) - ) - ) - + x - ) - - -class PatchEmbed(nn.Module): - r"""Image to Patch Embedding - Args: - img_size (int): Image size. Default: 224. - patch_size (int): Patch token size. Default: 4. - in_chans (int): Number of input image channels. Default: 3. - embed_dim (int): Number of linear projection output channels. Default: 96. - norm_layer (nn.Module, optional): Normalization layer. Default: None - """ - - def __init__( - self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None - ): - super().__init__() - img_size = to_2tuple(img_size) - patch_size = to_2tuple(patch_size) - patches_resolution = [ - img_size[0] // patch_size[0], # type: ignore - img_size[1] // patch_size[1], # type: ignore - ] - self.img_size = img_size - self.patch_size = patch_size - self.patches_resolution = patches_resolution - self.num_patches = patches_resolution[0] * patches_resolution[1] - - self.in_chans = in_chans - self.embed_dim = embed_dim - - if norm_layer is not None: - self.norm = norm_layer(embed_dim) - else: - self.norm = None - - def forward(self, x): - x = x.flatten(2).transpose(1, 2) # b Ph*Pw c - if self.norm is not None: - x = self.norm(x) - return x - - -class PatchUnEmbed(nn.Module): - r"""Image to Patch Unembedding - Args: - img_size (int): Image size. Default: 224. - patch_size (int): Patch token size. Default: 4. - in_chans (int): Number of input image channels. Default: 3. - embed_dim (int): Number of linear projection output channels. Default: 96. - norm_layer (nn.Module, optional): Normalization layer. Default: None - """ - - def __init__( - self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None - ): - super().__init__() - img_size = to_2tuple(img_size) - patch_size = to_2tuple(patch_size) - patches_resolution = [ - img_size[0] // patch_size[0], # type: ignore - img_size[1] // patch_size[1], # type: ignore - ] - self.img_size = img_size - self.patch_size = patch_size - self.patches_resolution = patches_resolution - self.num_patches = patches_resolution[0] * patches_resolution[1] - - self.in_chans = in_chans - self.embed_dim = embed_dim - - def forward(self, x, x_size): - x = ( - x.transpose(1, 2) - .contiguous() - .view(x.shape[0], self.embed_dim, x_size[0], x_size[1]) - ) # b Ph*Pw c - return x - - -class Upsample(nn.Sequential): - """Upsample module. - Args: - scale (int): Scale factor. Supported scales: 2^n and 3. - num_feat (int): Channel number of intermediate features. - """ - - def __init__(self, scale, num_feat): - m = [] - if (scale & (scale - 1)) == 0: # scale = 2^n - for _ in range(int(math.log(scale, 2))): - m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(2)) - elif scale == 3: - m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(3)) - else: - raise ValueError( - f"scale {scale} is not supported. " "Supported scales: 2^n and 3." - ) - super(Upsample, self).__init__(*m) - - -class HAT(nn.Module): - r"""Hybrid Attention Transformer - A PyTorch implementation of : `Activating More Pixels in Image Super-Resolution Transformer`. - Some codes are based on SwinIR. - Args: - img_size (int | tuple(int)): Input image size. Default 64 - patch_size (int | tuple(int)): Patch size. Default: 1 - in_chans (int): Number of input image channels. Default: 3 - embed_dim (int): Patch embedding dimension. Default: 96 - depths (tuple(int)): Depth of each Swin Transformer layer. - num_heads (tuple(int)): Number of attention heads in different layers. - window_size (int): Window size. Default: 7 - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4 - qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None - drop_rate (float): Dropout rate. Default: 0 - attn_drop_rate (float): Attention dropout rate. Default: 0 - drop_path_rate (float): Stochastic depth rate. Default: 0.1 - norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. - ape (bool): If True, add absolute position embedding to the patch embedding. Default: False - patch_norm (bool): If True, add normalization after patch embedding. Default: True - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False - upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction - img_range: Image range. 1. or 255. - upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None - resi_connection: The convolutional block before residual connection. '1conv'/'3conv' - """ - - def __init__( - self, - state_dict, - **kwargs, - ): - super(HAT, self).__init__() - - # Defaults - img_size = 64 - patch_size = 1 - in_chans = 3 - embed_dim = 96 - depths = (6, 6, 6, 6) - num_heads = (6, 6, 6, 6) - window_size = 7 - compress_ratio = 3 - squeeze_factor = 30 - conv_scale = 0.01 - overlap_ratio = 0.5 - mlp_ratio = 4.0 - qkv_bias = True - qk_scale = None - drop_rate = 0.0 - attn_drop_rate = 0.0 - drop_path_rate = 0.1 - norm_layer = nn.LayerNorm - ape = False - patch_norm = True - use_checkpoint = False - upscale = 2 - img_range = 1.0 - upsampler = "" - resi_connection = "1conv" - - self.state = state_dict - self.model_arch = "HAT" - self.sub_type = "SR" - self.supports_fp16 = False - self.support_bf16 = True - self.min_size_restriction = 16 - - state_keys = list(state_dict.keys()) - - num_feat = state_dict["conv_last.weight"].shape[1] - in_chans = state_dict["conv_first.weight"].shape[1] - num_out_ch = state_dict["conv_last.weight"].shape[0] - embed_dim = state_dict["conv_first.weight"].shape[0] - - if "conv_before_upsample.0.weight" in state_keys: - if "conv_up1.weight" in state_keys: - upsampler = "nearest+conv" - else: - upsampler = "pixelshuffle" - supports_fp16 = False - elif "upsample.0.weight" in state_keys: - upsampler = "pixelshuffledirect" - else: - upsampler = "" - upscale = 1 - if upsampler == "nearest+conv": - upsample_keys = [ - x for x in state_keys if "conv_up" in x and "bias" not in x - ] - - for upsample_key in upsample_keys: - upscale *= 2 - elif upsampler == "pixelshuffle": - upsample_keys = [ - x - for x in state_keys - if "upsample" in x and "conv" not in x and "bias" not in x - ] - for upsample_key in upsample_keys: - shape = self.state[upsample_key].shape[0] - upscale *= math.sqrt(shape // num_feat) - upscale = int(upscale) - elif upsampler == "pixelshuffledirect": - upscale = int( - math.sqrt(self.state["upsample.0.bias"].shape[0] // num_out_ch) - ) - - max_layer_num = 0 - max_block_num = 0 - for key in state_keys: - result = re.match( - r"layers.(\d*).residual_group.blocks.(\d*).conv_block.cab.0.weight", key - ) - if result: - layer_num, block_num = result.groups() - max_layer_num = max(max_layer_num, int(layer_num)) - max_block_num = max(max_block_num, int(block_num)) - - depths = [max_block_num + 1 for _ in range(max_layer_num + 1)] - - if ( - "layers.0.residual_group.blocks.0.attn.relative_position_bias_table" - in state_keys - ): - num_heads_num = self.state[ - "layers.0.residual_group.blocks.0.attn.relative_position_bias_table" - ].shape[-1] - num_heads = [num_heads_num for _ in range(max_layer_num + 1)] - else: - num_heads = depths - - mlp_ratio = float( - self.state["layers.0.residual_group.blocks.0.mlp.fc1.bias"].shape[0] - / embed_dim - ) - - # TODO: could actually count the layers, but this should do - if "layers.0.conv.4.weight" in state_keys: - resi_connection = "3conv" - else: - resi_connection = "1conv" - - window_size = int(math.sqrt(self.state["relative_position_index_SA"].shape[0])) - - # Not sure if this is needed or used at all anywhere in HAT's config - if "layers.0.residual_group.blocks.1.attn_mask" in state_keys: - img_size = int( - math.sqrt( - self.state["layers.0.residual_group.blocks.1.attn_mask"].shape[0] - ) - * window_size - ) - - self.window_size = window_size - self.shift_size = window_size // 2 - self.overlap_ratio = overlap_ratio - - self.in_nc = in_chans - self.out_nc = num_out_ch - self.num_feat = num_feat - self.embed_dim = embed_dim - self.num_heads = num_heads - self.depths = depths - self.window_size = window_size - self.mlp_ratio = mlp_ratio - self.scale = upscale - self.upsampler = upsampler - self.img_size = img_size - self.img_range = img_range - self.resi_connection = resi_connection - - num_in_ch = in_chans - # num_out_ch = in_chans - # num_feat = 64 - self.img_range = img_range - if in_chans == 3: - rgb_mean = (0.4488, 0.4371, 0.4040) - self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1) - else: - self.mean = torch.zeros(1, 1, 1, 1) - self.upscale = upscale - self.upsampler = upsampler - - # relative position index - relative_position_index_SA = self.calculate_rpi_sa() - relative_position_index_OCA = self.calculate_rpi_oca() - self.register_buffer("relative_position_index_SA", relative_position_index_SA) - self.register_buffer("relative_position_index_OCA", relative_position_index_OCA) - - # ------------------------- 1, shallow feature extraction ------------------------- # - self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1) - - # ------------------------- 2, deep feature extraction ------------------------- # - self.num_layers = len(depths) - self.embed_dim = embed_dim - self.ape = ape - self.patch_norm = patch_norm - self.num_features = embed_dim - self.mlp_ratio = mlp_ratio - - # split image into non-overlapping patches - self.patch_embed = PatchEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=embed_dim, - embed_dim=embed_dim, - norm_layer=norm_layer if self.patch_norm else None, - ) - num_patches = self.patch_embed.num_patches - patches_resolution = self.patch_embed.patches_resolution - self.patches_resolution = patches_resolution - - # merge non-overlapping patches into image - self.patch_unembed = PatchUnEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=embed_dim, - embed_dim=embed_dim, - norm_layer=norm_layer if self.patch_norm else None, - ) - - # absolute position embedding - if self.ape: - self.absolute_pos_embed = nn.Parameter( # type: ignore[arg-type] - torch.zeros(1, num_patches, embed_dim) - ) - trunc_normal_(self.absolute_pos_embed, std=0.02) - - self.pos_drop = nn.Dropout(p=drop_rate) - - # stochastic depth - dpr = [ - x.item() for x in torch.linspace(0, drop_path_rate, sum(depths)) - ] # stochastic depth decay rule - - # build Residual Hybrid Attention Groups (RHAG) - self.layers = nn.ModuleList() - for i_layer in range(self.num_layers): - layer = RHAG( - dim=embed_dim, - input_resolution=(patches_resolution[0], patches_resolution[1]), - depth=depths[i_layer], - num_heads=num_heads[i_layer], - window_size=window_size, - compress_ratio=compress_ratio, - squeeze_factor=squeeze_factor, - conv_scale=conv_scale, - overlap_ratio=overlap_ratio, - mlp_ratio=self.mlp_ratio, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop_rate, - attn_drop=attn_drop_rate, - drop_path=dpr[ - sum(depths[:i_layer]) : sum(depths[: i_layer + 1]) # type: ignore - ], # no impact on SR results - norm_layer=norm_layer, - downsample=None, - use_checkpoint=use_checkpoint, - img_size=img_size, - patch_size=patch_size, - resi_connection=resi_connection, - ) - self.layers.append(layer) - self.norm = norm_layer(self.num_features) - - # build the last conv layer in deep feature extraction - if resi_connection == "1conv": - self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1) - elif resi_connection == "identity": - self.conv_after_body = nn.Identity() - - # ------------------------- 3, high quality image reconstruction ------------------------- # - if self.upsampler == "pixelshuffle": - # for classical SR - self.conv_before_upsample = nn.Sequential( - nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.upsample = Upsample(upscale, num_feat) - self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - - self.apply(self._init_weights) - self.load_state_dict(self.state, strict=False) - - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=0.02) - if isinstance(m, nn.Linear) and m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.LayerNorm): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) - - def calculate_rpi_sa(self): - # calculate relative position index for SA - coords_h = torch.arange(self.window_size) - coords_w = torch.arange(self.window_size) - coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww - coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww - relative_coords = ( - coords_flatten[:, :, None] - coords_flatten[:, None, :] - ) # 2, Wh*Ww, Wh*Ww - relative_coords = relative_coords.permute( - 1, 2, 0 - ).contiguous() # Wh*Ww, Wh*Ww, 2 - relative_coords[:, :, 0] += self.window_size - 1 # shift to start from 0 - relative_coords[:, :, 1] += self.window_size - 1 - relative_coords[:, :, 0] *= 2 * self.window_size - 1 - relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww - return relative_position_index - - def calculate_rpi_oca(self): - # calculate relative position index for OCA - window_size_ori = self.window_size - window_size_ext = self.window_size + int(self.overlap_ratio * self.window_size) - - coords_h = torch.arange(window_size_ori) - coords_w = torch.arange(window_size_ori) - coords_ori = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, ws, ws - coords_ori_flatten = torch.flatten(coords_ori, 1) # 2, ws*ws - - coords_h = torch.arange(window_size_ext) - coords_w = torch.arange(window_size_ext) - coords_ext = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, wse, wse - coords_ext_flatten = torch.flatten(coords_ext, 1) # 2, wse*wse - - relative_coords = ( - coords_ext_flatten[:, None, :] - coords_ori_flatten[:, :, None] - ) # 2, ws*ws, wse*wse - - relative_coords = relative_coords.permute( - 1, 2, 0 - ).contiguous() # ws*ws, wse*wse, 2 - relative_coords[:, :, 0] += ( - window_size_ori - window_size_ext + 1 - ) # shift to start from 0 - relative_coords[:, :, 1] += window_size_ori - window_size_ext + 1 - - relative_coords[:, :, 0] *= window_size_ori + window_size_ext - 1 - relative_position_index = relative_coords.sum(-1) - return relative_position_index - - def calculate_mask(self, x_size): - # calculate attention mask for SW-MSA - h, w = x_size - img_mask = torch.zeros((1, h, w, 1)) # 1 h w 1 - h_slices = ( - slice(0, -self.window_size), - slice(-self.window_size, -self.shift_size), - slice(-self.shift_size, None), - ) - w_slices = ( - slice(0, -self.window_size), - slice(-self.window_size, -self.shift_size), - slice(-self.shift_size, None), - ) - cnt = 0 - for h in h_slices: - for w in w_slices: - img_mask[:, h, w, :] = cnt - cnt += 1 - - mask_windows = window_partition( - img_mask, self.window_size - ) # nw, window_size, window_size, 1 - mask_windows = mask_windows.view(-1, self.window_size * self.window_size) - attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) - attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill( - attn_mask == 0, float(0.0) - ) - - return attn_mask - - @torch.jit.ignore # type: ignore - def no_weight_decay(self): - return {"absolute_pos_embed"} - - @torch.jit.ignore # type: ignore - def no_weight_decay_keywords(self): - return {"relative_position_bias_table"} - - def check_image_size(self, x): - _, _, h, w = x.size() - mod_pad_h = (self.window_size - h % self.window_size) % self.window_size - mod_pad_w = (self.window_size - w % self.window_size) % self.window_size - x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect") - return x - - def forward_features(self, x): - x_size = (x.shape[2], x.shape[3]) - - # Calculate attention mask and relative position index in advance to speed up inference. - # The original code is very time-cosuming for large window size. - attn_mask = self.calculate_mask(x_size).to(x.device) - params = { - "attn_mask": attn_mask, - "rpi_sa": self.relative_position_index_SA, - "rpi_oca": self.relative_position_index_OCA, - } - - x = self.patch_embed(x) - if self.ape: - x = x + self.absolute_pos_embed - x = self.pos_drop(x) - - for layer in self.layers: - x = layer(x, x_size, params) - - x = self.norm(x) # b seq_len c - x = self.patch_unembed(x, x_size) - - return x - - def forward(self, x): - H, W = x.shape[2:] - self.mean = self.mean.type_as(x) - x = (x - self.mean) * self.img_range - x = self.check_image_size(x) - - if self.upsampler == "pixelshuffle": - # for classical SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.conv_before_upsample(x) - x = self.conv_last(self.upsample(x)) - - x = x / self.img_range + self.mean - - return x[:, :, : H * self.upscale, : W * self.upscale] diff --git a/comfy_extras/chainner_models/architecture/LICENSE-DAT b/comfy_extras/chainner_models/architecture/LICENSE-DAT deleted file mode 100644 index 261eeb9e9..000000000 --- a/comfy_extras/chainner_models/architecture/LICENSE-DAT +++ /dev/null @@ -1,201 +0,0 @@ - Apache License - Version 2.0, January 2004 - http://www.apache.org/licenses/ - - TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION - - 1. Definitions. - - "License" shall mean the terms and conditions for use, reproduction, - and distribution as defined by Sections 1 through 9 of this document. - - "Licensor" shall mean the copyright owner or entity authorized by - the copyright owner that is granting the License. - - "Legal Entity" shall mean the union of the acting entity and all - other entities that control, are controlled by, or are under common - control with that entity. For the purposes of this definition, - "control" means (i) the power, direct or indirect, to cause the - direction or management of such entity, whether by contract or - otherwise, or (ii) ownership of fifty percent (50%) or more of the - outstanding shares, or (iii) beneficial ownership of such entity. - - "You" (or "Your") shall mean an individual or Legal Entity - exercising permissions granted by this License. - - "Source" form shall mean the preferred form for making modifications, - including but not limited to software source code, documentation - source, and configuration files. - - "Object" form shall mean any form resulting from mechanical - transformation or translation of a Source form, including but - not limited to compiled object code, generated documentation, - and conversions to other media types. - - "Work" shall mean the work of authorship, whether in Source or - Object form, made available under the License, as indicated by a - copyright notice that is included in or attached to the work - (an example is provided in the Appendix below). - - "Derivative Works" shall mean any work, whether in Source or Object - form, that is based on (or derived from) the Work and for which the - editorial revisions, annotations, elaborations, or other modifications - represent, as a whole, an original work of authorship. For the purposes - of this License, Derivative Works shall not include works that remain - separable from, or merely link (or bind by name) to the interfaces of, - the Work and Derivative Works thereof. - - "Contribution" shall mean any work of authorship, including - the original version of the Work and any modifications or additions - to that Work or Derivative Works thereof, that is intentionally - submitted to Licensor for inclusion in the Work by the copyright owner - or by an individual or Legal Entity authorized to submit on behalf of - the copyright owner. For the purposes of this definition, "submitted" - means any form of electronic, verbal, or written communication sent - to the Licensor or its representatives, including but not limited to - communication on electronic mailing lists, source code control systems, - and issue tracking systems that are managed by, or on behalf of, the - Licensor for the purpose of discussing and improving the Work, but - excluding communication that is conspicuously marked or otherwise - designated in writing by the copyright owner as "Not a Contribution." - - "Contributor" shall mean Licensor and any individual or Legal Entity - on behalf of whom a Contribution has been received by Licensor and - subsequently incorporated within the Work. - - 2. Grant of Copyright License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - copyright license to reproduce, prepare Derivative Works of, - publicly display, publicly perform, sublicense, and distribute the - Work and such Derivative Works in Source or Object form. - - 3. Grant of Patent License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - (except as stated in this section) patent license to make, have made, - use, offer to sell, sell, import, and otherwise transfer the Work, - where such license applies only to those patent claims licensable - by such Contributor that are necessarily infringed by their - Contribution(s) alone or by combination of their Contribution(s) - with the Work to which such Contribution(s) was submitted. If You - institute patent litigation against any entity (including a - cross-claim or counterclaim in a lawsuit) alleging that the Work - or a Contribution incorporated within the Work constitutes direct - or contributory patent infringement, then any patent licenses - granted to You under this License for that Work shall terminate - as of the date such litigation is filed. - - 4. Redistribution. You may reproduce and distribute copies of the - Work or Derivative Works thereof in any medium, with or without - modifications, and in Source or Object form, provided that You - meet the following conditions: - - (a) You must give any other recipients of the Work or - Derivative Works a copy of this License; and - - (b) You must cause any modified files to carry prominent notices - stating that You changed the files; and - - (c) You must retain, in the Source form of any Derivative Works - that You distribute, all copyright, patent, trademark, and - attribution notices from the Source form of the Work, - excluding those notices that do not pertain to any part of - the Derivative Works; and - - (d) If the Work includes a "NOTICE" text file as part of its - distribution, then any Derivative Works that You distribute must - include a readable copy of the attribution notices contained - within such NOTICE file, excluding those notices that do not - pertain to any part of the Derivative Works, in at least one - of the following places: within a NOTICE text file distributed - as part of the Derivative Works; within the Source form or - documentation, if provided along with the Derivative Works; or, - within a display generated by the Derivative Works, if and - wherever such third-party notices normally appear. The contents - of the NOTICE file are for informational purposes only and - do not modify the License. You may add Your own attribution - notices within Derivative Works that You distribute, alongside - or as an addendum to the NOTICE text from the Work, provided - that such additional attribution notices cannot be construed - as modifying the License. - - You may add Your own copyright statement to Your modifications and - may provide additional or different license terms and conditions - for use, reproduction, or distribution of Your modifications, or - for any such Derivative Works as a whole, provided Your use, - reproduction, and distribution of the Work otherwise complies with - the conditions stated in this License. - - 5. Submission of Contributions. Unless You explicitly state otherwise, - any Contribution intentionally submitted for inclusion in the Work - by You to the Licensor shall be under the terms and conditions of - this License, without any additional terms or conditions. - Notwithstanding the above, nothing herein shall supersede or modify - the terms of any separate license agreement you may have executed - with Licensor regarding such Contributions. - - 6. Trademarks. This License does not grant permission to use the trade - names, trademarks, service marks, or product names of the Licensor, - except as required for reasonable and customary use in describing the - origin of the Work and reproducing the content of the NOTICE file. - - 7. Disclaimer of Warranty. Unless required by applicable law or - agreed to in writing, Licensor provides the Work (and each - Contributor provides its Contributions) on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or - implied, including, without limitation, any warranties or conditions - of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A - PARTICULAR PURPOSE. You are solely responsible for determining the - appropriateness of using or redistributing the Work and assume any - risks associated with Your exercise of permissions under this License. - - 8. Limitation of Liability. In no event and under no legal theory, - whether in tort (including negligence), contract, or otherwise, - unless required by applicable law (such as deliberate and grossly - negligent acts) or agreed to in writing, shall any Contributor be - liable to You for damages, including any direct, indirect, special, - incidental, or consequential damages of any character arising as a - result of this License or out of the use or inability to use the - Work (including but not limited to damages for loss of goodwill, - work stoppage, computer failure or malfunction, or any and all - other commercial damages or losses), even if such Contributor - has been advised of the possibility of such damages. - - 9. Accepting Warranty or Additional Liability. While redistributing - the Work or Derivative Works thereof, You may choose to offer, - and charge a fee for, acceptance of support, warranty, indemnity, - or other liability obligations and/or rights consistent with this - License. However, in accepting such obligations, You may act only - on Your own behalf and on Your sole responsibility, not on behalf - of any other Contributor, and only if You agree to indemnify, - defend, and hold each Contributor harmless for any liability - incurred by, or claims asserted against, such Contributor by reason - of your accepting any such warranty or additional liability. - - END OF TERMS AND CONDITIONS - - APPENDIX: How to apply the Apache License to your work. - - To apply the Apache License to your work, attach the following - boilerplate notice, with the fields enclosed by brackets "[]" - replaced with your own identifying information. (Don't include - the brackets!) The text should be enclosed in the appropriate - comment syntax for the file format. We also recommend that a - file or class name and description of purpose be included on the - same "printed page" as the copyright notice for easier - identification within third-party archives. - - Copyright [yyyy] [name of copyright owner] - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. diff --git a/comfy_extras/chainner_models/architecture/LICENSE-ESRGAN b/comfy_extras/chainner_models/architecture/LICENSE-ESRGAN deleted file mode 100644 index 261eeb9e9..000000000 --- a/comfy_extras/chainner_models/architecture/LICENSE-ESRGAN +++ /dev/null @@ -1,201 +0,0 @@ - Apache License - Version 2.0, January 2004 - http://www.apache.org/licenses/ - - TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION - - 1. Definitions. - - "License" shall mean the terms and conditions for use, reproduction, - and distribution as defined by Sections 1 through 9 of this document. - - "Licensor" shall mean the copyright owner or entity authorized by - the copyright owner that is granting the License. - - "Legal Entity" shall mean the union of the acting entity and all - other entities that control, are controlled by, or are under common - control with that entity. For the purposes of this definition, - "control" means (i) the power, direct or indirect, to cause the - direction or management of such entity, whether by contract or - otherwise, or (ii) ownership of fifty percent (50%) or more of the - outstanding shares, or (iii) beneficial ownership of such entity. - - "You" (or "Your") shall mean an individual or Legal Entity - exercising permissions granted by this License. - - "Source" form shall mean the preferred form for making modifications, - including but not limited to software source code, documentation - source, and configuration files. - - "Object" form shall mean any form resulting from mechanical - transformation or translation of a Source form, including but - not limited to compiled object code, generated documentation, - and conversions to other media types. - - "Work" shall mean the work of authorship, whether in Source or - Object form, made available under the License, as indicated by a - copyright notice that is included in or attached to the work - (an example is provided in the Appendix below). - - "Derivative Works" shall mean any work, whether in Source or Object - form, that is based on (or derived from) the Work and for which the - editorial revisions, annotations, elaborations, or other modifications - represent, as a whole, an original work of authorship. For the purposes - of this License, Derivative Works shall not include works that remain - separable from, or merely link (or bind by name) to the interfaces of, - the Work and Derivative Works thereof. - - "Contribution" shall mean any work of authorship, including - the original version of the Work and any modifications or additions - to that Work or Derivative Works thereof, that is intentionally - submitted to Licensor for inclusion in the Work by the copyright owner - or by an individual or Legal Entity authorized to submit on behalf of - the copyright owner. For the purposes of this definition, "submitted" - means any form of electronic, verbal, or written communication sent - to the Licensor or its representatives, including but not limited to - communication on electronic mailing lists, source code control systems, - and issue tracking systems that are managed by, or on behalf of, the - Licensor for the purpose of discussing and improving the Work, but - excluding communication that is conspicuously marked or otherwise - designated in writing by the copyright owner as "Not a Contribution." - - "Contributor" shall mean Licensor and any individual or Legal Entity - on behalf of whom a Contribution has been received by Licensor and - subsequently incorporated within the Work. - - 2. Grant of Copyright License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - copyright license to reproduce, prepare Derivative Works of, - publicly display, publicly perform, sublicense, and distribute the - Work and such Derivative Works in Source or Object form. - - 3. Grant of Patent License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - (except as stated in this section) patent license to make, have made, - use, offer to sell, sell, import, and otherwise transfer the Work, - where such license applies only to those patent claims licensable - by such Contributor that are necessarily infringed by their - Contribution(s) alone or by combination of their Contribution(s) - with the Work to which such Contribution(s) was submitted. If You - institute patent litigation against any entity (including a - cross-claim or counterclaim in a lawsuit) alleging that the Work - or a Contribution incorporated within the Work constitutes direct - or contributory patent infringement, then any patent licenses - granted to You under this License for that Work shall terminate - as of the date such litigation is filed. - - 4. Redistribution. You may reproduce and distribute copies of the - Work or Derivative Works thereof in any medium, with or without - modifications, and in Source or Object form, provided that You - meet the following conditions: - - (a) You must give any other recipients of the Work or - Derivative Works a copy of this License; and - - (b) You must cause any modified files to carry prominent notices - stating that You changed the files; and - - (c) You must retain, in the Source form of any Derivative Works - that You distribute, all copyright, patent, trademark, and - attribution notices from the Source form of the Work, - excluding those notices that do not pertain to any part of - the Derivative Works; and - - (d) If the Work includes a "NOTICE" text file as part of its - distribution, then any Derivative Works that You distribute must - include a readable copy of the attribution notices contained - within such NOTICE file, excluding those notices that do not - pertain to any part of the Derivative Works, in at least one - of the following places: within a NOTICE text file distributed - as part of the Derivative Works; within the Source form or - documentation, if provided along with the Derivative Works; or, - within a display generated by the Derivative Works, if and - wherever such third-party notices normally appear. The contents - of the NOTICE file are for informational purposes only and - do not modify the License. You may add Your own attribution - notices within Derivative Works that You distribute, alongside - or as an addendum to the NOTICE text from the Work, provided - that such additional attribution notices cannot be construed - as modifying the License. - - You may add Your own copyright statement to Your modifications and - may provide additional or different license terms and conditions - for use, reproduction, or distribution of Your modifications, or - for any such Derivative Works as a whole, provided Your use, - reproduction, and distribution of the Work otherwise complies with - the conditions stated in this License. - - 5. Submission of Contributions. Unless You explicitly state otherwise, - any Contribution intentionally submitted for inclusion in the Work - by You to the Licensor shall be under the terms and conditions of - this License, without any additional terms or conditions. - Notwithstanding the above, nothing herein shall supersede or modify - the terms of any separate license agreement you may have executed - with Licensor regarding such Contributions. - - 6. Trademarks. This License does not grant permission to use the trade - names, trademarks, service marks, or product names of the Licensor, - except as required for reasonable and customary use in describing the - origin of the Work and reproducing the content of the NOTICE file. - - 7. Disclaimer of Warranty. Unless required by applicable law or - agreed to in writing, Licensor provides the Work (and each - Contributor provides its Contributions) on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or - implied, including, without limitation, any warranties or conditions - of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A - PARTICULAR PURPOSE. You are solely responsible for determining the - appropriateness of using or redistributing the Work and assume any - risks associated with Your exercise of permissions under this License. - - 8. Limitation of Liability. In no event and under no legal theory, - whether in tort (including negligence), contract, or otherwise, - unless required by applicable law (such as deliberate and grossly - negligent acts) or agreed to in writing, shall any Contributor be - liable to You for damages, including any direct, indirect, special, - incidental, or consequential damages of any character arising as a - result of this License or out of the use or inability to use the - Work (including but not limited to damages for loss of goodwill, - work stoppage, computer failure or malfunction, or any and all - other commercial damages or losses), even if such Contributor - has been advised of the possibility of such damages. - - 9. Accepting Warranty or Additional Liability. While redistributing - the Work or Derivative Works thereof, You may choose to offer, - and charge a fee for, acceptance of support, warranty, indemnity, - or other liability obligations and/or rights consistent with this - License. However, in accepting such obligations, You may act only - on Your own behalf and on Your sole responsibility, not on behalf - of any other Contributor, and only if You agree to indemnify, - defend, and hold each Contributor harmless for any liability - incurred by, or claims asserted against, such Contributor by reason - of your accepting any such warranty or additional liability. - - END OF TERMS AND CONDITIONS - - APPENDIX: How to apply the Apache License to your work. - - To apply the Apache License to your work, attach the following - boilerplate notice, with the fields enclosed by brackets "[]" - replaced with your own identifying information. (Don't include - the brackets!) The text should be enclosed in the appropriate - comment syntax for the file format. We also recommend that a - file or class name and description of purpose be included on the - same "printed page" as the copyright notice for easier - identification within third-party archives. - - Copyright [yyyy] [name of copyright owner] - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. diff --git a/comfy_extras/chainner_models/architecture/LICENSE-HAT b/comfy_extras/chainner_models/architecture/LICENSE-HAT deleted file mode 100644 index 003e97e96..000000000 --- a/comfy_extras/chainner_models/architecture/LICENSE-HAT +++ /dev/null @@ -1,21 +0,0 @@ -MIT License - -Copyright (c) 2022 Xiangyu Chen - -Permission is hereby granted, free of charge, to any person obtaining a copy -of this software and associated documentation files (the "Software"), to deal -in the Software without restriction, including without limitation the rights -to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -copies of the Software, and to permit persons to whom the Software is -furnished to do so, subject to the following conditions: - -The above copyright notice and this permission notice shall be included in all -copies or substantial portions of the Software. - -THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -SOFTWARE. diff --git a/comfy_extras/chainner_models/architecture/LICENSE-RealESRGAN b/comfy_extras/chainner_models/architecture/LICENSE-RealESRGAN deleted file mode 100644 index 552a1eeaf..000000000 --- a/comfy_extras/chainner_models/architecture/LICENSE-RealESRGAN +++ /dev/null @@ -1,29 +0,0 @@ -BSD 3-Clause License - -Copyright (c) 2021, Xintao Wang -All rights reserved. - -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions are met: - -1. Redistributions of source code must retain the above copyright notice, this - list of conditions and the following disclaimer. - -2. Redistributions in binary form must reproduce the above copyright notice, - this list of conditions and the following disclaimer in the documentation - and/or other materials provided with the distribution. - -3. Neither the name of the copyright holder nor the names of its - contributors may be used to endorse or promote products derived from - this software without specific prior written permission. - -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE -DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE -FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL -DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR -SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER -CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, -OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. diff --git a/comfy_extras/chainner_models/architecture/LICENSE-SCUNet b/comfy_extras/chainner_models/architecture/LICENSE-SCUNet deleted file mode 100644 index ff75c988f..000000000 --- a/comfy_extras/chainner_models/architecture/LICENSE-SCUNet +++ /dev/null @@ -1,201 +0,0 @@ - Apache License - Version 2.0, January 2004 - http://www.apache.org/licenses/ - - TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION - - 1. Definitions. - - "License" shall mean the terms and conditions for use, reproduction, - and distribution as defined by Sections 1 through 9 of this document. - - "Licensor" shall mean the copyright owner or entity authorized by - the copyright owner that is granting the License. - - "Legal Entity" shall mean the union of the acting entity and all - other entities that control, are controlled by, or are under common - control with that entity. For the purposes of this definition, - "control" means (i) the power, direct or indirect, to cause the - direction or management of such entity, whether by contract or - otherwise, or (ii) ownership of fifty percent (50%) or more of the - outstanding shares, or (iii) beneficial ownership of such entity. - - "You" (or "Your") shall mean an individual or Legal Entity - exercising permissions granted by this License. - - "Source" form shall mean the preferred form for making modifications, - including but not limited to software source code, documentation - source, and configuration files. - - "Object" form shall mean any form resulting from mechanical - transformation or translation of a Source form, including but - not limited to compiled object code, generated documentation, - and conversions to other media types. - - "Work" shall mean the work of authorship, whether in Source or - Object form, made available under the License, as indicated by a - copyright notice that is included in or attached to the work - (an example is provided in the Appendix below). - - "Derivative Works" shall mean any work, whether in Source or Object - form, that is based on (or derived from) the Work and for which the - editorial revisions, annotations, elaborations, or other modifications - represent, as a whole, an original work of authorship. For the purposes - of this License, Derivative Works shall not include works that remain - separable from, or merely link (or bind by name) to the interfaces of, - the Work and Derivative Works thereof. - - "Contribution" shall mean any work of authorship, including - the original version of the Work and any modifications or additions - to that Work or Derivative Works thereof, that is intentionally - submitted to Licensor for inclusion in the Work by the copyright owner - or by an individual or Legal Entity authorized to submit on behalf of - the copyright owner. For the purposes of this definition, "submitted" - means any form of electronic, verbal, or written communication sent - to the Licensor or its representatives, including but not limited to - communication on electronic mailing lists, source code control systems, - and issue tracking systems that are managed by, or on behalf of, the - Licensor for the purpose of discussing and improving the Work, but - excluding communication that is conspicuously marked or otherwise - designated in writing by the copyright owner as "Not a Contribution." - - "Contributor" shall mean Licensor and any individual or Legal Entity - on behalf of whom a Contribution has been received by Licensor and - subsequently incorporated within the Work. - - 2. Grant of Copyright License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - copyright license to reproduce, prepare Derivative Works of, - publicly display, publicly perform, sublicense, and distribute the - Work and such Derivative Works in Source or Object form. - - 3. Grant of Patent License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - (except as stated in this section) patent license to make, have made, - use, offer to sell, sell, import, and otherwise transfer the Work, - where such license applies only to those patent claims licensable - by such Contributor that are necessarily infringed by their - Contribution(s) alone or by combination of their Contribution(s) - with the Work to which such Contribution(s) was submitted. If You - institute patent litigation against any entity (including a - cross-claim or counterclaim in a lawsuit) alleging that the Work - or a Contribution incorporated within the Work constitutes direct - or contributory patent infringement, then any patent licenses - granted to You under this License for that Work shall terminate - as of the date such litigation is filed. - - 4. Redistribution. You may reproduce and distribute copies of the - Work or Derivative Works thereof in any medium, with or without - modifications, and in Source or Object form, provided that You - meet the following conditions: - - (a) You must give any other recipients of the Work or - Derivative Works a copy of this License; and - - (b) You must cause any modified files to carry prominent notices - stating that You changed the files; and - - (c) You must retain, in the Source form of any Derivative Works - that You distribute, all copyright, patent, trademark, and - attribution notices from the Source form of the Work, - excluding those notices that do not pertain to any part of - the Derivative Works; and - - (d) If the Work includes a "NOTICE" text file as part of its - distribution, then any Derivative Works that You distribute must - include a readable copy of the attribution notices contained - within such NOTICE file, excluding those notices that do not - pertain to any part of the Derivative Works, in at least one - of the following places: within a NOTICE text file distributed - as part of the Derivative Works; within the Source form or - documentation, if provided along with the Derivative Works; or, - within a display generated by the Derivative Works, if and - wherever such third-party notices normally appear. The contents - of the NOTICE file are for informational purposes only and - do not modify the License. You may add Your own attribution - notices within Derivative Works that You distribute, alongside - or as an addendum to the NOTICE text from the Work, provided - that such additional attribution notices cannot be construed - as modifying the License. - - You may add Your own copyright statement to Your modifications and - may provide additional or different license terms and conditions - for use, reproduction, or distribution of Your modifications, or - for any such Derivative Works as a whole, provided Your use, - reproduction, and distribution of the Work otherwise complies with - the conditions stated in this License. - - 5. Submission of Contributions. Unless You explicitly state otherwise, - any Contribution intentionally submitted for inclusion in the Work - by You to the Licensor shall be under the terms and conditions of - this License, without any additional terms or conditions. - Notwithstanding the above, nothing herein shall supersede or modify - the terms of any separate license agreement you may have executed - with Licensor regarding such Contributions. - - 6. Trademarks. This License does not grant permission to use the trade - names, trademarks, service marks, or product names of the Licensor, - except as required for reasonable and customary use in describing the - origin of the Work and reproducing the content of the NOTICE file. - - 7. Disclaimer of Warranty. Unless required by applicable law or - agreed to in writing, Licensor provides the Work (and each - Contributor provides its Contributions) on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or - implied, including, without limitation, any warranties or conditions - of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A - PARTICULAR PURPOSE. You are solely responsible for determining the - appropriateness of using or redistributing the Work and assume any - risks associated with Your exercise of permissions under this License. - - 8. Limitation of Liability. In no event and under no legal theory, - whether in tort (including negligence), contract, or otherwise, - unless required by applicable law (such as deliberate and grossly - negligent acts) or agreed to in writing, shall any Contributor be - liable to You for damages, including any direct, indirect, special, - incidental, or consequential damages of any character arising as a - result of this License or out of the use or inability to use the - Work (including but not limited to damages for loss of goodwill, - work stoppage, computer failure or malfunction, or any and all - other commercial damages or losses), even if such Contributor - has been advised of the possibility of such damages. - - 9. Accepting Warranty or Additional Liability. While redistributing - the Work or Derivative Works thereof, You may choose to offer, - and charge a fee for, acceptance of support, warranty, indemnity, - or other liability obligations and/or rights consistent with this - License. However, in accepting such obligations, You may act only - on Your own behalf and on Your sole responsibility, not on behalf - of any other Contributor, and only if You agree to indemnify, - defend, and hold each Contributor harmless for any liability - incurred by, or claims asserted against, such Contributor by reason - of your accepting any such warranty or additional liability. - - END OF TERMS AND CONDITIONS - - APPENDIX: How to apply the Apache License to your work. - - To apply the Apache License to your work, attach the following - boilerplate notice, with the fields enclosed by brackets "[]" - replaced with your own identifying information. (Don't include - the brackets!) The text should be enclosed in the appropriate - comment syntax for the file format. We also recommend that a - file or class name and description of purpose be included on the - same "printed page" as the copyright notice for easier - identification within third-party archives. - - Copyright 2022 Kai Zhang (cskaizhang@gmail.com, https://cszn.github.io/). All rights reserved. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. diff --git a/comfy_extras/chainner_models/architecture/LICENSE-SPSR b/comfy_extras/chainner_models/architecture/LICENSE-SPSR deleted file mode 100644 index 3245f3f9e..000000000 --- a/comfy_extras/chainner_models/architecture/LICENSE-SPSR +++ /dev/null @@ -1,201 +0,0 @@ - Apache License - Version 2.0, January 2004 - http://www.apache.org/licenses/ - - TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION - - 1. Definitions. - - "License" shall mean the terms and conditions for use, reproduction, - and distribution as defined by Sections 1 through 9 of this document. - - "Licensor" shall mean the copyright owner or entity authorized by - the copyright owner that is granting the License. - - "Legal Entity" shall mean the union of the acting entity and all - other entities that control, are controlled by, or are under common - control with that entity. For the purposes of this definition, - "control" means (i) the power, direct or indirect, to cause the - direction or management of such entity, whether by contract or - otherwise, or (ii) ownership of fifty percent (50%) or more of the - outstanding shares, or (iii) beneficial ownership of such entity. - - "You" (or "Your") shall mean an individual or Legal Entity - exercising permissions granted by this License. - - "Source" form shall mean the preferred form for making modifications, - including but not limited to software source code, documentation - source, and configuration files. - - "Object" form shall mean any form resulting from mechanical - transformation or translation of a Source form, including but - not limited to compiled object code, generated documentation, - and conversions to other media types. - - "Work" shall mean the work of authorship, whether in Source or - Object form, made available under the License, as indicated by a - copyright notice that is included in or attached to the work - (an example is provided in the Appendix below). - - "Derivative Works" shall mean any work, whether in Source or Object - form, that is based on (or derived from) the Work and for which the - editorial revisions, annotations, elaborations, or other modifications - represent, as a whole, an original work of authorship. For the purposes - of this License, Derivative Works shall not include works that remain - separable from, or merely link (or bind by name) to the interfaces of, - the Work and Derivative Works thereof. - - "Contribution" shall mean any work of authorship, including - the original version of the Work and any modifications or additions - to that Work or Derivative Works thereof, that is intentionally - submitted to Licensor for inclusion in the Work by the copyright owner - or by an individual or Legal Entity authorized to submit on behalf of - the copyright owner. For the purposes of this definition, "submitted" - means any form of electronic, verbal, or written communication sent - to the Licensor or its representatives, including but not limited to - communication on electronic mailing lists, source code control systems, - and issue tracking systems that are managed by, or on behalf of, the - Licensor for the purpose of discussing and improving the Work, but - excluding communication that is conspicuously marked or otherwise - designated in writing by the copyright owner as "Not a Contribution." - - "Contributor" shall mean Licensor and any individual or Legal Entity - on behalf of whom a Contribution has been received by Licensor and - subsequently incorporated within the Work. - - 2. Grant of Copyright License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - copyright license to reproduce, prepare Derivative Works of, - publicly display, publicly perform, sublicense, and distribute the - Work and such Derivative Works in Source or Object form. - - 3. Grant of Patent License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - (except as stated in this section) patent license to make, have made, - use, offer to sell, sell, import, and otherwise transfer the Work, - where such license applies only to those patent claims licensable - by such Contributor that are necessarily infringed by their - Contribution(s) alone or by combination of their Contribution(s) - with the Work to which such Contribution(s) was submitted. If You - institute patent litigation against any entity (including a - cross-claim or counterclaim in a lawsuit) alleging that the Work - or a Contribution incorporated within the Work constitutes direct - or contributory patent infringement, then any patent licenses - granted to You under this License for that Work shall terminate - as of the date such litigation is filed. - - 4. Redistribution. You may reproduce and distribute copies of the - Work or Derivative Works thereof in any medium, with or without - modifications, and in Source or Object form, provided that You - meet the following conditions: - - (a) You must give any other recipients of the Work or - Derivative Works a copy of this License; and - - (b) You must cause any modified files to carry prominent notices - stating that You changed the files; and - - (c) You must retain, in the Source form of any Derivative Works - that You distribute, all copyright, patent, trademark, and - attribution notices from the Source form of the Work, - excluding those notices that do not pertain to any part of - the Derivative Works; and - - (d) If the Work includes a "NOTICE" text file as part of its - distribution, then any Derivative Works that You distribute must - include a readable copy of the attribution notices contained - within such NOTICE file, excluding those notices that do not - pertain to any part of the Derivative Works, in at least one - of the following places: within a NOTICE text file distributed - as part of the Derivative Works; within the Source form or - documentation, if provided along with the Derivative Works; or, - within a display generated by the Derivative Works, if and - wherever such third-party notices normally appear. The contents - of the NOTICE file are for informational purposes only and - do not modify the License. You may add Your own attribution - notices within Derivative Works that You distribute, alongside - or as an addendum to the NOTICE text from the Work, provided - that such additional attribution notices cannot be construed - as modifying the License. - - You may add Your own copyright statement to Your modifications and - may provide additional or different license terms and conditions - for use, reproduction, or distribution of Your modifications, or - for any such Derivative Works as a whole, provided Your use, - reproduction, and distribution of the Work otherwise complies with - the conditions stated in this License. - - 5. Submission of Contributions. Unless You explicitly state otherwise, - any Contribution intentionally submitted for inclusion in the Work - by You to the Licensor shall be under the terms and conditions of - this License, without any additional terms or conditions. - Notwithstanding the above, nothing herein shall supersede or modify - the terms of any separate license agreement you may have executed - with Licensor regarding such Contributions. - - 6. Trademarks. This License does not grant permission to use the trade - names, trademarks, service marks, or product names of the Licensor, - except as required for reasonable and customary use in describing the - origin of the Work and reproducing the content of the NOTICE file. - - 7. Disclaimer of Warranty. Unless required by applicable law or - agreed to in writing, Licensor provides the Work (and each - Contributor provides its Contributions) on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or - implied, including, without limitation, any warranties or conditions - of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A - PARTICULAR PURPOSE. You are solely responsible for determining the - appropriateness of using or redistributing the Work and assume any - risks associated with Your exercise of permissions under this License. - - 8. Limitation of Liability. In no event and under no legal theory, - whether in tort (including negligence), contract, or otherwise, - unless required by applicable law (such as deliberate and grossly - negligent acts) or agreed to in writing, shall any Contributor be - liable to You for damages, including any direct, indirect, special, - incidental, or consequential damages of any character arising as a - result of this License or out of the use or inability to use the - Work (including but not limited to damages for loss of goodwill, - work stoppage, computer failure or malfunction, or any and all - other commercial damages or losses), even if such Contributor - has been advised of the possibility of such damages. - - 9. Accepting Warranty or Additional Liability. While redistributing - the Work or Derivative Works thereof, You may choose to offer, - and charge a fee for, acceptance of support, warranty, indemnity, - or other liability obligations and/or rights consistent with this - License. However, in accepting such obligations, You may act only - on Your own behalf and on Your sole responsibility, not on behalf - of any other Contributor, and only if You agree to indemnify, - defend, and hold each Contributor harmless for any liability - incurred by, or claims asserted against, such Contributor by reason - of your accepting any such warranty or additional liability. - - END OF TERMS AND CONDITIONS - - APPENDIX: How to apply the Apache License to your work. - - To apply the Apache License to your work, attach the following - boilerplate notice, with the fields enclosed by brackets "[]" - replaced with your own identifying information. (Don't include - the brackets!) The text should be enclosed in the appropriate - comment syntax for the file format. We also recommend that a - file or class name and description of purpose be included on the - same "printed page" as the copyright notice for easier - identification within third-party archives. - - Copyright 2018-2022 BasicSR Authors - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. diff --git a/comfy_extras/chainner_models/architecture/LICENSE-SwiftSRGAN b/comfy_extras/chainner_models/architecture/LICENSE-SwiftSRGAN deleted file mode 100644 index 0e259d42c..000000000 --- a/comfy_extras/chainner_models/architecture/LICENSE-SwiftSRGAN +++ /dev/null @@ -1,121 +0,0 @@ -Creative Commons Legal Code - -CC0 1.0 Universal - - CREATIVE COMMONS CORPORATION IS NOT A LAW FIRM AND DOES NOT PROVIDE - LEGAL SERVICES. DISTRIBUTION OF THIS DOCUMENT DOES NOT CREATE AN - ATTORNEY-CLIENT RELATIONSHIP. CREATIVE COMMONS PROVIDES THIS - INFORMATION ON AN "AS-IS" BASIS. CREATIVE COMMONS MAKES NO WARRANTIES - REGARDING THE USE OF THIS DOCUMENT OR THE INFORMATION OR WORKS - PROVIDED HEREUNDER, AND DISCLAIMS LIABILITY FOR DAMAGES RESULTING FROM - THE USE OF THIS DOCUMENT OR THE INFORMATION OR WORKS PROVIDED - HEREUNDER. - -Statement of Purpose - -The laws of most jurisdictions throughout the world automatically confer -exclusive Copyright and Related Rights (defined below) upon the creator -and subsequent owner(s) (each and all, an "owner") of an original work of -authorship and/or a database (each, a "Work"). - -Certain owners wish to permanently relinquish those rights to a Work for -the purpose of contributing to a commons of creative, cultural and -scientific works ("Commons") that the public can reliably and without fear -of later claims of infringement build upon, modify, incorporate in other -works, reuse and redistribute as freely as possible in any form whatsoever -and for any purposes, including without limitation commercial purposes. -These owners may contribute to the Commons to promote the ideal of a free -culture and the further production of creative, cultural and scientific -works, or to gain reputation or greater distribution for their Work in -part through the use and efforts of others. - -For these and/or other purposes and motivations, and without any -expectation of additional consideration or compensation, the person -associating CC0 with a Work (the "Affirmer"), to the extent that he or she -is an owner of Copyright and Related Rights in the Work, voluntarily -elects to apply CC0 to the Work and publicly distribute the Work under its -terms, with knowledge of his or her Copyright and Related Rights in the -Work and the meaning and intended legal effect of CC0 on those rights. - -1. Copyright and Related Rights. A Work made available under CC0 may be -protected by copyright and related or neighboring rights ("Copyright and -Related Rights"). Copyright and Related Rights include, but are not -limited to, the following: - - i. the right to reproduce, adapt, distribute, perform, display, - communicate, and translate a Work; - ii. moral rights retained by the original author(s) and/or performer(s); -iii. publicity and privacy rights pertaining to a person's image or - likeness depicted in a Work; - iv. rights protecting against unfair competition in regards to a Work, - subject to the limitations in paragraph 4(a), below; - v. rights protecting the extraction, dissemination, use and reuse of data - in a Work; - vi. database rights (such as those arising under Directive 96/9/EC of the - European Parliament and of the Council of 11 March 1996 on the legal - protection of databases, and under any national implementation - thereof, including any amended or successor version of such - directive); and -vii. other similar, equivalent or corresponding rights throughout the - world based on applicable law or treaty, and any national - implementations thereof. - -2. Waiver. To the greatest extent permitted by, but not in contravention -of, applicable law, Affirmer hereby overtly, fully, permanently, -irrevocably and unconditionally waives, abandons, and surrenders all of -Affirmer's Copyright and Related Rights and associated claims and causes -of action, whether now known or unknown (including existing as well as -future claims and causes of action), in the Work (i) in all territories -worldwide, (ii) for the maximum duration provided by applicable law or -treaty (including future time extensions), (iii) in any current or future -medium and for any number of copies, and (iv) for any purpose whatsoever, -including without limitation commercial, advertising or promotional -purposes (the "Waiver"). Affirmer makes the Waiver for the benefit of each -member of the public at large and to the detriment of Affirmer's heirs and -successors, fully intending that such Waiver shall not be subject to -revocation, rescission, cancellation, termination, or any other legal or -equitable action to disrupt the quiet enjoyment of the Work by the public -as contemplated by Affirmer's express Statement of Purpose. - -3. Public License Fallback. Should any part of the Waiver for any reason -be judged legally invalid or ineffective under applicable law, then the -Waiver shall be preserved to the maximum extent permitted taking into -account Affirmer's express Statement of Purpose. In addition, to the -extent the Waiver is so judged Affirmer hereby grants to each affected -person a royalty-free, non transferable, non sublicensable, non exclusive, -irrevocable and unconditional license to exercise Affirmer's Copyright and -Related Rights in the Work (i) in all territories worldwide, (ii) for the -maximum duration provided by applicable law or treaty (including future -time extensions), (iii) in any current or future medium and for any number -of copies, and (iv) for any purpose whatsoever, including without -limitation commercial, advertising or promotional purposes (the -"License"). The License shall be deemed effective as of the date CC0 was -applied by Affirmer to the Work. Should any part of the License for any -reason be judged legally invalid or ineffective under applicable law, such -partial invalidity or ineffectiveness shall not invalidate the remainder -of the License, and in such case Affirmer hereby affirms that he or she -will not (i) exercise any of his or her remaining Copyright and Related -Rights in the Work or (ii) assert any associated claims and causes of -action with respect to the Work, in either case contrary to Affirmer's -express Statement of Purpose. - -4. Limitations and Disclaimers. - - a. No trademark or patent rights held by Affirmer are waived, abandoned, - surrendered, licensed or otherwise affected by this document. - b. Affirmer offers the Work as-is and makes no representations or - warranties of any kind concerning the Work, express, implied, - statutory or otherwise, including without limitation warranties of - title, merchantability, fitness for a particular purpose, non - infringement, or the absence of latent or other defects, accuracy, or - the present or absence of errors, whether or not discoverable, all to - the greatest extent permissible under applicable law. - c. Affirmer disclaims responsibility for clearing rights of other persons - that may apply to the Work or any use thereof, including without - limitation any person's Copyright and Related Rights in the Work. - Further, Affirmer disclaims responsibility for obtaining any necessary - consents, permissions or other rights required for any use of the - Work. - d. Affirmer understands and acknowledges that Creative Commons is not a - party to this document and has no duty or obligation with respect to - this CC0 or use of the Work. diff --git a/comfy_extras/chainner_models/architecture/LICENSE-Swin2SR b/comfy_extras/chainner_models/architecture/LICENSE-Swin2SR deleted file mode 100644 index e5e4ee061..000000000 --- a/comfy_extras/chainner_models/architecture/LICENSE-Swin2SR +++ /dev/null @@ -1,201 +0,0 @@ - Apache License - Version 2.0, January 2004 - http://www.apache.org/licenses/ - - TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION - - 1. Definitions. - - "License" shall mean the terms and conditions for use, reproduction, - and distribution as defined by Sections 1 through 9 of this document. - - "Licensor" shall mean the copyright owner or entity authorized by - the copyright owner that is granting the License. - - "Legal Entity" shall mean the union of the acting entity and all - other entities that control, are controlled by, or are under common - control with that entity. For the purposes of this definition, - "control" means (i) the power, direct or indirect, to cause the - direction or management of such entity, whether by contract or - otherwise, or (ii) ownership of fifty percent (50%) or more of the - outstanding shares, or (iii) beneficial ownership of such entity. - - "You" (or "Your") shall mean an individual or Legal Entity - exercising permissions granted by this License. - - "Source" form shall mean the preferred form for making modifications, - including but not limited to software source code, documentation - source, and configuration files. - - "Object" form shall mean any form resulting from mechanical - transformation or translation of a Source form, including but - not limited to compiled object code, generated documentation, - and conversions to other media types. - - "Work" shall mean the work of authorship, whether in Source or - Object form, made available under the License, as indicated by a - copyright notice that is included in or attached to the work - (an example is provided in the Appendix below). - - "Derivative Works" shall mean any work, whether in Source or Object - form, that is based on (or derived from) the Work and for which the - editorial revisions, annotations, elaborations, or other modifications - represent, as a whole, an original work of authorship. For the purposes - of this License, Derivative Works shall not include works that remain - separable from, or merely link (or bind by name) to the interfaces of, - the Work and Derivative Works thereof. - - "Contribution" shall mean any work of authorship, including - the original version of the Work and any modifications or additions - to that Work or Derivative Works thereof, that is intentionally - submitted to Licensor for inclusion in the Work by the copyright owner - or by an individual or Legal Entity authorized to submit on behalf of - the copyright owner. For the purposes of this definition, "submitted" - means any form of electronic, verbal, or written communication sent - to the Licensor or its representatives, including but not limited to - communication on electronic mailing lists, source code control systems, - and issue tracking systems that are managed by, or on behalf of, the - Licensor for the purpose of discussing and improving the Work, but - excluding communication that is conspicuously marked or otherwise - designated in writing by the copyright owner as "Not a Contribution." - - "Contributor" shall mean Licensor and any individual or Legal Entity - on behalf of whom a Contribution has been received by Licensor and - subsequently incorporated within the Work. - - 2. Grant of Copyright License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - copyright license to reproduce, prepare Derivative Works of, - publicly display, publicly perform, sublicense, and distribute the - Work and such Derivative Works in Source or Object form. - - 3. Grant of Patent License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - (except as stated in this section) patent license to make, have made, - use, offer to sell, sell, import, and otherwise transfer the Work, - where such license applies only to those patent claims licensable - by such Contributor that are necessarily infringed by their - Contribution(s) alone or by combination of their Contribution(s) - with the Work to which such Contribution(s) was submitted. If You - institute patent litigation against any entity (including a - cross-claim or counterclaim in a lawsuit) alleging that the Work - or a Contribution incorporated within the Work constitutes direct - or contributory patent infringement, then any patent licenses - granted to You under this License for that Work shall terminate - as of the date such litigation is filed. - - 4. Redistribution. You may reproduce and distribute copies of the - Work or Derivative Works thereof in any medium, with or without - modifications, and in Source or Object form, provided that You - meet the following conditions: - - (a) You must give any other recipients of the Work or - Derivative Works a copy of this License; and - - (b) You must cause any modified files to carry prominent notices - stating that You changed the files; and - - (c) You must retain, in the Source form of any Derivative Works - that You distribute, all copyright, patent, trademark, and - attribution notices from the Source form of the Work, - excluding those notices that do not pertain to any part of - the Derivative Works; and - - (d) If the Work includes a "NOTICE" text file as part of its - distribution, then any Derivative Works that You distribute must - include a readable copy of the attribution notices contained - within such NOTICE file, excluding those notices that do not - pertain to any part of the Derivative Works, in at least one - of the following places: within a NOTICE text file distributed - as part of the Derivative Works; within the Source form or - documentation, if provided along with the Derivative Works; or, - within a display generated by the Derivative Works, if and - wherever such third-party notices normally appear. The contents - of the NOTICE file are for informational purposes only and - do not modify the License. You may add Your own attribution - notices within Derivative Works that You distribute, alongside - or as an addendum to the NOTICE text from the Work, provided - that such additional attribution notices cannot be construed - as modifying the License. - - You may add Your own copyright statement to Your modifications and - may provide additional or different license terms and conditions - for use, reproduction, or distribution of Your modifications, or - for any such Derivative Works as a whole, provided Your use, - reproduction, and distribution of the Work otherwise complies with - the conditions stated in this License. - - 5. Submission of Contributions. Unless You explicitly state otherwise, - any Contribution intentionally submitted for inclusion in the Work - by You to the Licensor shall be under the terms and conditions of - this License, without any additional terms or conditions. - Notwithstanding the above, nothing herein shall supersede or modify - the terms of any separate license agreement you may have executed - with Licensor regarding such Contributions. - - 6. Trademarks. This License does not grant permission to use the trade - names, trademarks, service marks, or product names of the Licensor, - except as required for reasonable and customary use in describing the - origin of the Work and reproducing the content of the NOTICE file. - - 7. Disclaimer of Warranty. Unless required by applicable law or - agreed to in writing, Licensor provides the Work (and each - Contributor provides its Contributions) on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or - implied, including, without limitation, any warranties or conditions - of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A - PARTICULAR PURPOSE. You are solely responsible for determining the - appropriateness of using or redistributing the Work and assume any - risks associated with Your exercise of permissions under this License. - - 8. Limitation of Liability. In no event and under no legal theory, - whether in tort (including negligence), contract, or otherwise, - unless required by applicable law (such as deliberate and grossly - negligent acts) or agreed to in writing, shall any Contributor be - liable to You for damages, including any direct, indirect, special, - incidental, or consequential damages of any character arising as a - result of this License or out of the use or inability to use the - Work (including but not limited to damages for loss of goodwill, - work stoppage, computer failure or malfunction, or any and all - other commercial damages or losses), even if such Contributor - has been advised of the possibility of such damages. - - 9. Accepting Warranty or Additional Liability. While redistributing - the Work or Derivative Works thereof, You may choose to offer, - and charge a fee for, acceptance of support, warranty, indemnity, - or other liability obligations and/or rights consistent with this - License. However, in accepting such obligations, You may act only - on Your own behalf and on Your sole responsibility, not on behalf - of any other Contributor, and only if You agree to indemnify, - defend, and hold each Contributor harmless for any liability - incurred by, or claims asserted against, such Contributor by reason - of your accepting any such warranty or additional liability. - - END OF TERMS AND CONDITIONS - - APPENDIX: How to apply the Apache License to your work. - - To apply the Apache License to your work, attach the following - boilerplate notice, with the fields enclosed by brackets "[]" - replaced with your own identifying information. (Don't include - the brackets!) The text should be enclosed in the appropriate - comment syntax for the file format. We also recommend that a - file or class name and description of purpose be included on the - same "printed page" as the copyright notice for easier - identification within third-party archives. - - Copyright [2021] [SwinIR Authors] - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. diff --git a/comfy_extras/chainner_models/architecture/LICENSE-SwinIR b/comfy_extras/chainner_models/architecture/LICENSE-SwinIR deleted file mode 100644 index e5e4ee061..000000000 --- a/comfy_extras/chainner_models/architecture/LICENSE-SwinIR +++ /dev/null @@ -1,201 +0,0 @@ - Apache License - Version 2.0, January 2004 - http://www.apache.org/licenses/ - - TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION - - 1. Definitions. - - "License" shall mean the terms and conditions for use, reproduction, - and distribution as defined by Sections 1 through 9 of this document. - - "Licensor" shall mean the copyright owner or entity authorized by - the copyright owner that is granting the License. - - "Legal Entity" shall mean the union of the acting entity and all - other entities that control, are controlled by, or are under common - control with that entity. For the purposes of this definition, - "control" means (i) the power, direct or indirect, to cause the - direction or management of such entity, whether by contract or - otherwise, or (ii) ownership of fifty percent (50%) or more of the - outstanding shares, or (iii) beneficial ownership of such entity. - - "You" (or "Your") shall mean an individual or Legal Entity - exercising permissions granted by this License. - - "Source" form shall mean the preferred form for making modifications, - including but not limited to software source code, documentation - source, and configuration files. - - "Object" form shall mean any form resulting from mechanical - transformation or translation of a Source form, including but - not limited to compiled object code, generated documentation, - and conversions to other media types. - - "Work" shall mean the work of authorship, whether in Source or - Object form, made available under the License, as indicated by a - copyright notice that is included in or attached to the work - (an example is provided in the Appendix below). - - "Derivative Works" shall mean any work, whether in Source or Object - form, that is based on (or derived from) the Work and for which the - editorial revisions, annotations, elaborations, or other modifications - represent, as a whole, an original work of authorship. For the purposes - of this License, Derivative Works shall not include works that remain - separable from, or merely link (or bind by name) to the interfaces of, - the Work and Derivative Works thereof. - - "Contribution" shall mean any work of authorship, including - the original version of the Work and any modifications or additions - to that Work or Derivative Works thereof, that is intentionally - submitted to Licensor for inclusion in the Work by the copyright owner - or by an individual or Legal Entity authorized to submit on behalf of - the copyright owner. For the purposes of this definition, "submitted" - means any form of electronic, verbal, or written communication sent - to the Licensor or its representatives, including but not limited to - communication on electronic mailing lists, source code control systems, - and issue tracking systems that are managed by, or on behalf of, the - Licensor for the purpose of discussing and improving the Work, but - excluding communication that is conspicuously marked or otherwise - designated in writing by the copyright owner as "Not a Contribution." - - "Contributor" shall mean Licensor and any individual or Legal Entity - on behalf of whom a Contribution has been received by Licensor and - subsequently incorporated within the Work. - - 2. Grant of Copyright License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - copyright license to reproduce, prepare Derivative Works of, - publicly display, publicly perform, sublicense, and distribute the - Work and such Derivative Works in Source or Object form. - - 3. Grant of Patent License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - (except as stated in this section) patent license to make, have made, - use, offer to sell, sell, import, and otherwise transfer the Work, - where such license applies only to those patent claims licensable - by such Contributor that are necessarily infringed by their - Contribution(s) alone or by combination of their Contribution(s) - with the Work to which such Contribution(s) was submitted. If You - institute patent litigation against any entity (including a - cross-claim or counterclaim in a lawsuit) alleging that the Work - or a Contribution incorporated within the Work constitutes direct - or contributory patent infringement, then any patent licenses - granted to You under this License for that Work shall terminate - as of the date such litigation is filed. - - 4. Redistribution. You may reproduce and distribute copies of the - Work or Derivative Works thereof in any medium, with or without - modifications, and in Source or Object form, provided that You - meet the following conditions: - - (a) You must give any other recipients of the Work or - Derivative Works a copy of this License; and - - (b) You must cause any modified files to carry prominent notices - stating that You changed the files; and - - (c) You must retain, in the Source form of any Derivative Works - that You distribute, all copyright, patent, trademark, and - attribution notices from the Source form of the Work, - excluding those notices that do not pertain to any part of - the Derivative Works; and - - (d) If the Work includes a "NOTICE" text file as part of its - distribution, then any Derivative Works that You distribute must - include a readable copy of the attribution notices contained - within such NOTICE file, excluding those notices that do not - pertain to any part of the Derivative Works, in at least one - of the following places: within a NOTICE text file distributed - as part of the Derivative Works; within the Source form or - documentation, if provided along with the Derivative Works; or, - within a display generated by the Derivative Works, if and - wherever such third-party notices normally appear. The contents - of the NOTICE file are for informational purposes only and - do not modify the License. You may add Your own attribution - notices within Derivative Works that You distribute, alongside - or as an addendum to the NOTICE text from the Work, provided - that such additional attribution notices cannot be construed - as modifying the License. - - You may add Your own copyright statement to Your modifications and - may provide additional or different license terms and conditions - for use, reproduction, or distribution of Your modifications, or - for any such Derivative Works as a whole, provided Your use, - reproduction, and distribution of the Work otherwise complies with - the conditions stated in this License. - - 5. Submission of Contributions. Unless You explicitly state otherwise, - any Contribution intentionally submitted for inclusion in the Work - by You to the Licensor shall be under the terms and conditions of - this License, without any additional terms or conditions. - Notwithstanding the above, nothing herein shall supersede or modify - the terms of any separate license agreement you may have executed - with Licensor regarding such Contributions. - - 6. Trademarks. This License does not grant permission to use the trade - names, trademarks, service marks, or product names of the Licensor, - except as required for reasonable and customary use in describing the - origin of the Work and reproducing the content of the NOTICE file. - - 7. Disclaimer of Warranty. Unless required by applicable law or - agreed to in writing, Licensor provides the Work (and each - Contributor provides its Contributions) on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or - implied, including, without limitation, any warranties or conditions - of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A - PARTICULAR PURPOSE. You are solely responsible for determining the - appropriateness of using or redistributing the Work and assume any - risks associated with Your exercise of permissions under this License. - - 8. Limitation of Liability. In no event and under no legal theory, - whether in tort (including negligence), contract, or otherwise, - unless required by applicable law (such as deliberate and grossly - negligent acts) or agreed to in writing, shall any Contributor be - liable to You for damages, including any direct, indirect, special, - incidental, or consequential damages of any character arising as a - result of this License or out of the use or inability to use the - Work (including but not limited to damages for loss of goodwill, - work stoppage, computer failure or malfunction, or any and all - other commercial damages or losses), even if such Contributor - has been advised of the possibility of such damages. - - 9. Accepting Warranty or Additional Liability. While redistributing - the Work or Derivative Works thereof, You may choose to offer, - and charge a fee for, acceptance of support, warranty, indemnity, - or other liability obligations and/or rights consistent with this - License. However, in accepting such obligations, You may act only - on Your own behalf and on Your sole responsibility, not on behalf - of any other Contributor, and only if You agree to indemnify, - defend, and hold each Contributor harmless for any liability - incurred by, or claims asserted against, such Contributor by reason - of your accepting any such warranty or additional liability. - - END OF TERMS AND CONDITIONS - - APPENDIX: How to apply the Apache License to your work. - - To apply the Apache License to your work, attach the following - boilerplate notice, with the fields enclosed by brackets "[]" - replaced with your own identifying information. (Don't include - the brackets!) The text should be enclosed in the appropriate - comment syntax for the file format. We also recommend that a - file or class name and description of purpose be included on the - same "printed page" as the copyright notice for easier - identification within third-party archives. - - Copyright [2021] [SwinIR Authors] - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. diff --git a/comfy_extras/chainner_models/architecture/LICENSE-lama b/comfy_extras/chainner_models/architecture/LICENSE-lama deleted file mode 100644 index ca822bb5f..000000000 --- a/comfy_extras/chainner_models/architecture/LICENSE-lama +++ /dev/null @@ -1,201 +0,0 @@ - Apache License - Version 2.0, January 2004 - http://www.apache.org/licenses/ - - TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION - - 1. Definitions. - - "License" shall mean the terms and conditions for use, reproduction, - and distribution as defined by Sections 1 through 9 of this document. - - "Licensor" shall mean the copyright owner or entity authorized by - the copyright owner that is granting the License. - - "Legal Entity" shall mean the union of the acting entity and all - other entities that control, are controlled by, or are under common - control with that entity. For the purposes of this definition, - "control" means (i) the power, direct or indirect, to cause the - direction or management of such entity, whether by contract or - otherwise, or (ii) ownership of fifty percent (50%) or more of the - outstanding shares, or (iii) beneficial ownership of such entity. - - "You" (or "Your") shall mean an individual or Legal Entity - exercising permissions granted by this License. - - "Source" form shall mean the preferred form for making modifications, - including but not limited to software source code, documentation - source, and configuration files. - - "Object" form shall mean any form resulting from mechanical - transformation or translation of a Source form, including but - not limited to compiled object code, generated documentation, - and conversions to other media types. - - "Work" shall mean the work of authorship, whether in Source or - Object form, made available under the License, as indicated by a - copyright notice that is included in or attached to the work - (an example is provided in the Appendix below). - - "Derivative Works" shall mean any work, whether in Source or Object - form, that is based on (or derived from) the Work and for which the - editorial revisions, annotations, elaborations, or other modifications - represent, as a whole, an original work of authorship. For the purposes - of this License, Derivative Works shall not include works that remain - separable from, or merely link (or bind by name) to the interfaces of, - the Work and Derivative Works thereof. - - "Contribution" shall mean any work of authorship, including - the original version of the Work and any modifications or additions - to that Work or Derivative Works thereof, that is intentionally - submitted to Licensor for inclusion in the Work by the copyright owner - or by an individual or Legal Entity authorized to submit on behalf of - the copyright owner. For the purposes of this definition, "submitted" - means any form of electronic, verbal, or written communication sent - to the Licensor or its representatives, including but not limited to - communication on electronic mailing lists, source code control systems, - and issue tracking systems that are managed by, or on behalf of, the - Licensor for the purpose of discussing and improving the Work, but - excluding communication that is conspicuously marked or otherwise - designated in writing by the copyright owner as "Not a Contribution." - - "Contributor" shall mean Licensor and any individual or Legal Entity - on behalf of whom a Contribution has been received by Licensor and - subsequently incorporated within the Work. - - 2. Grant of Copyright License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - copyright license to reproduce, prepare Derivative Works of, - publicly display, publicly perform, sublicense, and distribute the - Work and such Derivative Works in Source or Object form. - - 3. Grant of Patent License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - (except as stated in this section) patent license to make, have made, - use, offer to sell, sell, import, and otherwise transfer the Work, - where such license applies only to those patent claims licensable - by such Contributor that are necessarily infringed by their - Contribution(s) alone or by combination of their Contribution(s) - with the Work to which such Contribution(s) was submitted. If You - institute patent litigation against any entity (including a - cross-claim or counterclaim in a lawsuit) alleging that the Work - or a Contribution incorporated within the Work constitutes direct - or contributory patent infringement, then any patent licenses - granted to You under this License for that Work shall terminate - as of the date such litigation is filed. - - 4. Redistribution. You may reproduce and distribute copies of the - Work or Derivative Works thereof in any medium, with or without - modifications, and in Source or Object form, provided that You - meet the following conditions: - - (a) You must give any other recipients of the Work or - Derivative Works a copy of this License; and - - (b) You must cause any modified files to carry prominent notices - stating that You changed the files; and - - (c) You must retain, in the Source form of any Derivative Works - that You distribute, all copyright, patent, trademark, and - attribution notices from the Source form of the Work, - excluding those notices that do not pertain to any part of - the Derivative Works; and - - (d) If the Work includes a "NOTICE" text file as part of its - distribution, then any Derivative Works that You distribute must - include a readable copy of the attribution notices contained - within such NOTICE file, excluding those notices that do not - pertain to any part of the Derivative Works, in at least one - of the following places: within a NOTICE text file distributed - as part of the Derivative Works; within the Source form or - documentation, if provided along with the Derivative Works; or, - within a display generated by the Derivative Works, if and - wherever such third-party notices normally appear. The contents - of the NOTICE file are for informational purposes only and - do not modify the License. You may add Your own attribution - notices within Derivative Works that You distribute, alongside - or as an addendum to the NOTICE text from the Work, provided - that such additional attribution notices cannot be construed - as modifying the License. - - You may add Your own copyright statement to Your modifications and - may provide additional or different license terms and conditions - for use, reproduction, or distribution of Your modifications, or - for any such Derivative Works as a whole, provided Your use, - reproduction, and distribution of the Work otherwise complies with - the conditions stated in this License. - - 5. Submission of Contributions. Unless You explicitly state otherwise, - any Contribution intentionally submitted for inclusion in the Work - by You to the Licensor shall be under the terms and conditions of - this License, without any additional terms or conditions. - Notwithstanding the above, nothing herein shall supersede or modify - the terms of any separate license agreement you may have executed - with Licensor regarding such Contributions. - - 6. Trademarks. This License does not grant permission to use the trade - names, trademarks, service marks, or product names of the Licensor, - except as required for reasonable and customary use in describing the - origin of the Work and reproducing the content of the NOTICE file. - - 7. Disclaimer of Warranty. Unless required by applicable law or - agreed to in writing, Licensor provides the Work (and each - Contributor provides its Contributions) on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or - implied, including, without limitation, any warranties or conditions - of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A - PARTICULAR PURPOSE. You are solely responsible for determining the - appropriateness of using or redistributing the Work and assume any - risks associated with Your exercise of permissions under this License. - - 8. Limitation of Liability. In no event and under no legal theory, - whether in tort (including negligence), contract, or otherwise, - unless required by applicable law (such as deliberate and grossly - negligent acts) or agreed to in writing, shall any Contributor be - liable to You for damages, including any direct, indirect, special, - incidental, or consequential damages of any character arising as a - result of this License or out of the use or inability to use the - Work (including but not limited to damages for loss of goodwill, - work stoppage, computer failure or malfunction, or any and all - other commercial damages or losses), even if such Contributor - has been advised of the possibility of such damages. - - 9. Accepting Warranty or Additional Liability. While redistributing - the Work or Derivative Works thereof, You may choose to offer, - and charge a fee for, acceptance of support, warranty, indemnity, - or other liability obligations and/or rights consistent with this - License. However, in accepting such obligations, You may act only - on Your own behalf and on Your sole responsibility, not on behalf - of any other Contributor, and only if You agree to indemnify, - defend, and hold each Contributor harmless for any liability - incurred by, or claims asserted against, such Contributor by reason - of your accepting any such warranty or additional liability. - - END OF TERMS AND CONDITIONS - - APPENDIX: How to apply the Apache License to your work. - - To apply the Apache License to your work, attach the following - boilerplate notice, with the fields enclosed by brackets "[]" - replaced with your own identifying information. (Don't include - the brackets!) The text should be enclosed in the appropriate - comment syntax for the file format. We also recommend that a - file or class name and description of purpose be included on the - same "printed page" as the copyright notice for easier - identification within third-party archives. - - Copyright [2021] Samsung Research - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. diff --git a/comfy_extras/chainner_models/architecture/LaMa.py b/comfy_extras/chainner_models/architecture/LaMa.py deleted file mode 100644 index a781f3e4d..000000000 --- a/comfy_extras/chainner_models/architecture/LaMa.py +++ /dev/null @@ -1,694 +0,0 @@ -# pylint: skip-file -""" -Model adapted from advimman's lama project: https://github.com/advimman/lama -""" - -# Fast Fourier Convolution NeurIPS 2020 -# original implementation https://github.com/pkumivision/FFC/blob/main/model_zoo/ffc.py -# paper https://proceedings.neurips.cc/paper/2020/file/2fd5d41ec6cfab47e32164d5624269b1-Paper.pdf - -from typing import List - -import torch -import torch.nn as nn -import torch.nn.functional as F -from torchvision.transforms.functional import InterpolationMode, rotate - - -class LearnableSpatialTransformWrapper(nn.Module): - def __init__(self, impl, pad_coef=0.5, angle_init_range=80, train_angle=True): - super().__init__() - self.impl = impl - self.angle = torch.rand(1) * angle_init_range - if train_angle: - self.angle = nn.Parameter(self.angle, requires_grad=True) - self.pad_coef = pad_coef - - def forward(self, x): - if torch.is_tensor(x): - return self.inverse_transform(self.impl(self.transform(x)), x) - elif isinstance(x, tuple): - x_trans = tuple(self.transform(elem) for elem in x) - y_trans = self.impl(x_trans) - return tuple( - self.inverse_transform(elem, orig_x) for elem, orig_x in zip(y_trans, x) - ) - else: - raise ValueError(f"Unexpected input type {type(x)}") - - def transform(self, x): - height, width = x.shape[2:] - pad_h, pad_w = int(height * self.pad_coef), int(width * self.pad_coef) - x_padded = F.pad(x, [pad_w, pad_w, pad_h, pad_h], mode="reflect") - x_padded_rotated = rotate( - x_padded, self.angle.to(x_padded), InterpolationMode.BILINEAR, fill=0 - ) - - return x_padded_rotated - - def inverse_transform(self, y_padded_rotated, orig_x): - height, width = orig_x.shape[2:] - pad_h, pad_w = int(height * self.pad_coef), int(width * self.pad_coef) - - y_padded = rotate( - y_padded_rotated, - -self.angle.to(y_padded_rotated), - InterpolationMode.BILINEAR, - fill=0, - ) - y_height, y_width = y_padded.shape[2:] - y = y_padded[:, :, pad_h : y_height - pad_h, pad_w : y_width - pad_w] - return y - - -class SELayer(nn.Module): - def __init__(self, channel, reduction=16): - super(SELayer, self).__init__() - self.avg_pool = nn.AdaptiveAvgPool2d(1) - self.fc = nn.Sequential( - nn.Linear(channel, channel // reduction, bias=False), - nn.ReLU(inplace=True), - nn.Linear(channel // reduction, channel, bias=False), - nn.Sigmoid(), - ) - - def forward(self, x): - b, c, _, _ = x.size() - y = self.avg_pool(x).view(b, c) - y = self.fc(y).view(b, c, 1, 1) - res = x * y.expand_as(x) - return res - - -class FourierUnit(nn.Module): - def __init__( - self, - in_channels, - out_channels, - groups=1, - spatial_scale_factor=None, - spatial_scale_mode="bilinear", - spectral_pos_encoding=False, - use_se=False, - se_kwargs=None, - ffc3d=False, - fft_norm="ortho", - ): - # bn_layer not used - super(FourierUnit, self).__init__() - self.groups = groups - - self.conv_layer = torch.nn.Conv2d( - in_channels=in_channels * 2 + (2 if spectral_pos_encoding else 0), - out_channels=out_channels * 2, - kernel_size=1, - stride=1, - padding=0, - groups=self.groups, - bias=False, - ) - self.bn = torch.nn.BatchNorm2d(out_channels * 2) - self.relu = torch.nn.ReLU(inplace=True) - - # squeeze and excitation block - self.use_se = use_se - if use_se: - if se_kwargs is None: - se_kwargs = {} - self.se = SELayer(self.conv_layer.in_channels, **se_kwargs) - - self.spatial_scale_factor = spatial_scale_factor - self.spatial_scale_mode = spatial_scale_mode - self.spectral_pos_encoding = spectral_pos_encoding - self.ffc3d = ffc3d - self.fft_norm = fft_norm - - def forward(self, x): - half_check = False - if x.type() == "torch.cuda.HalfTensor": - # half only works on gpu anyway - half_check = True - - batch = x.shape[0] - - if self.spatial_scale_factor is not None: - orig_size = x.shape[-2:] - x = F.interpolate( - x, - scale_factor=self.spatial_scale_factor, - mode=self.spatial_scale_mode, - align_corners=False, - ) - - # (batch, c, h, w/2+1, 2) - fft_dim = (-3, -2, -1) if self.ffc3d else (-2, -1) - if half_check == True: - ffted = torch.fft.rfftn( - x.float(), dim=fft_dim, norm=self.fft_norm - ) # .type(torch.cuda.HalfTensor) - else: - ffted = torch.fft.rfftn(x, dim=fft_dim, norm=self.fft_norm) - - ffted = torch.stack((ffted.real, ffted.imag), dim=-1) - ffted = ffted.permute(0, 1, 4, 2, 3).contiguous() # (batch, c, 2, h, w/2+1) - ffted = ffted.view( - ( - batch, - -1, - ) - + ffted.size()[3:] - ) - - if self.spectral_pos_encoding: - height, width = ffted.shape[-2:] - coords_vert = ( - torch.linspace(0, 1, height)[None, None, :, None] - .expand(batch, 1, height, width) - .to(ffted) - ) - coords_hor = ( - torch.linspace(0, 1, width)[None, None, None, :] - .expand(batch, 1, height, width) - .to(ffted) - ) - ffted = torch.cat((coords_vert, coords_hor, ffted), dim=1) - - if self.use_se: - ffted = self.se(ffted) - - if half_check == True: - ffted = self.conv_layer(ffted.half()) # (batch, c*2, h, w/2+1) - else: - ffted = self.conv_layer( - ffted - ) # .type(torch.cuda.FloatTensor) # (batch, c*2, h, w/2+1) - - ffted = self.relu(self.bn(ffted)) - # forcing to be always float - ffted = ffted.float() - - ffted = ( - ffted.view( - ( - batch, - -1, - 2, - ) - + ffted.size()[2:] - ) - .permute(0, 1, 3, 4, 2) - .contiguous() - ) # (batch,c, t, h, w/2+1, 2) - - ffted = torch.complex(ffted[..., 0], ffted[..., 1]) - - ifft_shape_slice = x.shape[-3:] if self.ffc3d else x.shape[-2:] - output = torch.fft.irfftn( - ffted, s=ifft_shape_slice, dim=fft_dim, norm=self.fft_norm - ) - - if half_check == True: - output = output.half() - - if self.spatial_scale_factor is not None: - output = F.interpolate( - output, - size=orig_size, - mode=self.spatial_scale_mode, - align_corners=False, - ) - - return output - - -class SpectralTransform(nn.Module): - def __init__( - self, - in_channels, - out_channels, - stride=1, - groups=1, - enable_lfu=True, - separable_fu=False, - **fu_kwargs, - ): - # bn_layer not used - super(SpectralTransform, self).__init__() - self.enable_lfu = enable_lfu - if stride == 2: - self.downsample = nn.AvgPool2d(kernel_size=(2, 2), stride=2) - else: - self.downsample = nn.Identity() - - self.stride = stride - self.conv1 = nn.Sequential( - nn.Conv2d( - in_channels, out_channels // 2, kernel_size=1, groups=groups, bias=False - ), - nn.BatchNorm2d(out_channels // 2), - nn.ReLU(inplace=True), - ) - fu_class = FourierUnit - self.fu = fu_class(out_channels // 2, out_channels // 2, groups, **fu_kwargs) - if self.enable_lfu: - self.lfu = fu_class(out_channels // 2, out_channels // 2, groups) - self.conv2 = torch.nn.Conv2d( - out_channels // 2, out_channels, kernel_size=1, groups=groups, bias=False - ) - - def forward(self, x): - x = self.downsample(x) - x = self.conv1(x) - output = self.fu(x) - - if self.enable_lfu: - _, c, h, _ = x.shape - split_no = 2 - split_s = h // split_no - xs = torch.cat( - torch.split(x[:, : c // 4], split_s, dim=-2), dim=1 - ).contiguous() - xs = torch.cat(torch.split(xs, split_s, dim=-1), dim=1).contiguous() - xs = self.lfu(xs) - xs = xs.repeat(1, 1, split_no, split_no).contiguous() - else: - xs = 0 - - output = self.conv2(x + output + xs) - - return output - - -class FFC(nn.Module): - def __init__( - self, - in_channels, - out_channels, - kernel_size, - ratio_gin, - ratio_gout, - stride=1, - padding=0, - dilation=1, - groups=1, - bias=False, - enable_lfu=True, - padding_type="reflect", - gated=False, - **spectral_kwargs, - ): - super(FFC, self).__init__() - - assert stride == 1 or stride == 2, "Stride should be 1 or 2." - self.stride = stride - - in_cg = int(in_channels * ratio_gin) - in_cl = in_channels - in_cg - out_cg = int(out_channels * ratio_gout) - out_cl = out_channels - out_cg - # groups_g = 1 if groups == 1 else int(groups * ratio_gout) - # groups_l = 1 if groups == 1 else groups - groups_g - - self.ratio_gin = ratio_gin - self.ratio_gout = ratio_gout - self.global_in_num = in_cg - - module = nn.Identity if in_cl == 0 or out_cl == 0 else nn.Conv2d - self.convl2l = module( - in_cl, - out_cl, - kernel_size, - stride, - padding, - dilation, - groups, - bias, - padding_mode=padding_type, - ) - module = nn.Identity if in_cl == 0 or out_cg == 0 else nn.Conv2d - self.convl2g = module( - in_cl, - out_cg, - kernel_size, - stride, - padding, - dilation, - groups, - bias, - padding_mode=padding_type, - ) - module = nn.Identity if in_cg == 0 or out_cl == 0 else nn.Conv2d - self.convg2l = module( - in_cg, - out_cl, - kernel_size, - stride, - padding, - dilation, - groups, - bias, - padding_mode=padding_type, - ) - module = nn.Identity if in_cg == 0 or out_cg == 0 else SpectralTransform - self.convg2g = module( - in_cg, - out_cg, - stride, - 1 if groups == 1 else groups // 2, - enable_lfu, - **spectral_kwargs, - ) - - self.gated = gated - module = ( - nn.Identity if in_cg == 0 or out_cl == 0 or not self.gated else nn.Conv2d - ) - self.gate = module(in_channels, 2, 1) - - def forward(self, x): - x_l, x_g = x if type(x) is tuple else (x, 0) - out_xl, out_xg = 0, 0 - - if self.gated: - total_input_parts = [x_l] - if torch.is_tensor(x_g): - total_input_parts.append(x_g) - total_input = torch.cat(total_input_parts, dim=1) - - gates = torch.sigmoid(self.gate(total_input)) - g2l_gate, l2g_gate = gates.chunk(2, dim=1) - else: - g2l_gate, l2g_gate = 1, 1 - - if self.ratio_gout != 1: - out_xl = self.convl2l(x_l) + self.convg2l(x_g) * g2l_gate - if self.ratio_gout != 0: - out_xg = self.convl2g(x_l) * l2g_gate + self.convg2g(x_g) - - return out_xl, out_xg - - -class FFC_BN_ACT(nn.Module): - def __init__( - self, - in_channels, - out_channels, - kernel_size, - ratio_gin, - ratio_gout, - stride=1, - padding=0, - dilation=1, - groups=1, - bias=False, - norm_layer=nn.BatchNorm2d, - activation_layer=nn.Identity, - padding_type="reflect", - enable_lfu=True, - **kwargs, - ): - super(FFC_BN_ACT, self).__init__() - self.ffc = FFC( - in_channels, - out_channels, - kernel_size, - ratio_gin, - ratio_gout, - stride, - padding, - dilation, - groups, - bias, - enable_lfu, - padding_type=padding_type, - **kwargs, - ) - lnorm = nn.Identity if ratio_gout == 1 else norm_layer - gnorm = nn.Identity if ratio_gout == 0 else norm_layer - global_channels = int(out_channels * ratio_gout) - self.bn_l = lnorm(out_channels - global_channels) - self.bn_g = gnorm(global_channels) - - lact = nn.Identity if ratio_gout == 1 else activation_layer - gact = nn.Identity if ratio_gout == 0 else activation_layer - self.act_l = lact(inplace=True) - self.act_g = gact(inplace=True) - - def forward(self, x): - x_l, x_g = self.ffc(x) - x_l = self.act_l(self.bn_l(x_l)) - x_g = self.act_g(self.bn_g(x_g)) - return x_l, x_g - - -class FFCResnetBlock(nn.Module): - def __init__( - self, - dim, - padding_type, - norm_layer, - activation_layer=nn.ReLU, - dilation=1, - spatial_transform_kwargs=None, - inline=False, - **conv_kwargs, - ): - super().__init__() - self.conv1 = FFC_BN_ACT( - dim, - dim, - kernel_size=3, - padding=dilation, - dilation=dilation, - norm_layer=norm_layer, - activation_layer=activation_layer, - padding_type=padding_type, - **conv_kwargs, - ) - self.conv2 = FFC_BN_ACT( - dim, - dim, - kernel_size=3, - padding=dilation, - dilation=dilation, - norm_layer=norm_layer, - activation_layer=activation_layer, - padding_type=padding_type, - **conv_kwargs, - ) - if spatial_transform_kwargs is not None: - self.conv1 = LearnableSpatialTransformWrapper( - self.conv1, **spatial_transform_kwargs - ) - self.conv2 = LearnableSpatialTransformWrapper( - self.conv2, **spatial_transform_kwargs - ) - self.inline = inline - - def forward(self, x): - if self.inline: - x_l, x_g = ( - x[:, : -self.conv1.ffc.global_in_num], - x[:, -self.conv1.ffc.global_in_num :], - ) - else: - x_l, x_g = x if type(x) is tuple else (x, 0) - - id_l, id_g = x_l, x_g - - x_l, x_g = self.conv1((x_l, x_g)) - x_l, x_g = self.conv2((x_l, x_g)) - - x_l, x_g = id_l + x_l, id_g + x_g - out = x_l, x_g - if self.inline: - out = torch.cat(out, dim=1) - return out - - -class ConcatTupleLayer(nn.Module): - def forward(self, x): - assert isinstance(x, tuple) - x_l, x_g = x - assert torch.is_tensor(x_l) or torch.is_tensor(x_g) - if not torch.is_tensor(x_g): - return x_l - return torch.cat(x, dim=1) - - -class FFCResNetGenerator(nn.Module): - def __init__( - self, - input_nc, - output_nc, - ngf=64, - n_downsampling=3, - n_blocks=18, - norm_layer=nn.BatchNorm2d, - padding_type="reflect", - activation_layer=nn.ReLU, - up_norm_layer=nn.BatchNorm2d, - up_activation=nn.ReLU(True), - init_conv_kwargs={}, - downsample_conv_kwargs={}, - resnet_conv_kwargs={}, - spatial_transform_layers=None, - spatial_transform_kwargs={}, - max_features=1024, - out_ffc=False, - out_ffc_kwargs={}, - ): - assert n_blocks >= 0 - super().__init__() - """ - init_conv_kwargs = {'ratio_gin': 0, 'ratio_gout': 0, 'enable_lfu': False} - downsample_conv_kwargs = {'ratio_gin': '${generator.init_conv_kwargs.ratio_gout}', 'ratio_gout': '${generator.downsample_conv_kwargs.ratio_gin}', 'enable_lfu': False} - resnet_conv_kwargs = {'ratio_gin': 0.75, 'ratio_gout': '${generator.resnet_conv_kwargs.ratio_gin}', 'enable_lfu': False} - spatial_transform_kwargs = {} - out_ffc_kwargs = {} - """ - """ - print(input_nc, output_nc, ngf, n_downsampling, n_blocks, norm_layer, - padding_type, activation_layer, - up_norm_layer, up_activation, - spatial_transform_layers, - add_out_act, max_features, out_ffc, file=sys.stderr) - - 4 3 64 3 18 - reflect - - ReLU(inplace=True) - None sigmoid 1024 False - """ - init_conv_kwargs = {"ratio_gin": 0, "ratio_gout": 0, "enable_lfu": False} - downsample_conv_kwargs = {"ratio_gin": 0, "ratio_gout": 0, "enable_lfu": False} - resnet_conv_kwargs = { - "ratio_gin": 0.75, - "ratio_gout": 0.75, - "enable_lfu": False, - } - spatial_transform_kwargs = {} - out_ffc_kwargs = {} - - model = [ - nn.ReflectionPad2d(3), - FFC_BN_ACT( - input_nc, - ngf, - kernel_size=7, - padding=0, - norm_layer=norm_layer, - activation_layer=activation_layer, - **init_conv_kwargs, - ), - ] - - ### downsample - for i in range(n_downsampling): - mult = 2**i - if i == n_downsampling - 1: - cur_conv_kwargs = dict(downsample_conv_kwargs) - cur_conv_kwargs["ratio_gout"] = resnet_conv_kwargs.get("ratio_gin", 0) - else: - cur_conv_kwargs = downsample_conv_kwargs - model += [ - FFC_BN_ACT( - min(max_features, ngf * mult), - min(max_features, ngf * mult * 2), - kernel_size=3, - stride=2, - padding=1, - norm_layer=norm_layer, - activation_layer=activation_layer, - **cur_conv_kwargs, - ) - ] - - mult = 2**n_downsampling - feats_num_bottleneck = min(max_features, ngf * mult) - - ### resnet blocks - for i in range(n_blocks): - cur_resblock = FFCResnetBlock( - feats_num_bottleneck, - padding_type=padding_type, - activation_layer=activation_layer, - norm_layer=norm_layer, - **resnet_conv_kwargs, - ) - if spatial_transform_layers is not None and i in spatial_transform_layers: - cur_resblock = LearnableSpatialTransformWrapper( - cur_resblock, **spatial_transform_kwargs - ) - model += [cur_resblock] - - model += [ConcatTupleLayer()] - - ### upsample - for i in range(n_downsampling): - mult = 2 ** (n_downsampling - i) - model += [ - nn.ConvTranspose2d( - min(max_features, ngf * mult), - min(max_features, int(ngf * mult / 2)), - kernel_size=3, - stride=2, - padding=1, - output_padding=1, - ), - up_norm_layer(min(max_features, int(ngf * mult / 2))), - up_activation, - ] - - if out_ffc: - model += [ - FFCResnetBlock( - ngf, - padding_type=padding_type, - activation_layer=activation_layer, - norm_layer=norm_layer, - inline=True, - **out_ffc_kwargs, - ) - ] - - model += [ - nn.ReflectionPad2d(3), - nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0), - ] - model.append(nn.Sigmoid()) - self.model = nn.Sequential(*model) - - def forward(self, image, mask): - return self.model(torch.cat([image, mask], dim=1)) - - -class LaMa(nn.Module): - def __init__(self, state_dict) -> None: - super(LaMa, self).__init__() - self.model_arch = "LaMa" - self.sub_type = "Inpaint" - self.in_nc = 4 - self.out_nc = 3 - self.scale = 1 - - self.min_size = None - self.pad_mod = 8 - self.pad_to_square = False - - self.model = FFCResNetGenerator(self.in_nc, self.out_nc) - self.state = { - k.replace("generator.model", "model.model"): v - for k, v in state_dict.items() - } - - self.supports_fp16 = False - self.support_bf16 = True - - self.load_state_dict(self.state, strict=False) - - def forward(self, img, mask): - masked_img = img * (1 - mask) - inpainted_mask = mask * self.model.forward(masked_img, mask) - result = inpainted_mask + (1 - mask) * img - return result diff --git a/comfy_extras/chainner_models/architecture/OmniSR/ChannelAttention.py b/comfy_extras/chainner_models/architecture/OmniSR/ChannelAttention.py deleted file mode 100644 index f4d52aa1e..000000000 --- a/comfy_extras/chainner_models/architecture/OmniSR/ChannelAttention.py +++ /dev/null @@ -1,110 +0,0 @@ -import math - -import torch.nn as nn - - -class CA_layer(nn.Module): - def __init__(self, channel, reduction=16): - super(CA_layer, self).__init__() - # global average pooling - self.gap = nn.AdaptiveAvgPool2d(1) - self.fc = nn.Sequential( - nn.Conv2d(channel, channel // reduction, kernel_size=(1, 1), bias=False), - nn.GELU(), - nn.Conv2d(channel // reduction, channel, kernel_size=(1, 1), bias=False), - # nn.Sigmoid() - ) - - def forward(self, x): - y = self.fc(self.gap(x)) - return x * y.expand_as(x) - - -class Simple_CA_layer(nn.Module): - def __init__(self, channel): - super(Simple_CA_layer, self).__init__() - self.gap = nn.AdaptiveAvgPool2d(1) - self.fc = nn.Conv2d( - in_channels=channel, - out_channels=channel, - kernel_size=1, - padding=0, - stride=1, - groups=1, - bias=True, - ) - - def forward(self, x): - return x * self.fc(self.gap(x)) - - -class ECA_layer(nn.Module): - """Constructs a ECA module. - Args: - channel: Number of channels of the input feature map - k_size: Adaptive selection of kernel size - """ - - def __init__(self, channel): - super(ECA_layer, self).__init__() - - b = 1 - gamma = 2 - k_size = int(abs(math.log(channel, 2) + b) / gamma) - k_size = k_size if k_size % 2 else k_size + 1 - self.avg_pool = nn.AdaptiveAvgPool2d(1) - self.conv = nn.Conv1d( - 1, 1, kernel_size=k_size, padding=(k_size - 1) // 2, bias=False - ) - # self.sigmoid = nn.Sigmoid() - - def forward(self, x): - # x: input features with shape [b, c, h, w] - # b, c, h, w = x.size() - - # feature descriptor on the global spatial information - y = self.avg_pool(x) - - # Two different branches of ECA module - y = self.conv(y.squeeze(-1).transpose(-1, -2)).transpose(-1, -2).unsqueeze(-1) - - # Multi-scale information fusion - # y = self.sigmoid(y) - - return x * y.expand_as(x) - - -class ECA_MaxPool_layer(nn.Module): - """Constructs a ECA module. - Args: - channel: Number of channels of the input feature map - k_size: Adaptive selection of kernel size - """ - - def __init__(self, channel): - super(ECA_MaxPool_layer, self).__init__() - - b = 1 - gamma = 2 - k_size = int(abs(math.log(channel, 2) + b) / gamma) - k_size = k_size if k_size % 2 else k_size + 1 - self.max_pool = nn.AdaptiveMaxPool2d(1) - self.conv = nn.Conv1d( - 1, 1, kernel_size=k_size, padding=(k_size - 1) // 2, bias=False - ) - # self.sigmoid = nn.Sigmoid() - - def forward(self, x): - # x: input features with shape [b, c, h, w] - # b, c, h, w = x.size() - - # feature descriptor on the global spatial information - y = self.max_pool(x) - - # Two different branches of ECA module - y = self.conv(y.squeeze(-1).transpose(-1, -2)).transpose(-1, -2).unsqueeze(-1) - - # Multi-scale information fusion - # y = self.sigmoid(y) - - return x * y.expand_as(x) diff --git a/comfy_extras/chainner_models/architecture/OmniSR/LICENSE b/comfy_extras/chainner_models/architecture/OmniSR/LICENSE deleted file mode 100644 index 261eeb9e9..000000000 --- a/comfy_extras/chainner_models/architecture/OmniSR/LICENSE +++ /dev/null @@ -1,201 +0,0 @@ - Apache License - Version 2.0, January 2004 - http://www.apache.org/licenses/ - - TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION - - 1. Definitions. - - "License" shall mean the terms and conditions for use, reproduction, - and distribution as defined by Sections 1 through 9 of this document. - - "Licensor" shall mean the copyright owner or entity authorized by - the copyright owner that is granting the License. - - "Legal Entity" shall mean the union of the acting entity and all - other entities that control, are controlled by, or are under common - control with that entity. For the purposes of this definition, - "control" means (i) the power, direct or indirect, to cause the - direction or management of such entity, whether by contract or - otherwise, or (ii) ownership of fifty percent (50%) or more of the - outstanding shares, or (iii) beneficial ownership of such entity. - - "You" (or "Your") shall mean an individual or Legal Entity - exercising permissions granted by this License. - - "Source" form shall mean the preferred form for making modifications, - including but not limited to software source code, documentation - source, and configuration files. - - "Object" form shall mean any form resulting from mechanical - transformation or translation of a Source form, including but - not limited to compiled object code, generated documentation, - and conversions to other media types. - - "Work" shall mean the work of authorship, whether in Source or - Object form, made available under the License, as indicated by a - copyright notice that is included in or attached to the work - (an example is provided in the Appendix below). - - "Derivative Works" shall mean any work, whether in Source or Object - form, that is based on (or derived from) the Work and for which the - editorial revisions, annotations, elaborations, or other modifications - represent, as a whole, an original work of authorship. For the purposes - of this License, Derivative Works shall not include works that remain - separable from, or merely link (or bind by name) to the interfaces of, - the Work and Derivative Works thereof. - - "Contribution" shall mean any work of authorship, including - the original version of the Work and any modifications or additions - to that Work or Derivative Works thereof, that is intentionally - submitted to Licensor for inclusion in the Work by the copyright owner - or by an individual or Legal Entity authorized to submit on behalf of - the copyright owner. For the purposes of this definition, "submitted" - means any form of electronic, verbal, or written communication sent - to the Licensor or its representatives, including but not limited to - communication on electronic mailing lists, source code control systems, - and issue tracking systems that are managed by, or on behalf of, the - Licensor for the purpose of discussing and improving the Work, but - excluding communication that is conspicuously marked or otherwise - designated in writing by the copyright owner as "Not a Contribution." - - "Contributor" shall mean Licensor and any individual or Legal Entity - on behalf of whom a Contribution has been received by Licensor and - subsequently incorporated within the Work. - - 2. Grant of Copyright License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - copyright license to reproduce, prepare Derivative Works of, - publicly display, publicly perform, sublicense, and distribute the - Work and such Derivative Works in Source or Object form. - - 3. Grant of Patent License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - (except as stated in this section) patent license to make, have made, - use, offer to sell, sell, import, and otherwise transfer the Work, - where such license applies only to those patent claims licensable - by such Contributor that are necessarily infringed by their - Contribution(s) alone or by combination of their Contribution(s) - with the Work to which such Contribution(s) was submitted. If You - institute patent litigation against any entity (including a - cross-claim or counterclaim in a lawsuit) alleging that the Work - or a Contribution incorporated within the Work constitutes direct - or contributory patent infringement, then any patent licenses - granted to You under this License for that Work shall terminate - as of the date such litigation is filed. - - 4. Redistribution. You may reproduce and distribute copies of the - Work or Derivative Works thereof in any medium, with or without - modifications, and in Source or Object form, provided that You - meet the following conditions: - - (a) You must give any other recipients of the Work or - Derivative Works a copy of this License; and - - (b) You must cause any modified files to carry prominent notices - stating that You changed the files; and - - (c) You must retain, in the Source form of any Derivative Works - that You distribute, all copyright, patent, trademark, and - attribution notices from the Source form of the Work, - excluding those notices that do not pertain to any part of - the Derivative Works; and - - (d) If the Work includes a "NOTICE" text file as part of its - distribution, then any Derivative Works that You distribute must - include a readable copy of the attribution notices contained - within such NOTICE file, excluding those notices that do not - pertain to any part of the Derivative Works, in at least one - of the following places: within a NOTICE text file distributed - as part of the Derivative Works; within the Source form or - documentation, if provided along with the Derivative Works; or, - within a display generated by the Derivative Works, if and - wherever such third-party notices normally appear. The contents - of the NOTICE file are for informational purposes only and - do not modify the License. You may add Your own attribution - notices within Derivative Works that You distribute, alongside - or as an addendum to the NOTICE text from the Work, provided - that such additional attribution notices cannot be construed - as modifying the License. - - You may add Your own copyright statement to Your modifications and - may provide additional or different license terms and conditions - for use, reproduction, or distribution of Your modifications, or - for any such Derivative Works as a whole, provided Your use, - reproduction, and distribution of the Work otherwise complies with - the conditions stated in this License. - - 5. Submission of Contributions. Unless You explicitly state otherwise, - any Contribution intentionally submitted for inclusion in the Work - by You to the Licensor shall be under the terms and conditions of - this License, without any additional terms or conditions. - Notwithstanding the above, nothing herein shall supersede or modify - the terms of any separate license agreement you may have executed - with Licensor regarding such Contributions. - - 6. Trademarks. This License does not grant permission to use the trade - names, trademarks, service marks, or product names of the Licensor, - except as required for reasonable and customary use in describing the - origin of the Work and reproducing the content of the NOTICE file. - - 7. Disclaimer of Warranty. Unless required by applicable law or - agreed to in writing, Licensor provides the Work (and each - Contributor provides its Contributions) on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or - implied, including, without limitation, any warranties or conditions - of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A - PARTICULAR PURPOSE. You are solely responsible for determining the - appropriateness of using or redistributing the Work and assume any - risks associated with Your exercise of permissions under this License. - - 8. Limitation of Liability. In no event and under no legal theory, - whether in tort (including negligence), contract, or otherwise, - unless required by applicable law (such as deliberate and grossly - negligent acts) or agreed to in writing, shall any Contributor be - liable to You for damages, including any direct, indirect, special, - incidental, or consequential damages of any character arising as a - result of this License or out of the use or inability to use the - Work (including but not limited to damages for loss of goodwill, - work stoppage, computer failure or malfunction, or any and all - other commercial damages or losses), even if such Contributor - has been advised of the possibility of such damages. - - 9. Accepting Warranty or Additional Liability. While redistributing - the Work or Derivative Works thereof, You may choose to offer, - and charge a fee for, acceptance of support, warranty, indemnity, - or other liability obligations and/or rights consistent with this - License. However, in accepting such obligations, You may act only - on Your own behalf and on Your sole responsibility, not on behalf - of any other Contributor, and only if You agree to indemnify, - defend, and hold each Contributor harmless for any liability - incurred by, or claims asserted against, such Contributor by reason - of your accepting any such warranty or additional liability. - - END OF TERMS AND CONDITIONS - - APPENDIX: How to apply the Apache License to your work. - - To apply the Apache License to your work, attach the following - boilerplate notice, with the fields enclosed by brackets "[]" - replaced with your own identifying information. (Don't include - the brackets!) The text should be enclosed in the appropriate - comment syntax for the file format. We also recommend that a - file or class name and description of purpose be included on the - same "printed page" as the copyright notice for easier - identification within third-party archives. - - Copyright [yyyy] [name of copyright owner] - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. diff --git a/comfy_extras/chainner_models/architecture/OmniSR/OSA.py b/comfy_extras/chainner_models/architecture/OmniSR/OSA.py deleted file mode 100644 index d7a129696..000000000 --- a/comfy_extras/chainner_models/architecture/OmniSR/OSA.py +++ /dev/null @@ -1,577 +0,0 @@ -#!/usr/bin/env python3 -# -*- coding:utf-8 -*- -############################################################# -# File: OSA.py -# Created Date: Tuesday April 28th 2022 -# Author: Chen Xuanhong -# Email: chenxuanhongzju@outlook.com -# Last Modified: Sunday, 23rd April 2023 3:07:42 pm -# Modified By: Chen Xuanhong -# Copyright (c) 2020 Shanghai Jiao Tong University -############################################################# - -import torch -import torch.nn.functional as F -from einops import rearrange, repeat -from einops.layers.torch import Rearrange, Reduce -from torch import einsum, nn - -from .layernorm import LayerNorm2d - -# helpers - - -def exists(val): - return val is not None - - -def default(val, d): - return val if exists(val) else d - - -def cast_tuple(val, length=1): - return val if isinstance(val, tuple) else ((val,) * length) - - -# helper classes - - -class PreNormResidual(nn.Module): - def __init__(self, dim, fn): - super().__init__() - self.norm = nn.LayerNorm(dim) - self.fn = fn - - def forward(self, x): - return self.fn(self.norm(x)) + x - - -class Conv_PreNormResidual(nn.Module): - def __init__(self, dim, fn): - super().__init__() - self.norm = LayerNorm2d(dim) - self.fn = fn - - def forward(self, x): - return self.fn(self.norm(x)) + x - - -class FeedForward(nn.Module): - def __init__(self, dim, mult=2, dropout=0.0): - super().__init__() - inner_dim = int(dim * mult) - self.net = nn.Sequential( - nn.Linear(dim, inner_dim), - nn.GELU(), - nn.Dropout(dropout), - nn.Linear(inner_dim, dim), - nn.Dropout(dropout), - ) - - def forward(self, x): - return self.net(x) - - -class Conv_FeedForward(nn.Module): - def __init__(self, dim, mult=2, dropout=0.0): - super().__init__() - inner_dim = int(dim * mult) - self.net = nn.Sequential( - nn.Conv2d(dim, inner_dim, 1, 1, 0), - nn.GELU(), - nn.Dropout(dropout), - nn.Conv2d(inner_dim, dim, 1, 1, 0), - nn.Dropout(dropout), - ) - - def forward(self, x): - return self.net(x) - - -class Gated_Conv_FeedForward(nn.Module): - def __init__(self, dim, mult=1, bias=False, dropout=0.0): - super().__init__() - - hidden_features = int(dim * mult) - - self.project_in = nn.Conv2d(dim, hidden_features * 2, kernel_size=1, bias=bias) - - self.dwconv = nn.Conv2d( - hidden_features * 2, - hidden_features * 2, - kernel_size=3, - stride=1, - padding=1, - groups=hidden_features * 2, - bias=bias, - ) - - self.project_out = nn.Conv2d(hidden_features, dim, kernel_size=1, bias=bias) - - def forward(self, x): - x = self.project_in(x) - x1, x2 = self.dwconv(x).chunk(2, dim=1) - x = F.gelu(x1) * x2 - x = self.project_out(x) - return x - - -# MBConv - - -class SqueezeExcitation(nn.Module): - def __init__(self, dim, shrinkage_rate=0.25): - super().__init__() - hidden_dim = int(dim * shrinkage_rate) - - self.gate = nn.Sequential( - Reduce("b c h w -> b c", "mean"), - nn.Linear(dim, hidden_dim, bias=False), - nn.SiLU(), - nn.Linear(hidden_dim, dim, bias=False), - nn.Sigmoid(), - Rearrange("b c -> b c 1 1"), - ) - - def forward(self, x): - return x * self.gate(x) - - -class MBConvResidual(nn.Module): - def __init__(self, fn, dropout=0.0): - super().__init__() - self.fn = fn - self.dropsample = Dropsample(dropout) - - def forward(self, x): - out = self.fn(x) - out = self.dropsample(out) - return out + x - - -class Dropsample(nn.Module): - def __init__(self, prob=0): - super().__init__() - self.prob = prob - - def forward(self, x): - device = x.device - - if self.prob == 0.0 or (not self.training): - return x - - keep_mask = ( - torch.FloatTensor((x.shape[0], 1, 1, 1), device=device).uniform_() - > self.prob - ) - return x * keep_mask / (1 - self.prob) - - -def MBConv( - dim_in, dim_out, *, downsample, expansion_rate=4, shrinkage_rate=0.25, dropout=0.0 -): - hidden_dim = int(expansion_rate * dim_out) - stride = 2 if downsample else 1 - - net = nn.Sequential( - nn.Conv2d(dim_in, hidden_dim, 1), - # nn.BatchNorm2d(hidden_dim), - nn.GELU(), - nn.Conv2d( - hidden_dim, hidden_dim, 3, stride=stride, padding=1, groups=hidden_dim - ), - # nn.BatchNorm2d(hidden_dim), - nn.GELU(), - SqueezeExcitation(hidden_dim, shrinkage_rate=shrinkage_rate), - nn.Conv2d(hidden_dim, dim_out, 1), - # nn.BatchNorm2d(dim_out) - ) - - if dim_in == dim_out and not downsample: - net = MBConvResidual(net, dropout=dropout) - - return net - - -# attention related classes -class Attention(nn.Module): - def __init__( - self, - dim, - dim_head=32, - dropout=0.0, - window_size=7, - with_pe=True, - ): - super().__init__() - assert ( - dim % dim_head - ) == 0, "dimension should be divisible by dimension per head" - - self.heads = dim // dim_head - self.scale = dim_head**-0.5 - self.with_pe = with_pe - - self.to_qkv = nn.Linear(dim, dim * 3, bias=False) - - self.attend = nn.Sequential(nn.Softmax(dim=-1), nn.Dropout(dropout)) - - self.to_out = nn.Sequential( - nn.Linear(dim, dim, bias=False), nn.Dropout(dropout) - ) - - # relative positional bias - if self.with_pe: - self.rel_pos_bias = nn.Embedding((2 * window_size - 1) ** 2, self.heads) - - pos = torch.arange(window_size) - grid = torch.stack(torch.meshgrid(pos, pos)) - grid = rearrange(grid, "c i j -> (i j) c") - rel_pos = rearrange(grid, "i ... -> i 1 ...") - rearrange( - grid, "j ... -> 1 j ..." - ) - rel_pos += window_size - 1 - rel_pos_indices = (rel_pos * torch.tensor([2 * window_size - 1, 1])).sum( - dim=-1 - ) - - self.register_buffer("rel_pos_indices", rel_pos_indices, persistent=False) - - def forward(self, x): - batch, height, width, window_height, window_width, _, device, h = ( - *x.shape, - x.device, - self.heads, - ) - - # flatten - - x = rearrange(x, "b x y w1 w2 d -> (b x y) (w1 w2) d") - - # project for queries, keys, values - - q, k, v = self.to_qkv(x).chunk(3, dim=-1) - - # split heads - - q, k, v = map(lambda t: rearrange(t, "b n (h d ) -> b h n d", h=h), (q, k, v)) - - # scale - - q = q * self.scale - - # sim - - sim = einsum("b h i d, b h j d -> b h i j", q, k) - - # add positional bias - if self.with_pe: - bias = self.rel_pos_bias(self.rel_pos_indices) - sim = sim + rearrange(bias, "i j h -> h i j") - - # attention - - attn = self.attend(sim) - - # aggregate - - out = einsum("b h i j, b h j d -> b h i d", attn, v) - - # merge heads - - out = rearrange( - out, "b h (w1 w2) d -> b w1 w2 (h d)", w1=window_height, w2=window_width - ) - - # combine heads out - - out = self.to_out(out) - return rearrange(out, "(b x y) ... -> b x y ...", x=height, y=width) - - -class Block_Attention(nn.Module): - def __init__( - self, - dim, - dim_head=32, - bias=False, - dropout=0.0, - window_size=7, - with_pe=True, - ): - super().__init__() - assert ( - dim % dim_head - ) == 0, "dimension should be divisible by dimension per head" - - self.heads = dim // dim_head - self.ps = window_size - self.scale = dim_head**-0.5 - self.with_pe = with_pe - - self.qkv = nn.Conv2d(dim, dim * 3, kernel_size=1, bias=bias) - self.qkv_dwconv = nn.Conv2d( - dim * 3, - dim * 3, - kernel_size=3, - stride=1, - padding=1, - groups=dim * 3, - bias=bias, - ) - - self.attend = nn.Sequential(nn.Softmax(dim=-1), nn.Dropout(dropout)) - - self.to_out = nn.Conv2d(dim, dim, kernel_size=1, bias=bias) - - def forward(self, x): - # project for queries, keys, values - b, c, h, w = x.shape - - qkv = self.qkv_dwconv(self.qkv(x)) - q, k, v = qkv.chunk(3, dim=1) - - # split heads - - q, k, v = map( - lambda t: rearrange( - t, - "b (h d) (x w1) (y w2) -> (b x y) h (w1 w2) d", - h=self.heads, - w1=self.ps, - w2=self.ps, - ), - (q, k, v), - ) - - # scale - - q = q * self.scale - - # sim - - sim = einsum("b h i d, b h j d -> b h i j", q, k) - - # attention - attn = self.attend(sim) - - # aggregate - - out = einsum("b h i j, b h j d -> b h i d", attn, v) - - # merge heads - out = rearrange( - out, - "(b x y) head (w1 w2) d -> b (head d) (x w1) (y w2)", - x=h // self.ps, - y=w // self.ps, - head=self.heads, - w1=self.ps, - w2=self.ps, - ) - - out = self.to_out(out) - return out - - -class Channel_Attention(nn.Module): - def __init__(self, dim, heads, bias=False, dropout=0.0, window_size=7): - super(Channel_Attention, self).__init__() - self.heads = heads - - self.temperature = nn.Parameter(torch.ones(heads, 1, 1)) - - self.ps = window_size - - self.qkv = nn.Conv2d(dim, dim * 3, kernel_size=1, bias=bias) - self.qkv_dwconv = nn.Conv2d( - dim * 3, - dim * 3, - kernel_size=3, - stride=1, - padding=1, - groups=dim * 3, - bias=bias, - ) - self.project_out = nn.Conv2d(dim, dim, kernel_size=1, bias=bias) - - def forward(self, x): - b, c, h, w = x.shape - - qkv = self.qkv_dwconv(self.qkv(x)) - qkv = qkv.chunk(3, dim=1) - - q, k, v = map( - lambda t: rearrange( - t, - "b (head d) (h ph) (w pw) -> b (h w) head d (ph pw)", - ph=self.ps, - pw=self.ps, - head=self.heads, - ), - qkv, - ) - - q = F.normalize(q, dim=-1) - k = F.normalize(k, dim=-1) - - attn = (q @ k.transpose(-2, -1)) * self.temperature - attn = attn.softmax(dim=-1) - out = attn @ v - - out = rearrange( - out, - "b (h w) head d (ph pw) -> b (head d) (h ph) (w pw)", - h=h // self.ps, - w=w // self.ps, - ph=self.ps, - pw=self.ps, - head=self.heads, - ) - - out = self.project_out(out) - - return out - - -class Channel_Attention_grid(nn.Module): - def __init__(self, dim, heads, bias=False, dropout=0.0, window_size=7): - super(Channel_Attention_grid, self).__init__() - self.heads = heads - - self.temperature = nn.Parameter(torch.ones(heads, 1, 1)) - - self.ps = window_size - - self.qkv = nn.Conv2d(dim, dim * 3, kernel_size=1, bias=bias) - self.qkv_dwconv = nn.Conv2d( - dim * 3, - dim * 3, - kernel_size=3, - stride=1, - padding=1, - groups=dim * 3, - bias=bias, - ) - self.project_out = nn.Conv2d(dim, dim, kernel_size=1, bias=bias) - - def forward(self, x): - b, c, h, w = x.shape - - qkv = self.qkv_dwconv(self.qkv(x)) - qkv = qkv.chunk(3, dim=1) - - q, k, v = map( - lambda t: rearrange( - t, - "b (head d) (h ph) (w pw) -> b (ph pw) head d (h w)", - ph=self.ps, - pw=self.ps, - head=self.heads, - ), - qkv, - ) - - q = F.normalize(q, dim=-1) - k = F.normalize(k, dim=-1) - - attn = (q @ k.transpose(-2, -1)) * self.temperature - attn = attn.softmax(dim=-1) - out = attn @ v - - out = rearrange( - out, - "b (ph pw) head d (h w) -> b (head d) (h ph) (w pw)", - h=h // self.ps, - w=w // self.ps, - ph=self.ps, - pw=self.ps, - head=self.heads, - ) - - out = self.project_out(out) - - return out - - -class OSA_Block(nn.Module): - def __init__( - self, - channel_num=64, - bias=True, - ffn_bias=True, - window_size=8, - with_pe=False, - dropout=0.0, - ): - super(OSA_Block, self).__init__() - - w = window_size - - self.layer = nn.Sequential( - MBConv( - channel_num, - channel_num, - downsample=False, - expansion_rate=1, - shrinkage_rate=0.25, - ), - Rearrange( - "b d (x w1) (y w2) -> b x y w1 w2 d", w1=w, w2=w - ), # block-like attention - PreNormResidual( - channel_num, - Attention( - dim=channel_num, - dim_head=channel_num // 4, - dropout=dropout, - window_size=window_size, - with_pe=with_pe, - ), - ), - Rearrange("b x y w1 w2 d -> b d (x w1) (y w2)"), - Conv_PreNormResidual( - channel_num, Gated_Conv_FeedForward(dim=channel_num, dropout=dropout) - ), - # channel-like attention - Conv_PreNormResidual( - channel_num, - Channel_Attention( - dim=channel_num, heads=4, dropout=dropout, window_size=window_size - ), - ), - Conv_PreNormResidual( - channel_num, Gated_Conv_FeedForward(dim=channel_num, dropout=dropout) - ), - Rearrange( - "b d (w1 x) (w2 y) -> b x y w1 w2 d", w1=w, w2=w - ), # grid-like attention - PreNormResidual( - channel_num, - Attention( - dim=channel_num, - dim_head=channel_num // 4, - dropout=dropout, - window_size=window_size, - with_pe=with_pe, - ), - ), - Rearrange("b x y w1 w2 d -> b d (w1 x) (w2 y)"), - Conv_PreNormResidual( - channel_num, Gated_Conv_FeedForward(dim=channel_num, dropout=dropout) - ), - # channel-like attention - Conv_PreNormResidual( - channel_num, - Channel_Attention_grid( - dim=channel_num, heads=4, dropout=dropout, window_size=window_size - ), - ), - Conv_PreNormResidual( - channel_num, Gated_Conv_FeedForward(dim=channel_num, dropout=dropout) - ), - ) - - def forward(self, x): - out = self.layer(x) - return out diff --git a/comfy_extras/chainner_models/architecture/OmniSR/OSAG.py b/comfy_extras/chainner_models/architecture/OmniSR/OSAG.py deleted file mode 100644 index 477e81f9d..000000000 --- a/comfy_extras/chainner_models/architecture/OmniSR/OSAG.py +++ /dev/null @@ -1,60 +0,0 @@ -#!/usr/bin/env python3 -# -*- coding:utf-8 -*- -############################################################# -# File: OSAG.py -# Created Date: Tuesday April 28th 2022 -# Author: Chen Xuanhong -# Email: chenxuanhongzju@outlook.com -# Last Modified: Sunday, 23rd April 2023 3:08:49 pm -# Modified By: Chen Xuanhong -# Copyright (c) 2020 Shanghai Jiao Tong University -############################################################# - - -import torch.nn as nn - -from .esa import ESA -from .OSA import OSA_Block - - -class OSAG(nn.Module): - def __init__( - self, - channel_num=64, - bias=True, - block_num=4, - ffn_bias=False, - window_size=0, - pe=False, - ): - super(OSAG, self).__init__() - - # print("window_size: %d" % (window_size)) - # print("with_pe", pe) - # print("ffn_bias: %d" % (ffn_bias)) - - # block_script_name = kwargs.get("block_script_name", "OSA") - # block_class_name = kwargs.get("block_class_name", "OSA_Block") - - # script_name = "." + block_script_name - # package = __import__(script_name, fromlist=True) - block_class = OSA_Block # getattr(package, block_class_name) - group_list = [] - for _ in range(block_num): - temp_res = block_class( - channel_num, - bias, - ffn_bias=ffn_bias, - window_size=window_size, - with_pe=pe, - ) - group_list.append(temp_res) - group_list.append(nn.Conv2d(channel_num, channel_num, 1, 1, 0, bias=bias)) - self.residual_layer = nn.Sequential(*group_list) - esa_channel = max(channel_num // 4, 16) - self.esa = ESA(esa_channel, channel_num) - - def forward(self, x): - out = self.residual_layer(x) - out = out + x - return self.esa(out) diff --git a/comfy_extras/chainner_models/architecture/OmniSR/OmniSR.py b/comfy_extras/chainner_models/architecture/OmniSR/OmniSR.py deleted file mode 100644 index 1e1c3f35e..000000000 --- a/comfy_extras/chainner_models/architecture/OmniSR/OmniSR.py +++ /dev/null @@ -1,143 +0,0 @@ -#!/usr/bin/env python3 -# -*- coding:utf-8 -*- -############################################################# -# File: OmniSR.py -# Created Date: Tuesday April 28th 2022 -# Author: Chen Xuanhong -# Email: chenxuanhongzju@outlook.com -# Last Modified: Sunday, 23rd April 2023 3:06:36 pm -# Modified By: Chen Xuanhong -# Copyright (c) 2020 Shanghai Jiao Tong University -############################################################# - -import math - -import torch -import torch.nn as nn -import torch.nn.functional as F - -from .OSAG import OSAG -from .pixelshuffle import pixelshuffle_block - - -class OmniSR(nn.Module): - def __init__( - self, - state_dict, - **kwargs, - ): - super(OmniSR, self).__init__() - self.state = state_dict - - bias = True # Fine to assume this for now - block_num = 1 # Fine to assume this for now - ffn_bias = True - pe = True - - num_feat = state_dict["input.weight"].shape[0] or 64 - num_in_ch = state_dict["input.weight"].shape[1] or 3 - num_out_ch = num_in_ch # we can just assume this for now. pixelshuffle smh - - pixelshuffle_shape = state_dict["up.0.weight"].shape[0] - up_scale = math.sqrt(pixelshuffle_shape / num_out_ch) - if up_scale - int(up_scale) > 0: - print( - "out_nc is probably different than in_nc, scale calculation might be wrong" - ) - up_scale = int(up_scale) - res_num = 0 - for key in state_dict.keys(): - if "residual_layer" in key: - temp_res_num = int(key.split(".")[1]) - if temp_res_num > res_num: - res_num = temp_res_num - res_num = res_num + 1 # zero-indexed - - residual_layer = [] - self.res_num = res_num - - if ( - "residual_layer.0.residual_layer.0.layer.2.fn.rel_pos_bias.weight" - in state_dict.keys() - ): - rel_pos_bias_weight = state_dict[ - "residual_layer.0.residual_layer.0.layer.2.fn.rel_pos_bias.weight" - ].shape[0] - self.window_size = int((math.sqrt(rel_pos_bias_weight) + 1) / 2) - else: - self.window_size = 8 - - self.up_scale = up_scale - - for _ in range(res_num): - temp_res = OSAG( - channel_num=num_feat, - bias=bias, - block_num=block_num, - ffn_bias=ffn_bias, - window_size=self.window_size, - pe=pe, - ) - residual_layer.append(temp_res) - self.residual_layer = nn.Sequential(*residual_layer) - self.input = nn.Conv2d( - in_channels=num_in_ch, - out_channels=num_feat, - kernel_size=3, - stride=1, - padding=1, - bias=bias, - ) - self.output = nn.Conv2d( - in_channels=num_feat, - out_channels=num_feat, - kernel_size=3, - stride=1, - padding=1, - bias=bias, - ) - self.up = pixelshuffle_block(num_feat, num_out_ch, up_scale, bias=bias) - - # self.tail = pixelshuffle_block(num_feat,num_out_ch,up_scale,bias=bias) - - # for m in self.modules(): - # if isinstance(m, nn.Conv2d): - # n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels - # m.weight.data.normal_(0, sqrt(2. / n)) - - # chaiNNer specific stuff - self.model_arch = "OmniSR" - self.sub_type = "SR" - self.in_nc = num_in_ch - self.out_nc = num_out_ch - self.num_feat = num_feat - self.scale = up_scale - - self.supports_fp16 = True # TODO: Test this - self.supports_bfp16 = True - self.min_size_restriction = 16 - - self.load_state_dict(state_dict, strict=False) - - def check_image_size(self, x): - _, _, h, w = x.size() - # import pdb; pdb.set_trace() - mod_pad_h = (self.window_size - h % self.window_size) % self.window_size - mod_pad_w = (self.window_size - w % self.window_size) % self.window_size - # x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect') - x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "constant", 0) - return x - - def forward(self, x): - H, W = x.shape[2:] - x = self.check_image_size(x) - - residual = self.input(x) - out = self.residual_layer(residual) - - # origin - out = torch.add(self.output(out), residual) - out = self.up(out) - - out = out[:, :, : H * self.up_scale, : W * self.up_scale] - return out diff --git a/comfy_extras/chainner_models/architecture/OmniSR/esa.py b/comfy_extras/chainner_models/architecture/OmniSR/esa.py deleted file mode 100644 index f9ce7f7a6..000000000 --- a/comfy_extras/chainner_models/architecture/OmniSR/esa.py +++ /dev/null @@ -1,294 +0,0 @@ -#!/usr/bin/env python3 -# -*- coding:utf-8 -*- -############################################################# -# File: esa.py -# Created Date: Tuesday April 28th 2022 -# Author: Chen Xuanhong -# Email: chenxuanhongzju@outlook.com -# Last Modified: Thursday, 20th April 2023 9:28:06 am -# Modified By: Chen Xuanhong -# Copyright (c) 2020 Shanghai Jiao Tong University -############################################################# - -import torch -import torch.nn as nn -import torch.nn.functional as F - -from .layernorm import LayerNorm2d - - -def moment(x, dim=(2, 3), k=2): - assert len(x.size()) == 4 - mean = torch.mean(x, dim=dim).unsqueeze(-1).unsqueeze(-1) - mk = (1 / (x.size(2) * x.size(3))) * torch.sum(torch.pow(x - mean, k), dim=dim) - return mk - - -class ESA(nn.Module): - """ - Modification of Enhanced Spatial Attention (ESA), which is proposed by - `Residual Feature Aggregation Network for Image Super-Resolution` - Note: `conv_max` and `conv3_` are NOT used here, so the corresponding codes - are deleted. - """ - - def __init__(self, esa_channels, n_feats, conv=nn.Conv2d): - super(ESA, self).__init__() - f = esa_channels - self.conv1 = conv(n_feats, f, kernel_size=1) - self.conv_f = conv(f, f, kernel_size=1) - self.conv2 = conv(f, f, kernel_size=3, stride=2, padding=0) - self.conv3 = conv(f, f, kernel_size=3, padding=1) - self.conv4 = conv(f, n_feats, kernel_size=1) - self.sigmoid = nn.Sigmoid() - self.relu = nn.ReLU(inplace=True) - - def forward(self, x): - c1_ = self.conv1(x) - c1 = self.conv2(c1_) - v_max = F.max_pool2d(c1, kernel_size=7, stride=3) - c3 = self.conv3(v_max) - c3 = F.interpolate( - c3, (x.size(2), x.size(3)), mode="bilinear", align_corners=False - ) - cf = self.conv_f(c1_) - c4 = self.conv4(c3 + cf) - m = self.sigmoid(c4) - return x * m - - -class LK_ESA(nn.Module): - def __init__( - self, esa_channels, n_feats, conv=nn.Conv2d, kernel_expand=1, bias=True - ): - super(LK_ESA, self).__init__() - f = esa_channels - self.conv1 = conv(n_feats, f, kernel_size=1) - self.conv_f = conv(f, f, kernel_size=1) - - kernel_size = 17 - kernel_expand = kernel_expand - padding = kernel_size // 2 - - self.vec_conv = nn.Conv2d( - in_channels=f * kernel_expand, - out_channels=f * kernel_expand, - kernel_size=(1, kernel_size), - padding=(0, padding), - groups=2, - bias=bias, - ) - self.vec_conv3x1 = nn.Conv2d( - in_channels=f * kernel_expand, - out_channels=f * kernel_expand, - kernel_size=(1, 3), - padding=(0, 1), - groups=2, - bias=bias, - ) - - self.hor_conv = nn.Conv2d( - in_channels=f * kernel_expand, - out_channels=f * kernel_expand, - kernel_size=(kernel_size, 1), - padding=(padding, 0), - groups=2, - bias=bias, - ) - self.hor_conv1x3 = nn.Conv2d( - in_channels=f * kernel_expand, - out_channels=f * kernel_expand, - kernel_size=(3, 1), - padding=(1, 0), - groups=2, - bias=bias, - ) - - self.conv4 = conv(f, n_feats, kernel_size=1) - self.sigmoid = nn.Sigmoid() - self.relu = nn.ReLU(inplace=True) - - def forward(self, x): - c1_ = self.conv1(x) - - res = self.vec_conv(c1_) + self.vec_conv3x1(c1_) - res = self.hor_conv(res) + self.hor_conv1x3(res) - - cf = self.conv_f(c1_) - c4 = self.conv4(res + cf) - m = self.sigmoid(c4) - return x * m - - -class LK_ESA_LN(nn.Module): - def __init__( - self, esa_channels, n_feats, conv=nn.Conv2d, kernel_expand=1, bias=True - ): - super(LK_ESA_LN, self).__init__() - f = esa_channels - self.conv1 = conv(n_feats, f, kernel_size=1) - self.conv_f = conv(f, f, kernel_size=1) - - kernel_size = 17 - kernel_expand = kernel_expand - padding = kernel_size // 2 - - self.norm = LayerNorm2d(n_feats) - - self.vec_conv = nn.Conv2d( - in_channels=f * kernel_expand, - out_channels=f * kernel_expand, - kernel_size=(1, kernel_size), - padding=(0, padding), - groups=2, - bias=bias, - ) - self.vec_conv3x1 = nn.Conv2d( - in_channels=f * kernel_expand, - out_channels=f * kernel_expand, - kernel_size=(1, 3), - padding=(0, 1), - groups=2, - bias=bias, - ) - - self.hor_conv = nn.Conv2d( - in_channels=f * kernel_expand, - out_channels=f * kernel_expand, - kernel_size=(kernel_size, 1), - padding=(padding, 0), - groups=2, - bias=bias, - ) - self.hor_conv1x3 = nn.Conv2d( - in_channels=f * kernel_expand, - out_channels=f * kernel_expand, - kernel_size=(3, 1), - padding=(1, 0), - groups=2, - bias=bias, - ) - - self.conv4 = conv(f, n_feats, kernel_size=1) - self.sigmoid = nn.Sigmoid() - self.relu = nn.ReLU(inplace=True) - - def forward(self, x): - c1_ = self.norm(x) - c1_ = self.conv1(c1_) - - res = self.vec_conv(c1_) + self.vec_conv3x1(c1_) - res = self.hor_conv(res) + self.hor_conv1x3(res) - - cf = self.conv_f(c1_) - c4 = self.conv4(res + cf) - m = self.sigmoid(c4) - return x * m - - -class AdaGuidedFilter(nn.Module): - def __init__( - self, esa_channels, n_feats, conv=nn.Conv2d, kernel_expand=1, bias=True - ): - super(AdaGuidedFilter, self).__init__() - - self.gap = nn.AdaptiveAvgPool2d(1) - self.fc = nn.Conv2d( - in_channels=n_feats, - out_channels=1, - kernel_size=1, - padding=0, - stride=1, - groups=1, - bias=True, - ) - - self.r = 5 - - def box_filter(self, x, r): - channel = x.shape[1] - kernel_size = 2 * r + 1 - weight = 1.0 / (kernel_size**2) - box_kernel = weight * torch.ones( - (channel, 1, kernel_size, kernel_size), dtype=torch.float32, device=x.device - ) - output = F.conv2d(x, weight=box_kernel, stride=1, padding=r, groups=channel) - return output - - def forward(self, x): - _, _, H, W = x.shape - N = self.box_filter( - torch.ones((1, 1, H, W), dtype=x.dtype, device=x.device), self.r - ) - - # epsilon = self.fc(self.gap(x)) - # epsilon = torch.pow(epsilon, 2) - epsilon = 1e-2 - - mean_x = self.box_filter(x, self.r) / N - var_x = self.box_filter(x * x, self.r) / N - mean_x * mean_x - - A = var_x / (var_x + epsilon) - b = (1 - A) * mean_x - m = A * x + b - - # mean_A = self.box_filter(A, self.r) / N - # mean_b = self.box_filter(b, self.r) / N - # m = mean_A * x + mean_b - return x * m - - -class AdaConvGuidedFilter(nn.Module): - def __init__( - self, esa_channels, n_feats, conv=nn.Conv2d, kernel_expand=1, bias=True - ): - super(AdaConvGuidedFilter, self).__init__() - f = esa_channels - - self.conv_f = conv(f, f, kernel_size=1) - - kernel_size = 17 - kernel_expand = kernel_expand - padding = kernel_size // 2 - - self.vec_conv = nn.Conv2d( - in_channels=f, - out_channels=f, - kernel_size=(1, kernel_size), - padding=(0, padding), - groups=f, - bias=bias, - ) - - self.hor_conv = nn.Conv2d( - in_channels=f, - out_channels=f, - kernel_size=(kernel_size, 1), - padding=(padding, 0), - groups=f, - bias=bias, - ) - - self.gap = nn.AdaptiveAvgPool2d(1) - self.fc = nn.Conv2d( - in_channels=f, - out_channels=f, - kernel_size=1, - padding=0, - stride=1, - groups=1, - bias=True, - ) - - def forward(self, x): - y = self.vec_conv(x) - y = self.hor_conv(y) - - sigma = torch.pow(y, 2) - epsilon = self.fc(self.gap(y)) - - weight = sigma / (sigma + epsilon) - - m = weight * x + (1 - weight) - - return x * m diff --git a/comfy_extras/chainner_models/architecture/OmniSR/layernorm.py b/comfy_extras/chainner_models/architecture/OmniSR/layernorm.py deleted file mode 100644 index 731a25f75..000000000 --- a/comfy_extras/chainner_models/architecture/OmniSR/layernorm.py +++ /dev/null @@ -1,70 +0,0 @@ -#!/usr/bin/env python3 -# -*- coding:utf-8 -*- -############################################################# -# File: layernorm.py -# Created Date: Tuesday April 28th 2022 -# Author: Chen Xuanhong -# Email: chenxuanhongzju@outlook.com -# Last Modified: Thursday, 20th April 2023 9:28:20 am -# Modified By: Chen Xuanhong -# Copyright (c) 2020 Shanghai Jiao Tong University -############################################################# - -import torch -import torch.nn as nn - - -class LayerNormFunction(torch.autograd.Function): - @staticmethod - def forward(ctx, x, weight, bias, eps): - ctx.eps = eps - N, C, H, W = x.size() - mu = x.mean(1, keepdim=True) - var = (x - mu).pow(2).mean(1, keepdim=True) - y = (x - mu) / (var + eps).sqrt() - ctx.save_for_backward(y, var, weight) - y = weight.view(1, C, 1, 1) * y + bias.view(1, C, 1, 1) - return y - - @staticmethod - def backward(ctx, grad_output): - eps = ctx.eps - - N, C, H, W = grad_output.size() - y, var, weight = ctx.saved_variables - g = grad_output * weight.view(1, C, 1, 1) - mean_g = g.mean(dim=1, keepdim=True) - - mean_gy = (g * y).mean(dim=1, keepdim=True) - gx = 1.0 / torch.sqrt(var + eps) * (g - y * mean_gy - mean_g) - return ( - gx, - (grad_output * y).sum(dim=3).sum(dim=2).sum(dim=0), - grad_output.sum(dim=3).sum(dim=2).sum(dim=0), - None, - ) - - -class LayerNorm2d(nn.Module): - def __init__(self, channels, eps=1e-6): - super(LayerNorm2d, self).__init__() - self.register_parameter("weight", nn.Parameter(torch.ones(channels))) - self.register_parameter("bias", nn.Parameter(torch.zeros(channels))) - self.eps = eps - - def forward(self, x): - return LayerNormFunction.apply(x, self.weight, self.bias, self.eps) - - -class GRN(nn.Module): - """GRN (Global Response Normalization) layer""" - - def __init__(self, dim): - super().__init__() - self.gamma = nn.Parameter(torch.zeros(1, dim, 1, 1)) - self.beta = nn.Parameter(torch.zeros(1, dim, 1, 1)) - - def forward(self, x): - Gx = torch.norm(x, p=2, dim=(2, 3), keepdim=True) - Nx = Gx / (Gx.mean(dim=1, keepdim=True) + 1e-6) - return self.gamma * (x * Nx) + self.beta + x diff --git a/comfy_extras/chainner_models/architecture/OmniSR/pixelshuffle.py b/comfy_extras/chainner_models/architecture/OmniSR/pixelshuffle.py deleted file mode 100644 index 4260fb7c9..000000000 --- a/comfy_extras/chainner_models/architecture/OmniSR/pixelshuffle.py +++ /dev/null @@ -1,31 +0,0 @@ -#!/usr/bin/env python3 -# -*- coding:utf-8 -*- -############################################################# -# File: pixelshuffle.py -# Created Date: Friday July 1st 2022 -# Author: Chen Xuanhong -# Email: chenxuanhongzju@outlook.com -# Last Modified: Friday, 1st July 2022 10:18:39 am -# Modified By: Chen Xuanhong -# Copyright (c) 2022 Shanghai Jiao Tong University -############################################################# - -import torch.nn as nn - - -def pixelshuffle_block( - in_channels, out_channels, upscale_factor=2, kernel_size=3, bias=False -): - """ - Upsample features according to `upscale_factor`. - """ - padding = kernel_size // 2 - conv = nn.Conv2d( - in_channels, - out_channels * (upscale_factor**2), - kernel_size, - padding=1, - bias=bias, - ) - pixel_shuffle = nn.PixelShuffle(upscale_factor) - return nn.Sequential(*[conv, pixel_shuffle]) diff --git a/comfy_extras/chainner_models/architecture/RRDB.py b/comfy_extras/chainner_models/architecture/RRDB.py deleted file mode 100644 index b50db7c24..000000000 --- a/comfy_extras/chainner_models/architecture/RRDB.py +++ /dev/null @@ -1,296 +0,0 @@ -#!/usr/bin/env python3 -# -*- coding: utf-8 -*- - -import functools -import math -import re -from collections import OrderedDict - -import torch -import torch.nn as nn -import torch.nn.functional as F - -from . import block as B - - -# Borrowed from https://github.com/rlaphoenix/VSGAN/blob/master/vsgan/archs/ESRGAN.py -# Which enhanced stuff that was already here -class RRDBNet(nn.Module): - def __init__( - self, - state_dict, - norm=None, - act: str = "leakyrelu", - upsampler: str = "upconv", - mode: B.ConvMode = "CNA", - ) -> None: - """ - ESRGAN - Enhanced Super-Resolution Generative Adversarial Networks. - By Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, - and Chen Change Loy. - This is old-arch Residual in Residual Dense Block Network and is not - the newest revision that's available at github.com/xinntao/ESRGAN. - This is on purpose, the newest Network has severely limited the - potential use of the Network with no benefits. - This network supports model files from both new and old-arch. - Args: - norm: Normalization layer - act: Activation layer - upsampler: Upsample layer. upconv, pixel_shuffle - mode: Convolution mode - """ - super(RRDBNet, self).__init__() - self.model_arch = "ESRGAN" - self.sub_type = "SR" - - self.state = state_dict - self.norm = norm - self.act = act - self.upsampler = upsampler - self.mode = mode - - self.state_map = { - # currently supports old, new, and newer RRDBNet arch models - # ESRGAN, BSRGAN/RealSR, Real-ESRGAN - "model.0.weight": ("conv_first.weight",), - "model.0.bias": ("conv_first.bias",), - "model.1.sub./NB/.weight": ("trunk_conv.weight", "conv_body.weight"), - "model.1.sub./NB/.bias": ("trunk_conv.bias", "conv_body.bias"), - r"model.1.sub.\1.RDB\2.conv\3.0.\4": ( - r"RRDB_trunk\.(\d+)\.RDB(\d)\.conv(\d+)\.(weight|bias)", - r"body\.(\d+)\.rdb(\d)\.conv(\d+)\.(weight|bias)", - ), - } - if "params_ema" in self.state: - self.state = self.state["params_ema"] - # self.model_arch = "RealESRGAN" - self.num_blocks = self.get_num_blocks() - self.plus = any("conv1x1" in k for k in self.state.keys()) - if self.plus: - self.model_arch = "ESRGAN+" - - self.state = self.new_to_old_arch(self.state) - - self.key_arr = list(self.state.keys()) - - self.in_nc: int = self.state[self.key_arr[0]].shape[1] - self.out_nc: int = self.state[self.key_arr[-1]].shape[0] - - self.scale: int = self.get_scale() - self.num_filters: int = self.state[self.key_arr[0]].shape[0] - - c2x2 = False - if self.state["model.0.weight"].shape[-2] == 2: - c2x2 = True - self.scale = round(math.sqrt(self.scale / 4)) - self.model_arch = "ESRGAN-2c2" - - self.supports_fp16 = True - self.supports_bfp16 = True - self.min_size_restriction = None - - # Detect if pixelunshuffle was used (Real-ESRGAN) - if self.in_nc in (self.out_nc * 4, self.out_nc * 16) and self.out_nc in ( - self.in_nc / 4, - self.in_nc / 16, - ): - self.shuffle_factor = int(math.sqrt(self.in_nc / self.out_nc)) - else: - self.shuffle_factor = None - - upsample_block = { - "upconv": B.upconv_block, - "pixel_shuffle": B.pixelshuffle_block, - }.get(self.upsampler) - if upsample_block is None: - raise NotImplementedError(f"Upsample mode [{self.upsampler}] is not found") - - if self.scale == 3: - upsample_blocks = upsample_block( - in_nc=self.num_filters, - out_nc=self.num_filters, - upscale_factor=3, - act_type=self.act, - c2x2=c2x2, - ) - else: - upsample_blocks = [ - upsample_block( - in_nc=self.num_filters, - out_nc=self.num_filters, - act_type=self.act, - c2x2=c2x2, - ) - for _ in range(int(math.log(self.scale, 2))) - ] - - self.model = B.sequential( - # fea conv - B.conv_block( - in_nc=self.in_nc, - out_nc=self.num_filters, - kernel_size=3, - norm_type=None, - act_type=None, - c2x2=c2x2, - ), - B.ShortcutBlock( - B.sequential( - # rrdb blocks - *[ - B.RRDB( - nf=self.num_filters, - kernel_size=3, - gc=32, - stride=1, - bias=True, - pad_type="zero", - norm_type=self.norm, - act_type=self.act, - mode="CNA", - plus=self.plus, - c2x2=c2x2, - ) - for _ in range(self.num_blocks) - ], - # lr conv - B.conv_block( - in_nc=self.num_filters, - out_nc=self.num_filters, - kernel_size=3, - norm_type=self.norm, - act_type=None, - mode=self.mode, - c2x2=c2x2, - ), - ) - ), - *upsample_blocks, - # hr_conv0 - B.conv_block( - in_nc=self.num_filters, - out_nc=self.num_filters, - kernel_size=3, - norm_type=None, - act_type=self.act, - c2x2=c2x2, - ), - # hr_conv1 - B.conv_block( - in_nc=self.num_filters, - out_nc=self.out_nc, - kernel_size=3, - norm_type=None, - act_type=None, - c2x2=c2x2, - ), - ) - - # Adjust these properties for calculations outside of the model - if self.shuffle_factor: - self.in_nc //= self.shuffle_factor**2 - self.scale //= self.shuffle_factor - - self.load_state_dict(self.state, strict=False) - - def new_to_old_arch(self, state): - """Convert a new-arch model state dictionary to an old-arch dictionary.""" - if "params_ema" in state: - state = state["params_ema"] - - if "conv_first.weight" not in state: - # model is already old arch, this is a loose check, but should be sufficient - return state - - # add nb to state keys - for kind in ("weight", "bias"): - self.state_map[f"model.1.sub.{self.num_blocks}.{kind}"] = self.state_map[ - f"model.1.sub./NB/.{kind}" - ] - del self.state_map[f"model.1.sub./NB/.{kind}"] - - old_state = OrderedDict() - for old_key, new_keys in self.state_map.items(): - for new_key in new_keys: - if r"\1" in old_key: - for k, v in state.items(): - sub = re.sub(new_key, old_key, k) - if sub != k: - old_state[sub] = v - else: - if new_key in state: - old_state[old_key] = state[new_key] - - # upconv layers - max_upconv = 0 - for key in state.keys(): - match = re.match(r"(upconv|conv_up)(\d)\.(weight|bias)", key) - if match is not None: - _, key_num, key_type = match.groups() - old_state[f"model.{int(key_num) * 3}.{key_type}"] = state[key] - max_upconv = max(max_upconv, int(key_num) * 3) - - # final layers - for key in state.keys(): - if key in ("HRconv.weight", "conv_hr.weight"): - old_state[f"model.{max_upconv + 2}.weight"] = state[key] - elif key in ("HRconv.bias", "conv_hr.bias"): - old_state[f"model.{max_upconv + 2}.bias"] = state[key] - elif key in ("conv_last.weight",): - old_state[f"model.{max_upconv + 4}.weight"] = state[key] - elif key in ("conv_last.bias",): - old_state[f"model.{max_upconv + 4}.bias"] = state[key] - - # Sort by first numeric value of each layer - def compare(item1, item2): - parts1 = item1.split(".") - parts2 = item2.split(".") - int1 = int(parts1[1]) - int2 = int(parts2[1]) - return int1 - int2 - - sorted_keys = sorted(old_state.keys(), key=functools.cmp_to_key(compare)) - - # Rebuild the output dict in the right order - out_dict = OrderedDict((k, old_state[k]) for k in sorted_keys) - - return out_dict - - def get_scale(self, min_part: int = 6) -> int: - n = 0 - for part in list(self.state): - parts = part.split(".")[1:] - if len(parts) == 2: - part_num = int(parts[0]) - if part_num > min_part and parts[1] == "weight": - n += 1 - return 2**n - - def get_num_blocks(self) -> int: - nbs = [] - state_keys = self.state_map[r"model.1.sub.\1.RDB\2.conv\3.0.\4"] + ( - r"model\.\d+\.sub\.(\d+)\.RDB(\d+)\.conv(\d+)\.0\.(weight|bias)", - ) - for state_key in state_keys: - for k in self.state: - m = re.search(state_key, k) - if m: - nbs.append(int(m.group(1))) - if nbs: - break - return max(*nbs) + 1 - - def forward(self, x): - if self.shuffle_factor: - _, _, h, w = x.size() - mod_pad_h = ( - self.shuffle_factor - h % self.shuffle_factor - ) % self.shuffle_factor - mod_pad_w = ( - self.shuffle_factor - w % self.shuffle_factor - ) % self.shuffle_factor - x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect") - x = torch.pixel_unshuffle(x, downscale_factor=self.shuffle_factor) - x = self.model(x) - return x[:, :, : h * self.scale, : w * self.scale] - return self.model(x) diff --git a/comfy_extras/chainner_models/architecture/SCUNet.py b/comfy_extras/chainner_models/architecture/SCUNet.py deleted file mode 100644 index b8354a873..000000000 --- a/comfy_extras/chainner_models/architecture/SCUNet.py +++ /dev/null @@ -1,455 +0,0 @@ -# pylint: skip-file -# ----------------------------------------------------------------------------------- -# SCUNet: Practical Blind Denoising via Swin-Conv-UNet and Data Synthesis, https://arxiv.org/abs/2203.13278 -# Zhang, Kai and Li, Yawei and Liang, Jingyun and Cao, Jiezhang and Zhang, Yulun and Tang, Hao and Timofte, Radu and Van Gool, Luc -# ----------------------------------------------------------------------------------- - -import numpy as np -import torch -import torch.nn as nn -import torch.nn.functional as F -from einops import rearrange -from einops.layers.torch import Rearrange - -from .timm.drop import DropPath -from .timm.weight_init import trunc_normal_ - - -# Borrowed from https://github.com/cszn/SCUNet/blob/main/models/network_scunet.py -class WMSA(nn.Module): - """Self-attention module in Swin Transformer""" - - def __init__(self, input_dim, output_dim, head_dim, window_size, type): - super(WMSA, self).__init__() - self.input_dim = input_dim - self.output_dim = output_dim - self.head_dim = head_dim - self.scale = self.head_dim**-0.5 - self.n_heads = input_dim // head_dim - self.window_size = window_size - self.type = type - self.embedding_layer = nn.Linear(self.input_dim, 3 * self.input_dim, bias=True) - - self.relative_position_params = nn.Parameter( - torch.zeros((2 * window_size - 1) * (2 * window_size - 1), self.n_heads) - ) - # TODO recover - # self.relative_position_params = nn.Parameter(torch.zeros(self.n_heads, 2 * window_size - 1, 2 * window_size -1)) - self.relative_position_params = nn.Parameter( - torch.zeros((2 * window_size - 1) * (2 * window_size - 1), self.n_heads) - ) - - self.linear = nn.Linear(self.input_dim, self.output_dim) - - trunc_normal_(self.relative_position_params, std=0.02) - self.relative_position_params = torch.nn.Parameter( - self.relative_position_params.view( - 2 * window_size - 1, 2 * window_size - 1, self.n_heads - ) - .transpose(1, 2) - .transpose(0, 1) - ) - - def generate_mask(self, h, w, p, shift): - """generating the mask of SW-MSA - Args: - shift: shift parameters in CyclicShift. - Returns: - attn_mask: should be (1 1 w p p), - """ - # supporting square. - attn_mask = torch.zeros( - h, - w, - p, - p, - p, - p, - dtype=torch.bool, - device=self.relative_position_params.device, - ) - if self.type == "W": - return attn_mask - - s = p - shift - attn_mask[-1, :, :s, :, s:, :] = True - attn_mask[-1, :, s:, :, :s, :] = True - attn_mask[:, -1, :, :s, :, s:] = True - attn_mask[:, -1, :, s:, :, :s] = True - attn_mask = rearrange( - attn_mask, "w1 w2 p1 p2 p3 p4 -> 1 1 (w1 w2) (p1 p2) (p3 p4)" - ) - return attn_mask - - def forward(self, x): - """Forward pass of Window Multi-head Self-attention module. - Args: - x: input tensor with shape of [b h w c]; - attn_mask: attention mask, fill -inf where the value is True; - Returns: - output: tensor shape [b h w c] - """ - if self.type != "W": - x = torch.roll( - x, - shifts=(-(self.window_size // 2), -(self.window_size // 2)), - dims=(1, 2), - ) - - x = rearrange( - x, - "b (w1 p1) (w2 p2) c -> b w1 w2 p1 p2 c", - p1=self.window_size, - p2=self.window_size, - ) - h_windows = x.size(1) - w_windows = x.size(2) - # square validation - # assert h_windows == w_windows - - x = rearrange( - x, - "b w1 w2 p1 p2 c -> b (w1 w2) (p1 p2) c", - p1=self.window_size, - p2=self.window_size, - ) - qkv = self.embedding_layer(x) - q, k, v = rearrange( - qkv, "b nw np (threeh c) -> threeh b nw np c", c=self.head_dim - ).chunk(3, dim=0) - sim = torch.einsum("hbwpc,hbwqc->hbwpq", q, k) * self.scale - # Adding learnable relative embedding - sim = sim + rearrange(self.relative_embedding(), "h p q -> h 1 1 p q") - # Using Attn Mask to distinguish different subwindows. - if self.type != "W": - attn_mask = self.generate_mask( - h_windows, w_windows, self.window_size, shift=self.window_size // 2 - ) - sim = sim.masked_fill_(attn_mask, float("-inf")) - - probs = nn.functional.softmax(sim, dim=-1) - output = torch.einsum("hbwij,hbwjc->hbwic", probs, v) - output = rearrange(output, "h b w p c -> b w p (h c)") - output = self.linear(output) - output = rearrange( - output, - "b (w1 w2) (p1 p2) c -> b (w1 p1) (w2 p2) c", - w1=h_windows, - p1=self.window_size, - ) - - if self.type != "W": - output = torch.roll( - output, - shifts=(self.window_size // 2, self.window_size // 2), - dims=(1, 2), - ) - - return output - - def relative_embedding(self): - cord = torch.tensor( - np.array( - [ - [i, j] - for i in range(self.window_size) - for j in range(self.window_size) - ] - ) - ) - relation = cord[:, None, :] - cord[None, :, :] + self.window_size - 1 - # negative is allowed - return self.relative_position_params[ - :, relation[:, :, 0].long(), relation[:, :, 1].long() - ] - - -class Block(nn.Module): - def __init__( - self, - input_dim, - output_dim, - head_dim, - window_size, - drop_path, - type="W", - input_resolution=None, - ): - """SwinTransformer Block""" - super(Block, self).__init__() - self.input_dim = input_dim - self.output_dim = output_dim - assert type in ["W", "SW"] - self.type = type - if input_resolution <= window_size: - self.type = "W" - - self.ln1 = nn.LayerNorm(input_dim) - self.msa = WMSA(input_dim, input_dim, head_dim, window_size, self.type) - self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() - self.ln2 = nn.LayerNorm(input_dim) - self.mlp = nn.Sequential( - nn.Linear(input_dim, 4 * input_dim), - nn.GELU(), - nn.Linear(4 * input_dim, output_dim), - ) - - def forward(self, x): - x = x + self.drop_path(self.msa(self.ln1(x))) - x = x + self.drop_path(self.mlp(self.ln2(x))) - return x - - -class ConvTransBlock(nn.Module): - def __init__( - self, - conv_dim, - trans_dim, - head_dim, - window_size, - drop_path, - type="W", - input_resolution=None, - ): - """SwinTransformer and Conv Block""" - super(ConvTransBlock, self).__init__() - self.conv_dim = conv_dim - self.trans_dim = trans_dim - self.head_dim = head_dim - self.window_size = window_size - self.drop_path = drop_path - self.type = type - self.input_resolution = input_resolution - - assert self.type in ["W", "SW"] - if self.input_resolution <= self.window_size: - self.type = "W" - - self.trans_block = Block( - self.trans_dim, - self.trans_dim, - self.head_dim, - self.window_size, - self.drop_path, - self.type, - self.input_resolution, - ) - self.conv1_1 = nn.Conv2d( - self.conv_dim + self.trans_dim, - self.conv_dim + self.trans_dim, - 1, - 1, - 0, - bias=True, - ) - self.conv1_2 = nn.Conv2d( - self.conv_dim + self.trans_dim, - self.conv_dim + self.trans_dim, - 1, - 1, - 0, - bias=True, - ) - - self.conv_block = nn.Sequential( - nn.Conv2d(self.conv_dim, self.conv_dim, 3, 1, 1, bias=False), - nn.ReLU(True), - nn.Conv2d(self.conv_dim, self.conv_dim, 3, 1, 1, bias=False), - ) - - def forward(self, x): - conv_x, trans_x = torch.split( - self.conv1_1(x), (self.conv_dim, self.trans_dim), dim=1 - ) - conv_x = self.conv_block(conv_x) + conv_x - trans_x = Rearrange("b c h w -> b h w c")(trans_x) - trans_x = self.trans_block(trans_x) - trans_x = Rearrange("b h w c -> b c h w")(trans_x) - res = self.conv1_2(torch.cat((conv_x, trans_x), dim=1)) - x = x + res - - return x - - -class SCUNet(nn.Module): - def __init__( - self, - state_dict, - in_nc=3, - config=[4, 4, 4, 4, 4, 4, 4], - dim=64, - drop_path_rate=0.0, - input_resolution=256, - ): - super(SCUNet, self).__init__() - self.model_arch = "SCUNet" - self.sub_type = "SR" - - self.num_filters: int = 0 - - self.state = state_dict - self.config = config - self.dim = dim - self.head_dim = 32 - self.window_size = 8 - - self.in_nc = in_nc - self.out_nc = self.in_nc - self.scale = 1 - self.supports_fp16 = True - - # drop path rate for each layer - dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(config))] - - self.m_head = [nn.Conv2d(in_nc, dim, 3, 1, 1, bias=False)] - - begin = 0 - self.m_down1 = [ - ConvTransBlock( - dim // 2, - dim // 2, - self.head_dim, - self.window_size, - dpr[i + begin], - "W" if not i % 2 else "SW", - input_resolution, - ) - for i in range(config[0]) - ] + [nn.Conv2d(dim, 2 * dim, 2, 2, 0, bias=False)] - - begin += config[0] - self.m_down2 = [ - ConvTransBlock( - dim, - dim, - self.head_dim, - self.window_size, - dpr[i + begin], - "W" if not i % 2 else "SW", - input_resolution // 2, - ) - for i in range(config[1]) - ] + [nn.Conv2d(2 * dim, 4 * dim, 2, 2, 0, bias=False)] - - begin += config[1] - self.m_down3 = [ - ConvTransBlock( - 2 * dim, - 2 * dim, - self.head_dim, - self.window_size, - dpr[i + begin], - "W" if not i % 2 else "SW", - input_resolution // 4, - ) - for i in range(config[2]) - ] + [nn.Conv2d(4 * dim, 8 * dim, 2, 2, 0, bias=False)] - - begin += config[2] - self.m_body = [ - ConvTransBlock( - 4 * dim, - 4 * dim, - self.head_dim, - self.window_size, - dpr[i + begin], - "W" if not i % 2 else "SW", - input_resolution // 8, - ) - for i in range(config[3]) - ] - - begin += config[3] - self.m_up3 = [ - nn.ConvTranspose2d(8 * dim, 4 * dim, 2, 2, 0, bias=False), - ] + [ - ConvTransBlock( - 2 * dim, - 2 * dim, - self.head_dim, - self.window_size, - dpr[i + begin], - "W" if not i % 2 else "SW", - input_resolution // 4, - ) - for i in range(config[4]) - ] - - begin += config[4] - self.m_up2 = [ - nn.ConvTranspose2d(4 * dim, 2 * dim, 2, 2, 0, bias=False), - ] + [ - ConvTransBlock( - dim, - dim, - self.head_dim, - self.window_size, - dpr[i + begin], - "W" if not i % 2 else "SW", - input_resolution // 2, - ) - for i in range(config[5]) - ] - - begin += config[5] - self.m_up1 = [ - nn.ConvTranspose2d(2 * dim, dim, 2, 2, 0, bias=False), - ] + [ - ConvTransBlock( - dim // 2, - dim // 2, - self.head_dim, - self.window_size, - dpr[i + begin], - "W" if not i % 2 else "SW", - input_resolution, - ) - for i in range(config[6]) - ] - - self.m_tail = [nn.Conv2d(dim, in_nc, 3, 1, 1, bias=False)] - - self.m_head = nn.Sequential(*self.m_head) - self.m_down1 = nn.Sequential(*self.m_down1) - self.m_down2 = nn.Sequential(*self.m_down2) - self.m_down3 = nn.Sequential(*self.m_down3) - self.m_body = nn.Sequential(*self.m_body) - self.m_up3 = nn.Sequential(*self.m_up3) - self.m_up2 = nn.Sequential(*self.m_up2) - self.m_up1 = nn.Sequential(*self.m_up1) - self.m_tail = nn.Sequential(*self.m_tail) - # self.apply(self._init_weights) - self.load_state_dict(state_dict, strict=True) - - def check_image_size(self, x): - _, _, h, w = x.size() - mod_pad_h = (64 - h % 64) % 64 - mod_pad_w = (64 - w % 64) % 64 - x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect") - return x - - def forward(self, x0): - h, w = x0.size()[-2:] - x0 = self.check_image_size(x0) - - x1 = self.m_head(x0) - x2 = self.m_down1(x1) - x3 = self.m_down2(x2) - x4 = self.m_down3(x3) - x = self.m_body(x4) - x = self.m_up3(x + x4) - x = self.m_up2(x + x3) - x = self.m_up1(x + x2) - x = self.m_tail(x + x1) - - x = x[:, :, :h, :w] - return x - - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=0.02) - if m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.LayerNorm): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) diff --git a/comfy_extras/chainner_models/architecture/SPSR.py b/comfy_extras/chainner_models/architecture/SPSR.py deleted file mode 100644 index c3cefff19..000000000 --- a/comfy_extras/chainner_models/architecture/SPSR.py +++ /dev/null @@ -1,383 +0,0 @@ -#!/usr/bin/env python3 -# -*- coding: utf-8 -*- - -import math - -import torch -import torch.nn as nn -import torch.nn.functional as F - -from . import block as B - - -class Get_gradient_nopadding(nn.Module): - def __init__(self): - super(Get_gradient_nopadding, self).__init__() - kernel_v = [[0, -1, 0], [0, 0, 0], [0, 1, 0]] - kernel_h = [[0, 0, 0], [-1, 0, 1], [0, 0, 0]] - kernel_h = torch.FloatTensor(kernel_h).unsqueeze(0).unsqueeze(0) - kernel_v = torch.FloatTensor(kernel_v).unsqueeze(0).unsqueeze(0) - self.weight_h = nn.Parameter(data=kernel_h, requires_grad=False) # type: ignore - - self.weight_v = nn.Parameter(data=kernel_v, requires_grad=False) # type: ignore - - def forward(self, x): - x_list = [] - for i in range(x.shape[1]): - x_i = x[:, i] - x_i_v = F.conv2d(x_i.unsqueeze(1), self.weight_v, padding=1) - x_i_h = F.conv2d(x_i.unsqueeze(1), self.weight_h, padding=1) - x_i = torch.sqrt(torch.pow(x_i_v, 2) + torch.pow(x_i_h, 2) + 1e-6) - x_list.append(x_i) - - x = torch.cat(x_list, dim=1) - - return x - - -class SPSRNet(nn.Module): - def __init__( - self, - state_dict, - norm=None, - act: str = "leakyrelu", - upsampler: str = "upconv", - mode: B.ConvMode = "CNA", - ): - super(SPSRNet, self).__init__() - self.model_arch = "SPSR" - self.sub_type = "SR" - - self.state = state_dict - self.norm = norm - self.act = act - self.upsampler = upsampler - self.mode = mode - - self.num_blocks = self.get_num_blocks() - - self.in_nc: int = self.state["model.0.weight"].shape[1] - self.out_nc: int = self.state["f_HR_conv1.0.bias"].shape[0] - - self.scale = self.get_scale(4) - self.num_filters: int = self.state["model.0.weight"].shape[0] - - self.supports_fp16 = True - self.supports_bfp16 = True - self.min_size_restriction = None - - n_upscale = int(math.log(self.scale, 2)) - if self.scale == 3: - n_upscale = 1 - - fea_conv = B.conv_block( - self.in_nc, self.num_filters, kernel_size=3, norm_type=None, act_type=None - ) - rb_blocks = [ - B.RRDB( - self.num_filters, - kernel_size=3, - gc=32, - stride=1, - bias=True, - pad_type="zero", - norm_type=norm, - act_type=act, - mode="CNA", - ) - for _ in range(self.num_blocks) - ] - LR_conv = B.conv_block( - self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=norm, - act_type=None, - mode=mode, - ) - - if upsampler == "upconv": - upsample_block = B.upconv_block - elif upsampler == "pixelshuffle": - upsample_block = B.pixelshuffle_block - else: - raise NotImplementedError(f"upsample mode [{upsampler}] is not found") - if self.scale == 3: - a_upsampler = upsample_block( - self.num_filters, self.num_filters, 3, act_type=act - ) - else: - a_upsampler = [ - upsample_block(self.num_filters, self.num_filters, act_type=act) - for _ in range(n_upscale) - ] - self.HR_conv0_new = B.conv_block( - self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=None, - act_type=act, - ) - self.HR_conv1_new = B.conv_block( - self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=None, - act_type=None, - ) - - self.model = B.sequential( - fea_conv, - B.ShortcutBlockSPSR(B.sequential(*rb_blocks, LR_conv)), - *a_upsampler, - self.HR_conv0_new, - ) - - self.get_g_nopadding = Get_gradient_nopadding() - - self.b_fea_conv = B.conv_block( - self.in_nc, self.num_filters, kernel_size=3, norm_type=None, act_type=None - ) - - self.b_concat_1 = B.conv_block( - 2 * self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=None, - act_type=None, - ) - self.b_block_1 = B.RRDB( - self.num_filters * 2, - kernel_size=3, - gc=32, - stride=1, - bias=True, - pad_type="zero", - norm_type=norm, - act_type=act, - mode="CNA", - ) - - self.b_concat_2 = B.conv_block( - 2 * self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=None, - act_type=None, - ) - self.b_block_2 = B.RRDB( - self.num_filters * 2, - kernel_size=3, - gc=32, - stride=1, - bias=True, - pad_type="zero", - norm_type=norm, - act_type=act, - mode="CNA", - ) - - self.b_concat_3 = B.conv_block( - 2 * self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=None, - act_type=None, - ) - self.b_block_3 = B.RRDB( - self.num_filters * 2, - kernel_size=3, - gc=32, - stride=1, - bias=True, - pad_type="zero", - norm_type=norm, - act_type=act, - mode="CNA", - ) - - self.b_concat_4 = B.conv_block( - 2 * self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=None, - act_type=None, - ) - self.b_block_4 = B.RRDB( - self.num_filters * 2, - kernel_size=3, - gc=32, - stride=1, - bias=True, - pad_type="zero", - norm_type=norm, - act_type=act, - mode="CNA", - ) - - self.b_LR_conv = B.conv_block( - self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=norm, - act_type=None, - mode=mode, - ) - - if upsampler == "upconv": - upsample_block = B.upconv_block - elif upsampler == "pixelshuffle": - upsample_block = B.pixelshuffle_block - else: - raise NotImplementedError(f"upsample mode [{upsampler}] is not found") - if self.scale == 3: - b_upsampler = upsample_block( - self.num_filters, self.num_filters, 3, act_type=act - ) - else: - b_upsampler = [ - upsample_block(self.num_filters, self.num_filters, act_type=act) - for _ in range(n_upscale) - ] - - b_HR_conv0 = B.conv_block( - self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=None, - act_type=act, - ) - b_HR_conv1 = B.conv_block( - self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=None, - act_type=None, - ) - - self.b_module = B.sequential(*b_upsampler, b_HR_conv0, b_HR_conv1) - - self.conv_w = B.conv_block( - self.num_filters, self.out_nc, kernel_size=1, norm_type=None, act_type=None - ) - - self.f_concat = B.conv_block( - self.num_filters * 2, - self.num_filters, - kernel_size=3, - norm_type=None, - act_type=None, - ) - - self.f_block = B.RRDB( - self.num_filters * 2, - kernel_size=3, - gc=32, - stride=1, - bias=True, - pad_type="zero", - norm_type=norm, - act_type=act, - mode="CNA", - ) - - self.f_HR_conv0 = B.conv_block( - self.num_filters, - self.num_filters, - kernel_size=3, - norm_type=None, - act_type=act, - ) - self.f_HR_conv1 = B.conv_block( - self.num_filters, self.out_nc, kernel_size=3, norm_type=None, act_type=None - ) - - self.load_state_dict(self.state, strict=False) - - def get_scale(self, min_part: int = 4) -> int: - n = 0 - for part in list(self.state): - parts = part.split(".") - if len(parts) == 3: - part_num = int(parts[1]) - if part_num > min_part and parts[0] == "model" and parts[2] == "weight": - n += 1 - return 2**n - - def get_num_blocks(self) -> int: - nb = 0 - for part in list(self.state): - parts = part.split(".") - n_parts = len(parts) - if n_parts == 5 and parts[2] == "sub": - nb = int(parts[3]) - return nb - - def forward(self, x): - x_grad = self.get_g_nopadding(x) - x = self.model[0](x) - - x, block_list = self.model[1](x) - - x_ori = x - for i in range(5): - x = block_list[i](x) - x_fea1 = x - - for i in range(5): - x = block_list[i + 5](x) - x_fea2 = x - - for i in range(5): - x = block_list[i + 10](x) - x_fea3 = x - - for i in range(5): - x = block_list[i + 15](x) - x_fea4 = x - - x = block_list[20:](x) - # short cut - x = x_ori + x - x = self.model[2:](x) - x = self.HR_conv1_new(x) - - x_b_fea = self.b_fea_conv(x_grad) - x_cat_1 = torch.cat([x_b_fea, x_fea1], dim=1) - - x_cat_1 = self.b_block_1(x_cat_1) - x_cat_1 = self.b_concat_1(x_cat_1) - - x_cat_2 = torch.cat([x_cat_1, x_fea2], dim=1) - - x_cat_2 = self.b_block_2(x_cat_2) - x_cat_2 = self.b_concat_2(x_cat_2) - - x_cat_3 = torch.cat([x_cat_2, x_fea3], dim=1) - - x_cat_3 = self.b_block_3(x_cat_3) - x_cat_3 = self.b_concat_3(x_cat_3) - - x_cat_4 = torch.cat([x_cat_3, x_fea4], dim=1) - - x_cat_4 = self.b_block_4(x_cat_4) - x_cat_4 = self.b_concat_4(x_cat_4) - - x_cat_4 = self.b_LR_conv(x_cat_4) - - # short cut - x_cat_4 = x_cat_4 + x_b_fea - x_branch = self.b_module(x_cat_4) - - # x_out_branch = self.conv_w(x_branch) - ######## - x_branch_d = x_branch - x_f_cat = torch.cat([x_branch_d, x], dim=1) - x_f_cat = self.f_block(x_f_cat) - x_out = self.f_concat(x_f_cat) - x_out = self.f_HR_conv0(x_out) - x_out = self.f_HR_conv1(x_out) - - ######### - # return x_out_branch, x_out, x_grad - return x_out diff --git a/comfy_extras/chainner_models/architecture/SRVGG.py b/comfy_extras/chainner_models/architecture/SRVGG.py deleted file mode 100644 index 7a8ec37ae..000000000 --- a/comfy_extras/chainner_models/architecture/SRVGG.py +++ /dev/null @@ -1,114 +0,0 @@ -#!/usr/bin/env python3 -# -*- coding: utf-8 -*- - -import math - -import torch.nn as nn -import torch.nn.functional as F - - -class SRVGGNetCompact(nn.Module): - """A compact VGG-style network structure for super-resolution. - It is a compact network structure, which performs upsampling in the last layer and no convolution is - conducted on the HR feature space. - Args: - num_in_ch (int): Channel number of inputs. Default: 3. - num_out_ch (int): Channel number of outputs. Default: 3. - num_feat (int): Channel number of intermediate features. Default: 64. - num_conv (int): Number of convolution layers in the body network. Default: 16. - upscale (int): Upsampling factor. Default: 4. - act_type (str): Activation type, options: 'relu', 'prelu', 'leakyrelu'. Default: prelu. - """ - - def __init__( - self, - state_dict, - act_type: str = "prelu", - ): - super(SRVGGNetCompact, self).__init__() - self.model_arch = "SRVGG (RealESRGAN)" - self.sub_type = "SR" - - self.act_type = act_type - - self.state = state_dict - - if "params" in self.state: - self.state = self.state["params"] - - self.key_arr = list(self.state.keys()) - - self.in_nc = self.get_in_nc() - self.num_feat = self.get_num_feats() - self.num_conv = self.get_num_conv() - self.out_nc = self.in_nc # :( - self.pixelshuffle_shape = None # Defined in get_scale() - self.scale = self.get_scale() - - self.supports_fp16 = True - self.supports_bfp16 = True - self.min_size_restriction = None - - self.body = nn.ModuleList() - # the first conv - self.body.append(nn.Conv2d(self.in_nc, self.num_feat, 3, 1, 1)) - # the first activation - if act_type == "relu": - activation = nn.ReLU(inplace=True) - elif act_type == "prelu": - activation = nn.PReLU(num_parameters=self.num_feat) - elif act_type == "leakyrelu": - activation = nn.LeakyReLU(negative_slope=0.1, inplace=True) - self.body.append(activation) # type: ignore - - # the body structure - for _ in range(self.num_conv): - self.body.append(nn.Conv2d(self.num_feat, self.num_feat, 3, 1, 1)) - # activation - if act_type == "relu": - activation = nn.ReLU(inplace=True) - elif act_type == "prelu": - activation = nn.PReLU(num_parameters=self.num_feat) - elif act_type == "leakyrelu": - activation = nn.LeakyReLU(negative_slope=0.1, inplace=True) - self.body.append(activation) # type: ignore - - # the last conv - self.body.append(nn.Conv2d(self.num_feat, self.pixelshuffle_shape, 3, 1, 1)) # type: ignore - # upsample - self.upsampler = nn.PixelShuffle(self.scale) - - self.load_state_dict(self.state, strict=False) - - def get_num_conv(self) -> int: - return (int(self.key_arr[-1].split(".")[1]) - 2) // 2 - - def get_num_feats(self) -> int: - return self.state[self.key_arr[0]].shape[0] - - def get_in_nc(self) -> int: - return self.state[self.key_arr[0]].shape[1] - - def get_scale(self) -> int: - self.pixelshuffle_shape = self.state[self.key_arr[-1]].shape[0] - # Assume out_nc is the same as in_nc - # I cant think of a better way to do that - self.out_nc = self.in_nc - scale = math.sqrt(self.pixelshuffle_shape / self.out_nc) - if scale - int(scale) > 0: - print( - "out_nc is probably different than in_nc, scale calculation might be wrong" - ) - scale = int(scale) - return scale - - def forward(self, x): - out = x - for i in range(0, len(self.body)): - out = self.body[i](out) - - out = self.upsampler(out) - # add the nearest upsampled image, so that the network learns the residual - base = F.interpolate(x, scale_factor=self.scale, mode="nearest") - out += base - return out diff --git a/comfy_extras/chainner_models/architecture/SwiftSRGAN.py b/comfy_extras/chainner_models/architecture/SwiftSRGAN.py deleted file mode 100644 index dbb7725b0..000000000 --- a/comfy_extras/chainner_models/architecture/SwiftSRGAN.py +++ /dev/null @@ -1,161 +0,0 @@ -# From https://github.com/Koushik0901/Swift-SRGAN/blob/master/swift-srgan/models.py - -import torch -from torch import nn - - -class SeperableConv2d(nn.Module): - def __init__( - self, in_channels, out_channels, kernel_size, stride=1, padding=1, bias=True - ): - super(SeperableConv2d, self).__init__() - self.depthwise = nn.Conv2d( - in_channels, - in_channels, - kernel_size=kernel_size, - stride=stride, - groups=in_channels, - bias=bias, - padding=padding, - ) - self.pointwise = nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=bias) - - def forward(self, x): - return self.pointwise(self.depthwise(x)) - - -class ConvBlock(nn.Module): - def __init__( - self, - in_channels, - out_channels, - use_act=True, - use_bn=True, - discriminator=False, - **kwargs, - ): - super(ConvBlock, self).__init__() - - self.use_act = use_act - self.cnn = SeperableConv2d(in_channels, out_channels, **kwargs, bias=not use_bn) - self.bn = nn.BatchNorm2d(out_channels) if use_bn else nn.Identity() - self.act = ( - nn.LeakyReLU(0.2, inplace=True) - if discriminator - else nn.PReLU(num_parameters=out_channels) - ) - - def forward(self, x): - return self.act(self.bn(self.cnn(x))) if self.use_act else self.bn(self.cnn(x)) - - -class UpsampleBlock(nn.Module): - def __init__(self, in_channels, scale_factor): - super(UpsampleBlock, self).__init__() - - self.conv = SeperableConv2d( - in_channels, - in_channels * scale_factor**2, - kernel_size=3, - stride=1, - padding=1, - ) - self.ps = nn.PixelShuffle( - scale_factor - ) # (in_channels * 4, H, W) -> (in_channels, H*2, W*2) - self.act = nn.PReLU(num_parameters=in_channels) - - def forward(self, x): - return self.act(self.ps(self.conv(x))) - - -class ResidualBlock(nn.Module): - def __init__(self, in_channels): - super(ResidualBlock, self).__init__() - - self.block1 = ConvBlock( - in_channels, in_channels, kernel_size=3, stride=1, padding=1 - ) - self.block2 = ConvBlock( - in_channels, in_channels, kernel_size=3, stride=1, padding=1, use_act=False - ) - - def forward(self, x): - out = self.block1(x) - out = self.block2(out) - return out + x - - -class Generator(nn.Module): - """Swift-SRGAN Generator - Args: - in_channels (int): number of input image channels. - num_channels (int): number of hidden channels. - num_blocks (int): number of residual blocks. - upscale_factor (int): factor to upscale the image [2x, 4x, 8x]. - Returns: - torch.Tensor: super resolution image - """ - - def __init__( - self, - state_dict, - ): - super(Generator, self).__init__() - self.model_arch = "Swift-SRGAN" - self.sub_type = "SR" - self.state = state_dict - if "model" in self.state: - self.state = self.state["model"] - - self.in_nc: int = self.state["initial.cnn.depthwise.weight"].shape[0] - self.out_nc: int = self.state["final_conv.pointwise.weight"].shape[0] - self.num_filters: int = self.state["initial.cnn.pointwise.weight"].shape[0] - self.num_blocks = len( - set([x.split(".")[1] for x in self.state.keys() if "residual" in x]) - ) - self.scale: int = 2 ** len( - set([x.split(".")[1] for x in self.state.keys() if "upsampler" in x]) - ) - - in_channels = self.in_nc - num_channels = self.num_filters - num_blocks = self.num_blocks - upscale_factor = self.scale - - self.supports_fp16 = True - self.supports_bfp16 = True - self.min_size_restriction = None - - self.initial = ConvBlock( - in_channels, num_channels, kernel_size=9, stride=1, padding=4, use_bn=False - ) - self.residual = nn.Sequential( - *[ResidualBlock(num_channels) for _ in range(num_blocks)] - ) - self.convblock = ConvBlock( - num_channels, - num_channels, - kernel_size=3, - stride=1, - padding=1, - use_act=False, - ) - self.upsampler = nn.Sequential( - *[ - UpsampleBlock(num_channels, scale_factor=2) - for _ in range(upscale_factor // 2) - ] - ) - self.final_conv = SeperableConv2d( - num_channels, in_channels, kernel_size=9, stride=1, padding=4 - ) - - self.load_state_dict(self.state, strict=False) - - def forward(self, x): - initial = self.initial(x) - x = self.residual(initial) - x = self.convblock(x) + initial - x = self.upsampler(x) - return (torch.tanh(self.final_conv(x)) + 1) / 2 diff --git a/comfy_extras/chainner_models/architecture/Swin2SR.py b/comfy_extras/chainner_models/architecture/Swin2SR.py deleted file mode 100644 index cb57ecfc4..000000000 --- a/comfy_extras/chainner_models/architecture/Swin2SR.py +++ /dev/null @@ -1,1377 +0,0 @@ -# pylint: skip-file -# ----------------------------------------------------------------------------------- -# Swin2SR: Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration, https://arxiv.org/abs/2209.11345 -# Written by Conde and Choi et al. -# From: https://raw.githubusercontent.com/mv-lab/swin2sr/main/models/network_swin2sr.py -# ----------------------------------------------------------------------------------- - -import math -import re - -import numpy as np -import torch -import torch.nn as nn -import torch.nn.functional as F -import torch.utils.checkpoint as checkpoint - -# Originally from the timm package -from .timm.drop import DropPath -from .timm.helpers import to_2tuple -from .timm.weight_init import trunc_normal_ - - -class Mlp(nn.Module): - def __init__( - self, - in_features, - hidden_features=None, - out_features=None, - act_layer=nn.GELU, - drop=0.0, - ): - super().__init__() - out_features = out_features or in_features - hidden_features = hidden_features or in_features - self.fc1 = nn.Linear(in_features, hidden_features) - self.act = act_layer() - self.fc2 = nn.Linear(hidden_features, out_features) - self.drop = nn.Dropout(drop) - - def forward(self, x): - x = self.fc1(x) - x = self.act(x) - x = self.drop(x) - x = self.fc2(x) - x = self.drop(x) - return x - - -def window_partition(x, window_size): - """ - Args: - x: (B, H, W, C) - window_size (int): window size - Returns: - windows: (num_windows*B, window_size, window_size, C) - """ - B, H, W, C = x.shape - x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) - windows = ( - x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) - ) - return windows - - -def window_reverse(windows, window_size, H, W): - """ - Args: - windows: (num_windows*B, window_size, window_size, C) - window_size (int): Window size - H (int): Height of image - W (int): Width of image - Returns: - x: (B, H, W, C) - """ - B = int(windows.shape[0] / (H * W / window_size / window_size)) - x = windows.view( - B, H // window_size, W // window_size, window_size, window_size, -1 - ) - x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) - return x - - -class WindowAttention(nn.Module): - r"""Window based multi-head self attention (W-MSA) module with relative position bias. - It supports both of shifted and non-shifted window. - Args: - dim (int): Number of input channels. - window_size (tuple[int]): The height and width of the window. - num_heads (int): Number of attention heads. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 - proj_drop (float, optional): Dropout ratio of output. Default: 0.0 - pretrained_window_size (tuple[int]): The height and width of the window in pre-training. - """ - - def __init__( - self, - dim, - window_size, - num_heads, - qkv_bias=True, - attn_drop=0.0, - proj_drop=0.0, - pretrained_window_size=[0, 0], - ): - super().__init__() - self.dim = dim - self.window_size = window_size # Wh, Ww - self.pretrained_window_size = pretrained_window_size - self.num_heads = num_heads - - self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True) # type: ignore - - # mlp to generate continuous relative position bias - self.cpb_mlp = nn.Sequential( - nn.Linear(2, 512, bias=True), - nn.ReLU(inplace=True), - nn.Linear(512, num_heads, bias=False), - ) - - # get relative_coords_table - relative_coords_h = torch.arange( - -(self.window_size[0] - 1), self.window_size[0], dtype=torch.float32 - ) - relative_coords_w = torch.arange( - -(self.window_size[1] - 1), self.window_size[1], dtype=torch.float32 - ) - relative_coords_table = ( - torch.stack(torch.meshgrid([relative_coords_h, relative_coords_w])) - .permute(1, 2, 0) - .contiguous() - .unsqueeze(0) - ) # 1, 2*Wh-1, 2*Ww-1, 2 - if pretrained_window_size[0] > 0: - relative_coords_table[:, :, :, 0] /= pretrained_window_size[0] - 1 - relative_coords_table[:, :, :, 1] /= pretrained_window_size[1] - 1 - else: - relative_coords_table[:, :, :, 0] /= self.window_size[0] - 1 - relative_coords_table[:, :, :, 1] /= self.window_size[1] - 1 - relative_coords_table *= 8 # normalize to -8, 8 - relative_coords_table = ( - torch.sign(relative_coords_table) - * torch.log2(torch.abs(relative_coords_table) + 1.0) - / np.log2(8) - ) - - self.register_buffer("relative_coords_table", relative_coords_table) - - # get pair-wise relative position index for each token inside the window - coords_h = torch.arange(self.window_size[0]) - coords_w = torch.arange(self.window_size[1]) - coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww - coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww - relative_coords = ( - coords_flatten[:, :, None] - coords_flatten[:, None, :] - ) # 2, Wh*Ww, Wh*Ww - relative_coords = relative_coords.permute( - 1, 2, 0 - ).contiguous() # Wh*Ww, Wh*Ww, 2 - relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0 - relative_coords[:, :, 1] += self.window_size[1] - 1 - relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 - relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww - self.register_buffer("relative_position_index", relative_position_index) - - self.qkv = nn.Linear(dim, dim * 3, bias=False) - if qkv_bias: - self.q_bias = nn.Parameter(torch.zeros(dim)) # type: ignore - self.v_bias = nn.Parameter(torch.zeros(dim)) # type: ignore - else: - self.q_bias = None - self.v_bias = None - self.attn_drop = nn.Dropout(attn_drop) - self.proj = nn.Linear(dim, dim) - self.proj_drop = nn.Dropout(proj_drop) - self.softmax = nn.Softmax(dim=-1) - - def forward(self, x, mask=None): - """ - Args: - x: input features with shape of (num_windows*B, N, C) - mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None - """ - B_, N, C = x.shape - qkv_bias = None - if self.q_bias is not None: - qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias)) # type: ignore - qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias) - qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) - q, k, v = ( - qkv[0], - qkv[1], - qkv[2], - ) # make torchscript happy (cannot use tensor as tuple) - - # cosine attention - attn = F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1) - logit_scale = torch.clamp( - self.logit_scale, - max=torch.log(torch.tensor(1.0 / 0.01)).to(self.logit_scale.device), - ).exp() - attn = attn * logit_scale - - relative_position_bias_table = self.cpb_mlp(self.relative_coords_table).view( - -1, self.num_heads - ) - relative_position_bias = relative_position_bias_table[self.relative_position_index.view(-1)].view( # type: ignore - self.window_size[0] * self.window_size[1], - self.window_size[0] * self.window_size[1], - -1, - ) # Wh*Ww,Wh*Ww,nH - relative_position_bias = relative_position_bias.permute( - 2, 0, 1 - ).contiguous() # nH, Wh*Ww, Wh*Ww - relative_position_bias = 16 * torch.sigmoid(relative_position_bias) - attn = attn + relative_position_bias.unsqueeze(0) - - if mask is not None: - nW = mask.shape[0] - attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze( - 1 - ).unsqueeze(0) - attn = attn.view(-1, self.num_heads, N, N) - attn = self.softmax(attn) - else: - attn = self.softmax(attn) - - attn = self.attn_drop(attn) - - x = (attn @ v).transpose(1, 2).reshape(B_, N, C) - x = self.proj(x) - x = self.proj_drop(x) - return x - - def extra_repr(self) -> str: - return ( - f"dim={self.dim}, window_size={self.window_size}, " - f"pretrained_window_size={self.pretrained_window_size}, num_heads={self.num_heads}" - ) - - def flops(self, N): - # calculate flops for 1 window with token length of N - flops = 0 - # qkv = self.qkv(x) - flops += N * self.dim * 3 * self.dim - # attn = (q @ k.transpose(-2, -1)) - flops += self.num_heads * N * (self.dim // self.num_heads) * N - # x = (attn @ v) - flops += self.num_heads * N * N * (self.dim // self.num_heads) - # x = self.proj(x) - flops += N * self.dim * self.dim - return flops - - -class SwinTransformerBlock(nn.Module): - r"""Swin Transformer Block. - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resulotion. - num_heads (int): Number of attention heads. - window_size (int): Window size. - shift_size (int): Shift size for SW-MSA. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float, optional): Stochastic depth rate. Default: 0.0 - act_layer (nn.Module, optional): Activation layer. Default: nn.GELU - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - pretrained_window_size (int): Window size in pre-training. - """ - - def __init__( - self, - dim, - input_resolution, - num_heads, - window_size=7, - shift_size=0, - mlp_ratio=4.0, - qkv_bias=True, - drop=0.0, - attn_drop=0.0, - drop_path=0.0, - act_layer=nn.GELU, - norm_layer=nn.LayerNorm, - pretrained_window_size=0, - ): - super().__init__() - self.dim = dim - self.input_resolution = input_resolution - self.num_heads = num_heads - self.window_size = window_size - self.shift_size = shift_size - self.mlp_ratio = mlp_ratio - if min(self.input_resolution) <= self.window_size: - # if window size is larger than input resolution, we don't partition windows - self.shift_size = 0 - self.window_size = min(self.input_resolution) - assert ( - 0 <= self.shift_size < self.window_size - ), "shift_size must in 0-window_size" - - self.norm1 = norm_layer(dim) - self.attn = WindowAttention( - dim, - window_size=to_2tuple(self.window_size), - num_heads=num_heads, - qkv_bias=qkv_bias, - attn_drop=attn_drop, - proj_drop=drop, - pretrained_window_size=to_2tuple(pretrained_window_size), - ) - - self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() - self.norm2 = norm_layer(dim) - mlp_hidden_dim = int(dim * mlp_ratio) - self.mlp = Mlp( - in_features=dim, - hidden_features=mlp_hidden_dim, - act_layer=act_layer, - drop=drop, - ) - - if self.shift_size > 0: - attn_mask = self.calculate_mask(self.input_resolution) - else: - attn_mask = None - - self.register_buffer("attn_mask", attn_mask) - - def calculate_mask(self, x_size): - # calculate attention mask for SW-MSA - H, W = x_size - img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1 - h_slices = ( - slice(0, -self.window_size), - slice(-self.window_size, -self.shift_size), - slice(-self.shift_size, None), - ) - w_slices = ( - slice(0, -self.window_size), - slice(-self.window_size, -self.shift_size), - slice(-self.shift_size, None), - ) - cnt = 0 - for h in h_slices: - for w in w_slices: - img_mask[:, h, w, :] = cnt - cnt += 1 - - mask_windows = window_partition( - img_mask, self.window_size - ) # nW, window_size, window_size, 1 - mask_windows = mask_windows.view(-1, self.window_size * self.window_size) - attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) - attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill( - attn_mask == 0, float(0.0) - ) - - return attn_mask - - def forward(self, x, x_size): - H, W = x_size - B, L, C = x.shape - # assert L == H * W, "input feature has wrong size" - - shortcut = x - x = x.view(B, H, W, C) - - # cyclic shift - if self.shift_size > 0: - shifted_x = torch.roll( - x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2) - ) - else: - shifted_x = x - - # partition windows - x_windows = window_partition( - shifted_x, self.window_size - ) # nW*B, window_size, window_size, C - x_windows = x_windows.view( - -1, self.window_size * self.window_size, C - ) # nW*B, window_size*window_size, C - - # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size - if self.input_resolution == x_size: - attn_windows = self.attn( - x_windows, mask=self.attn_mask - ) # nW*B, window_size*window_size, C - else: - attn_windows = self.attn( - x_windows, mask=self.calculate_mask(x_size).to(x.device) - ) - - # merge windows - attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) - shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C - - # reverse cyclic shift - if self.shift_size > 0: - x = torch.roll( - shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2) - ) - else: - x = shifted_x - x = x.view(B, H * W, C) - x = shortcut + self.drop_path(self.norm1(x)) - - # FFN - x = x + self.drop_path(self.norm2(self.mlp(x))) - - return x - - def extra_repr(self) -> str: - return ( - f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " - f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}" - ) - - def flops(self): - flops = 0 - H, W = self.input_resolution - # norm1 - flops += self.dim * H * W - # W-MSA/SW-MSA - nW = H * W / self.window_size / self.window_size - flops += nW * self.attn.flops(self.window_size * self.window_size) - # mlp - flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio - # norm2 - flops += self.dim * H * W - return flops - - -class PatchMerging(nn.Module): - r"""Patch Merging Layer. - Args: - input_resolution (tuple[int]): Resolution of input feature. - dim (int): Number of input channels. - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - """ - - def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm): - super().__init__() - self.input_resolution = input_resolution - self.dim = dim - self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) - self.norm = norm_layer(2 * dim) - - def forward(self, x): - """ - x: B, H*W, C - """ - H, W = self.input_resolution - B, L, C = x.shape - assert L == H * W, "input feature has wrong size" - assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even." - - x = x.view(B, H, W, C) - - x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C - x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C - x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C - x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C - x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C - x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C - - x = self.reduction(x) - x = self.norm(x) - - return x - - def extra_repr(self) -> str: - return f"input_resolution={self.input_resolution}, dim={self.dim}" - - def flops(self): - H, W = self.input_resolution - flops = (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim - flops += H * W * self.dim // 2 - return flops - - -class BasicLayer(nn.Module): - """A basic Swin Transformer layer for one stage. - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resolution. - depth (int): Number of blocks. - num_heads (int): Number of attention heads. - window_size (int): Local window size. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. - pretrained_window_size (int): Local window size in pre-training. - """ - - def __init__( - self, - dim, - input_resolution, - depth, - num_heads, - window_size, - mlp_ratio=4.0, - qkv_bias=True, - drop=0.0, - attn_drop=0.0, - drop_path=0.0, - norm_layer=nn.LayerNorm, - downsample=None, - use_checkpoint=False, - pretrained_window_size=0, - ): - super().__init__() - self.dim = dim - self.input_resolution = input_resolution - self.depth = depth - self.use_checkpoint = use_checkpoint - - # build blocks - self.blocks = nn.ModuleList( - [ - SwinTransformerBlock( - dim=dim, - input_resolution=input_resolution, - num_heads=num_heads, - window_size=window_size, - shift_size=0 if (i % 2 == 0) else window_size // 2, - mlp_ratio=mlp_ratio, - qkv_bias=qkv_bias, - drop=drop, - attn_drop=attn_drop, - drop_path=drop_path[i] - if isinstance(drop_path, list) - else drop_path, - norm_layer=norm_layer, - pretrained_window_size=pretrained_window_size, - ) - for i in range(depth) - ] - ) - - # patch merging layer - if downsample is not None: - self.downsample = downsample( - input_resolution, dim=dim, norm_layer=norm_layer - ) - else: - self.downsample = None - - def forward(self, x, x_size): - for blk in self.blocks: - if self.use_checkpoint: - x = checkpoint.checkpoint(blk, x, x_size) - else: - x = blk(x, x_size) - if self.downsample is not None: - x = self.downsample(x) - return x - - def extra_repr(self) -> str: - return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}" - - def flops(self): - flops = 0 - for blk in self.blocks: - flops += blk.flops() # type: ignore - if self.downsample is not None: - flops += self.downsample.flops() - return flops - - def _init_respostnorm(self): - for blk in self.blocks: - nn.init.constant_(blk.norm1.bias, 0) # type: ignore - nn.init.constant_(blk.norm1.weight, 0) # type: ignore - nn.init.constant_(blk.norm2.bias, 0) # type: ignore - nn.init.constant_(blk.norm2.weight, 0) # type: ignore - - -class PatchEmbed(nn.Module): - r"""Image to Patch Embedding - Args: - img_size (int): Image size. Default: 224. - patch_size (int): Patch token size. Default: 4. - in_chans (int): Number of input image channels. Default: 3. - embed_dim (int): Number of linear projection output channels. Default: 96. - norm_layer (nn.Module, optional): Normalization layer. Default: None - """ - - def __init__( - self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None - ): - super().__init__() - img_size = to_2tuple(img_size) - patch_size = to_2tuple(patch_size) - patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] # type: ignore - self.img_size = img_size - self.patch_size = patch_size - self.patches_resolution = patches_resolution - self.num_patches = patches_resolution[0] * patches_resolution[1] - - self.in_chans = in_chans - self.embed_dim = embed_dim - - self.proj = nn.Conv2d( - in_chans, embed_dim, kernel_size=patch_size, stride=patch_size # type: ignore - ) - if norm_layer is not None: - self.norm = norm_layer(embed_dim) - else: - self.norm = None - - def forward(self, x): - B, C, H, W = x.shape - # FIXME look at relaxing size constraints - # assert H == self.img_size[0] and W == self.img_size[1], - # f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})." - x = self.proj(x).flatten(2).transpose(1, 2) # B Ph*Pw C - if self.norm is not None: - x = self.norm(x) - return x - - def flops(self): - Ho, Wo = self.patches_resolution - flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1]) # type: ignore - if self.norm is not None: - flops += Ho * Wo * self.embed_dim - return flops - - -class RSTB(nn.Module): - """Residual Swin Transformer Block (RSTB). - - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resolution. - depth (int): Number of blocks. - num_heads (int): Number of attention heads. - window_size (int): Local window size. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. - img_size: Input image size. - patch_size: Patch size. - resi_connection: The convolutional block before residual connection. - """ - - def __init__( - self, - dim, - input_resolution, - depth, - num_heads, - window_size, - mlp_ratio=4.0, - qkv_bias=True, - drop=0.0, - attn_drop=0.0, - drop_path=0.0, - norm_layer=nn.LayerNorm, - downsample=None, - use_checkpoint=False, - img_size=224, - patch_size=4, - resi_connection="1conv", - ): - super(RSTB, self).__init__() - - self.dim = dim - self.input_resolution = input_resolution - - self.residual_group = BasicLayer( - dim=dim, - input_resolution=input_resolution, - depth=depth, - num_heads=num_heads, - window_size=window_size, - mlp_ratio=mlp_ratio, - qkv_bias=qkv_bias, - drop=drop, - attn_drop=attn_drop, - drop_path=drop_path, - norm_layer=norm_layer, - downsample=downsample, - use_checkpoint=use_checkpoint, - ) - - if resi_connection == "1conv": - self.conv = nn.Conv2d(dim, dim, 3, 1, 1) - elif resi_connection == "3conv": - # to save parameters and memory - self.conv = nn.Sequential( - nn.Conv2d(dim, dim // 4, 3, 1, 1), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(dim // 4, dim // 4, 1, 1, 0), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(dim // 4, dim, 3, 1, 1), - ) - - self.patch_embed = PatchEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=dim, - embed_dim=dim, - norm_layer=None, - ) - - self.patch_unembed = PatchUnEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=dim, - embed_dim=dim, - norm_layer=None, - ) - - def forward(self, x, x_size): - return ( - self.patch_embed( - self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size)) - ) - + x - ) - - def flops(self): - flops = 0 - flops += self.residual_group.flops() - H, W = self.input_resolution - flops += H * W * self.dim * self.dim * 9 - flops += self.patch_embed.flops() - flops += self.patch_unembed.flops() - - return flops - - -class PatchUnEmbed(nn.Module): - r"""Image to Patch Unembedding - - Args: - img_size (int): Image size. Default: 224. - patch_size (int): Patch token size. Default: 4. - in_chans (int): Number of input image channels. Default: 3. - embed_dim (int): Number of linear projection output channels. Default: 96. - norm_layer (nn.Module, optional): Normalization layer. Default: None - """ - - def __init__( - self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None - ): - super().__init__() - img_size = to_2tuple(img_size) - patch_size = to_2tuple(patch_size) - patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] # type: ignore - self.img_size = img_size - self.patch_size = patch_size - self.patches_resolution = patches_resolution - self.num_patches = patches_resolution[0] * patches_resolution[1] - - self.in_chans = in_chans - self.embed_dim = embed_dim - - def forward(self, x, x_size): - B, HW, C = x.shape - x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C - return x - - def flops(self): - flops = 0 - return flops - - -class Upsample(nn.Sequential): - """Upsample module. - - Args: - scale (int): Scale factor. Supported scales: 2^n and 3. - num_feat (int): Channel number of intermediate features. - """ - - def __init__(self, scale, num_feat): - m = [] - if (scale & (scale - 1)) == 0: # scale = 2^n - for _ in range(int(math.log(scale, 2))): - m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(2)) - elif scale == 3: - m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(3)) - else: - raise ValueError( - f"scale {scale} is not supported. " "Supported scales: 2^n and 3." - ) - super(Upsample, self).__init__(*m) - - -class Upsample_hf(nn.Sequential): - """Upsample module. - - Args: - scale (int): Scale factor. Supported scales: 2^n and 3. - num_feat (int): Channel number of intermediate features. - """ - - def __init__(self, scale, num_feat): - m = [] - if (scale & (scale - 1)) == 0: # scale = 2^n - for _ in range(int(math.log(scale, 2))): - m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(2)) - elif scale == 3: - m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(3)) - else: - raise ValueError( - f"scale {scale} is not supported. " "Supported scales: 2^n and 3." - ) - super(Upsample_hf, self).__init__(*m) - - -class UpsampleOneStep(nn.Sequential): - """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle) - Used in lightweight SR to save parameters. - - Args: - scale (int): Scale factor. Supported scales: 2^n and 3. - num_feat (int): Channel number of intermediate features. - - """ - - def __init__(self, scale, num_feat, num_out_ch, input_resolution=None): - self.num_feat = num_feat - self.input_resolution = input_resolution - m = [] - m.append(nn.Conv2d(num_feat, (scale**2) * num_out_ch, 3, 1, 1)) - m.append(nn.PixelShuffle(scale)) - super(UpsampleOneStep, self).__init__(*m) - - def flops(self): - H, W = self.input_resolution # type: ignore - flops = H * W * self.num_feat * 3 * 9 - return flops - - -class Swin2SR(nn.Module): - r"""Swin2SR - A PyTorch impl of : `Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration`. - - Args: - img_size (int | tuple(int)): Input image size. Default 64 - patch_size (int | tuple(int)): Patch size. Default: 1 - in_chans (int): Number of input image channels. Default: 3 - embed_dim (int): Patch embedding dimension. Default: 96 - depths (tuple(int)): Depth of each Swin Transformer layer. - num_heads (tuple(int)): Number of attention heads in different layers. - window_size (int): Window size. Default: 7 - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4 - qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True - drop_rate (float): Dropout rate. Default: 0 - attn_drop_rate (float): Attention dropout rate. Default: 0 - drop_path_rate (float): Stochastic depth rate. Default: 0.1 - norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. - ape (bool): If True, add absolute position embedding to the patch embedding. Default: False - patch_norm (bool): If True, add normalization after patch embedding. Default: True - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False - upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction - img_range: Image range. 1. or 255. - upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None - resi_connection: The convolutional block before residual connection. '1conv'/'3conv' - """ - - def __init__( - self, - state_dict, - **kwargs, - ): - super(Swin2SR, self).__init__() - - # Defaults - img_size = 128 - patch_size = 1 - in_chans = 3 - embed_dim = 96 - depths = [6, 6, 6, 6] - num_heads = [6, 6, 6, 6] - window_size = 7 - mlp_ratio = 4.0 - qkv_bias = True - drop_rate = 0.0 - attn_drop_rate = 0.0 - drop_path_rate = 0.1 - norm_layer = nn.LayerNorm - ape = False - patch_norm = True - use_checkpoint = False - upscale = 2 - img_range = 1.0 - upsampler = "" - resi_connection = "1conv" - num_in_ch = in_chans - num_out_ch = in_chans - num_feat = 64 - - self.model_arch = "Swin2SR" - self.sub_type = "SR" - self.state = state_dict - if "params_ema" in self.state: - self.state = self.state["params_ema"] - elif "params" in self.state: - self.state = self.state["params"] - - state_keys = self.state.keys() - - if "conv_before_upsample.0.weight" in state_keys: - if "conv_aux.weight" in state_keys: - upsampler = "pixelshuffle_aux" - elif "conv_up1.weight" in state_keys: - upsampler = "nearest+conv" - else: - upsampler = "pixelshuffle" - supports_fp16 = False - elif "upsample.0.weight" in state_keys: - upsampler = "pixelshuffledirect" - else: - upsampler = "" - - num_feat = ( - self.state.get("conv_before_upsample.0.weight", None).shape[1] - if self.state.get("conv_before_upsample.weight", None) - else 64 - ) - - num_in_ch = self.state["conv_first.weight"].shape[1] - in_chans = num_in_ch - if "conv_last.weight" in state_keys: - num_out_ch = self.state["conv_last.weight"].shape[0] - else: - num_out_ch = num_in_ch - - upscale = 1 - if upsampler == "nearest+conv": - upsample_keys = [ - x for x in state_keys if "conv_up" in x and "bias" not in x - ] - - for upsample_key in upsample_keys: - upscale *= 2 - elif upsampler == "pixelshuffle" or upsampler == "pixelshuffle_aux": - upsample_keys = [ - x - for x in state_keys - if "upsample" in x and "conv" not in x and "bias" not in x - ] - for upsample_key in upsample_keys: - shape = self.state[upsample_key].shape[0] - upscale *= math.sqrt(shape // num_feat) - upscale = int(upscale) - elif upsampler == "pixelshuffledirect": - upscale = int( - math.sqrt(self.state["upsample.0.bias"].shape[0] // num_out_ch) - ) - - max_layer_num = 0 - max_block_num = 0 - for key in state_keys: - result = re.match( - r"layers.(\d*).residual_group.blocks.(\d*).norm1.weight", key - ) - if result: - layer_num, block_num = result.groups() - max_layer_num = max(max_layer_num, int(layer_num)) - max_block_num = max(max_block_num, int(block_num)) - - depths = [max_block_num + 1 for _ in range(max_layer_num + 1)] - - if ( - "layers.0.residual_group.blocks.0.attn.relative_position_bias_table" - in state_keys - ): - num_heads_num = self.state[ - "layers.0.residual_group.blocks.0.attn.relative_position_bias_table" - ].shape[-1] - num_heads = [num_heads_num for _ in range(max_layer_num + 1)] - else: - num_heads = depths - - embed_dim = self.state["conv_first.weight"].shape[0] - - mlp_ratio = float( - self.state["layers.0.residual_group.blocks.0.mlp.fc1.bias"].shape[0] - / embed_dim - ) - - # TODO: could actually count the layers, but this should do - if "layers.0.conv.4.weight" in state_keys: - resi_connection = "3conv" - else: - resi_connection = "1conv" - - window_size = int( - math.sqrt( - self.state[ - "layers.0.residual_group.blocks.0.attn.relative_position_index" - ].shape[0] - ) - ) - - if "layers.0.residual_group.blocks.1.attn_mask" in state_keys: - img_size = int( - math.sqrt( - self.state["layers.0.residual_group.blocks.1.attn_mask"].shape[0] - ) - * window_size - ) - - # The JPEG models are the only ones with window-size 7, and they also use this range - img_range = 255.0 if window_size == 7 else 1.0 - - self.in_nc = num_in_ch - self.out_nc = num_out_ch - self.num_feat = num_feat - self.embed_dim = embed_dim - self.num_heads = num_heads - self.depths = depths - self.window_size = window_size - self.mlp_ratio = mlp_ratio - self.scale = upscale - self.upsampler = upsampler - self.img_size = img_size - self.img_range = img_range - self.resi_connection = resi_connection - - self.supports_fp16 = False # Too much weirdness to support this at the moment - self.supports_bfp16 = True - self.min_size_restriction = 16 - - ## END AUTO DETECTION - - if in_chans == 3: - rgb_mean = (0.4488, 0.4371, 0.4040) - self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1) - else: - self.mean = torch.zeros(1, 1, 1, 1) - self.upscale = upscale - self.upsampler = upsampler - self.window_size = window_size - - ##################################################################################################### - ################################### 1, shallow feature extraction ################################### - self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1) - - ##################################################################################################### - ################################### 2, deep feature extraction ###################################### - self.num_layers = len(depths) - self.embed_dim = embed_dim - self.ape = ape - self.patch_norm = patch_norm - self.num_features = embed_dim - self.mlp_ratio = mlp_ratio - - # split image into non-overlapping patches - self.patch_embed = PatchEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=embed_dim, - embed_dim=embed_dim, - norm_layer=norm_layer if self.patch_norm else None, - ) - num_patches = self.patch_embed.num_patches - patches_resolution = self.patch_embed.patches_resolution - self.patches_resolution = patches_resolution - - # merge non-overlapping patches into image - self.patch_unembed = PatchUnEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=embed_dim, - embed_dim=embed_dim, - norm_layer=norm_layer if self.patch_norm else None, - ) - - # absolute position embedding - if self.ape: - self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim)) # type: ignore - trunc_normal_(self.absolute_pos_embed, std=0.02) - - self.pos_drop = nn.Dropout(p=drop_rate) - - # stochastic depth - dpr = [ - x.item() for x in torch.linspace(0, drop_path_rate, sum(depths)) - ] # stochastic depth decay rule - - # build Residual Swin Transformer blocks (RSTB) - self.layers = nn.ModuleList() - for i_layer in range(self.num_layers): - layer = RSTB( - dim=embed_dim, - input_resolution=(patches_resolution[0], patches_resolution[1]), - depth=depths[i_layer], - num_heads=num_heads[i_layer], - window_size=window_size, - mlp_ratio=self.mlp_ratio, - qkv_bias=qkv_bias, - drop=drop_rate, - attn_drop=attn_drop_rate, - drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])], # type: ignore # no impact on SR results - norm_layer=norm_layer, - downsample=None, - use_checkpoint=use_checkpoint, - img_size=img_size, - patch_size=patch_size, - resi_connection=resi_connection, - ) - self.layers.append(layer) - - if self.upsampler == "pixelshuffle_hf": - self.layers_hf = nn.ModuleList() - for i_layer in range(self.num_layers): - layer = RSTB( - dim=embed_dim, - input_resolution=(patches_resolution[0], patches_resolution[1]), - depth=depths[i_layer], - num_heads=num_heads[i_layer], - window_size=window_size, - mlp_ratio=self.mlp_ratio, - qkv_bias=qkv_bias, - drop=drop_rate, - attn_drop=attn_drop_rate, - drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])], # type: ignore # no impact on SR results # type: ignore - norm_layer=norm_layer, - downsample=None, - use_checkpoint=use_checkpoint, - img_size=img_size, - patch_size=patch_size, - resi_connection=resi_connection, - ) - self.layers_hf.append(layer) - - self.norm = norm_layer(self.num_features) - - # build the last conv layer in deep feature extraction - if resi_connection == "1conv": - self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1) - elif resi_connection == "3conv": - # to save parameters and memory - self.conv_after_body = nn.Sequential( - nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1), - ) - - ##################################################################################################### - ################################ 3, high quality image reconstruction ################################ - if self.upsampler == "pixelshuffle": - # for classical SR - self.conv_before_upsample = nn.Sequential( - nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.upsample = Upsample(upscale, num_feat) - self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - elif self.upsampler == "pixelshuffle_aux": - self.conv_bicubic = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1) - self.conv_before_upsample = nn.Sequential( - nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.conv_aux = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - self.conv_after_aux = nn.Sequential( - nn.Conv2d(3, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.upsample = Upsample(upscale, num_feat) - self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - - elif self.upsampler == "pixelshuffle_hf": - self.conv_before_upsample = nn.Sequential( - nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.upsample = Upsample(upscale, num_feat) - self.upsample_hf = Upsample_hf(upscale, num_feat) - self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - self.conv_first_hf = nn.Sequential( - nn.Conv2d(num_feat, embed_dim, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.conv_after_body_hf = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1) - self.conv_before_upsample_hf = nn.Sequential( - nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.conv_last_hf = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - - elif self.upsampler == "pixelshuffledirect": - # for lightweight SR (to save parameters) - self.upsample = UpsampleOneStep( - upscale, - embed_dim, - num_out_ch, - (patches_resolution[0], patches_resolution[1]), - ) - elif self.upsampler == "nearest+conv": - # for real-world SR (less artifacts) - assert self.upscale == 4, "only support x4 now." - self.conv_before_upsample = nn.Sequential( - nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) - else: - # for image denoising and JPEG compression artifact reduction - self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1) - - self.apply(self._init_weights) - - self.load_state_dict(state_dict) - - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=0.02) - if isinstance(m, nn.Linear) and m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.LayerNorm): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) - - @torch.jit.ignore # type: ignore - def no_weight_decay(self): - return {"absolute_pos_embed"} - - @torch.jit.ignore # type: ignore - def no_weight_decay_keywords(self): - return {"relative_position_bias_table"} - - def check_image_size(self, x): - _, _, h, w = x.size() - mod_pad_h = (self.window_size - h % self.window_size) % self.window_size - mod_pad_w = (self.window_size - w % self.window_size) % self.window_size - x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect") - return x - - def forward_features(self, x): - x_size = (x.shape[2], x.shape[3]) - x = self.patch_embed(x) - if self.ape: - x = x + self.absolute_pos_embed - x = self.pos_drop(x) - - for layer in self.layers: - x = layer(x, x_size) - - x = self.norm(x) # B L C - x = self.patch_unembed(x, x_size) - - return x - - def forward_features_hf(self, x): - x_size = (x.shape[2], x.shape[3]) - x = self.patch_embed(x) - if self.ape: - x = x + self.absolute_pos_embed - x = self.pos_drop(x) - - for layer in self.layers_hf: - x = layer(x, x_size) - - x = self.norm(x) # B L C - x = self.patch_unembed(x, x_size) - - return x - - def forward(self, x): - H, W = x.shape[2:] - x = self.check_image_size(x) - - self.mean = self.mean.type_as(x) - x = (x - self.mean) * self.img_range - - if self.upsampler == "pixelshuffle": - # for classical SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.conv_before_upsample(x) - x = self.conv_last(self.upsample(x)) - elif self.upsampler == "pixelshuffle_aux": - bicubic = F.interpolate( - x, - size=(H * self.upscale, W * self.upscale), - mode="bicubic", - align_corners=False, - ) - bicubic = self.conv_bicubic(bicubic) - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.conv_before_upsample(x) - aux = self.conv_aux(x) # b, 3, LR_H, LR_W - x = self.conv_after_aux(aux) - x = ( - self.upsample(x)[:, :, : H * self.upscale, : W * self.upscale] - + bicubic[:, :, : H * self.upscale, : W * self.upscale] - ) - x = self.conv_last(x) - aux = aux / self.img_range + self.mean - elif self.upsampler == "pixelshuffle_hf": - # for classical SR with HF - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x_before = self.conv_before_upsample(x) - x_out = self.conv_last(self.upsample(x_before)) - - x_hf = self.conv_first_hf(x_before) - x_hf = self.conv_after_body_hf(self.forward_features_hf(x_hf)) + x_hf - x_hf = self.conv_before_upsample_hf(x_hf) - x_hf = self.conv_last_hf(self.upsample_hf(x_hf)) - x = x_out + x_hf - x_hf = x_hf / self.img_range + self.mean - - elif self.upsampler == "pixelshuffledirect": - # for lightweight SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.upsample(x) - elif self.upsampler == "nearest+conv": - # for real-world SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.conv_before_upsample(x) - x = self.lrelu( - self.conv_up1( - torch.nn.functional.interpolate(x, scale_factor=2, mode="nearest") - ) - ) - x = self.lrelu( - self.conv_up2( - torch.nn.functional.interpolate(x, scale_factor=2, mode="nearest") - ) - ) - x = self.conv_last(self.lrelu(self.conv_hr(x))) - else: - # for image denoising and JPEG compression artifact reduction - x_first = self.conv_first(x) - res = self.conv_after_body(self.forward_features(x_first)) + x_first - x = x + self.conv_last(res) - - x = x / self.img_range + self.mean - if self.upsampler == "pixelshuffle_aux": - # NOTE: I removed an "aux" output here. not sure what that was for - return x[:, :, : H * self.upscale, : W * self.upscale] # type: ignore - - elif self.upsampler == "pixelshuffle_hf": - x_out = x_out / self.img_range + self.mean # type: ignore - return x_out[:, :, : H * self.upscale, : W * self.upscale], x[:, :, : H * self.upscale, : W * self.upscale], x_hf[:, :, : H * self.upscale, : W * self.upscale] # type: ignore - - else: - return x[:, :, : H * self.upscale, : W * self.upscale] - - def flops(self): - flops = 0 - H, W = self.patches_resolution - flops += H * W * 3 * self.embed_dim * 9 - flops += self.patch_embed.flops() - for i, layer in enumerate(self.layers): - flops += layer.flops() # type: ignore - flops += H * W * 3 * self.embed_dim * self.embed_dim - flops += self.upsample.flops() # type: ignore - return flops diff --git a/comfy_extras/chainner_models/architecture/SwinIR.py b/comfy_extras/chainner_models/architecture/SwinIR.py deleted file mode 100644 index 439dcbcb2..000000000 --- a/comfy_extras/chainner_models/architecture/SwinIR.py +++ /dev/null @@ -1,1224 +0,0 @@ -# pylint: skip-file -# ----------------------------------------------------------------------------------- -# SwinIR: Image Restoration Using Swin Transformer, https://arxiv.org/abs/2108.10257 -# Originally Written by Ze Liu, Modified by Jingyun Liang. -# ----------------------------------------------------------------------------------- - -import math -import re - -import torch -import torch.nn as nn -import torch.nn.functional as F -import torch.utils.checkpoint as checkpoint - -# Originally from the timm package -from .timm.drop import DropPath -from .timm.helpers import to_2tuple -from .timm.weight_init import trunc_normal_ - - -class Mlp(nn.Module): - def __init__( - self, - in_features, - hidden_features=None, - out_features=None, - act_layer=nn.GELU, - drop=0.0, - ): - super().__init__() - out_features = out_features or in_features - hidden_features = hidden_features or in_features - self.fc1 = nn.Linear(in_features, hidden_features) - self.act = act_layer() - self.fc2 = nn.Linear(hidden_features, out_features) - self.drop = nn.Dropout(drop) - - def forward(self, x): - x = self.fc1(x) - x = self.act(x) - x = self.drop(x) - x = self.fc2(x) - x = self.drop(x) - return x - - -def window_partition(x, window_size): - """ - Args: - x: (B, H, W, C) - window_size (int): window size - - Returns: - windows: (num_windows*B, window_size, window_size, C) - """ - B, H, W, C = x.shape - x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) - windows = ( - x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) - ) - return windows - - -def window_reverse(windows, window_size, H, W): - """ - Args: - windows: (num_windows*B, window_size, window_size, C) - window_size (int): Window size - H (int): Height of image - W (int): Width of image - - Returns: - x: (B, H, W, C) - """ - B = int(windows.shape[0] / (H * W / window_size / window_size)) - x = windows.view( - B, H // window_size, W // window_size, window_size, window_size, -1 - ) - x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) - return x - - -class WindowAttention(nn.Module): - r"""Window based multi-head self attention (W-MSA) module with relative position bias. - It supports both of shifted and non-shifted window. - - Args: - dim (int): Number of input channels. - window_size (tuple[int]): The height and width of the window. - num_heads (int): Number of attention heads. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set - attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 - proj_drop (float, optional): Dropout ratio of output. Default: 0.0 - """ - - def __init__( - self, - dim, - window_size, - num_heads, - qkv_bias=True, - qk_scale=None, - attn_drop=0.0, - proj_drop=0.0, - ): - super().__init__() - self.dim = dim - self.window_size = window_size # Wh, Ww - self.num_heads = num_heads - head_dim = dim // num_heads - self.scale = qk_scale or head_dim**-0.5 - - # define a parameter table of relative position bias - self.relative_position_bias_table = nn.Parameter( # type: ignore - torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads) - ) # 2*Wh-1 * 2*Ww-1, nH - - # get pair-wise relative position index for each token inside the window - coords_h = torch.arange(self.window_size[0]) - coords_w = torch.arange(self.window_size[1]) - coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww - coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww - relative_coords = ( - coords_flatten[:, :, None] - coords_flatten[:, None, :] - ) # 2, Wh*Ww, Wh*Ww - relative_coords = relative_coords.permute( - 1, 2, 0 - ).contiguous() # Wh*Ww, Wh*Ww, 2 - relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0 - relative_coords[:, :, 1] += self.window_size[1] - 1 - relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 - relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww - self.register_buffer("relative_position_index", relative_position_index) - - self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) - self.attn_drop = nn.Dropout(attn_drop) - self.proj = nn.Linear(dim, dim) - - self.proj_drop = nn.Dropout(proj_drop) - - trunc_normal_(self.relative_position_bias_table, std=0.02) - self.softmax = nn.Softmax(dim=-1) - - def forward(self, x, mask=None): - """ - Args: - x: input features with shape of (num_windows*B, N, C) - mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None - """ - B_, N, C = x.shape - qkv = ( - self.qkv(x) - .reshape(B_, N, 3, self.num_heads, C // self.num_heads) - .permute(2, 0, 3, 1, 4) - ) - q, k, v = ( - qkv[0], - qkv[1], - qkv[2], - ) # make torchscript happy (cannot use tensor as tuple) - - q = q * self.scale - attn = q @ k.transpose(-2, -1) - - relative_position_bias = self.relative_position_bias_table[ - self.relative_position_index.view(-1) # type: ignore - ].view( - self.window_size[0] * self.window_size[1], - self.window_size[0] * self.window_size[1], - -1, - ) # Wh*Ww,Wh*Ww,nH - relative_position_bias = relative_position_bias.permute( - 2, 0, 1 - ).contiguous() # nH, Wh*Ww, Wh*Ww - attn = attn + relative_position_bias.unsqueeze(0) - - if mask is not None: - nW = mask.shape[0] - attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze( - 1 - ).unsqueeze(0) - attn = attn.view(-1, self.num_heads, N, N) - attn = self.softmax(attn) - else: - attn = self.softmax(attn) - - attn = self.attn_drop(attn) - - x = (attn @ v).transpose(1, 2).reshape(B_, N, C) - x = self.proj(x) - x = self.proj_drop(x) - return x - - def extra_repr(self) -> str: - return f"dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}" - - def flops(self, N): - # calculate flops for 1 window with token length of N - flops = 0 - # qkv = self.qkv(x) - flops += N * self.dim * 3 * self.dim - # attn = (q @ k.transpose(-2, -1)) - flops += self.num_heads * N * (self.dim // self.num_heads) * N - # x = (attn @ v) - flops += self.num_heads * N * N * (self.dim // self.num_heads) - # x = self.proj(x) - flops += N * self.dim * self.dim - return flops - - -class SwinTransformerBlock(nn.Module): - r"""Swin Transformer Block. - - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resulotion. - num_heads (int): Number of attention heads. - window_size (int): Window size. - shift_size (int): Shift size for SW-MSA. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float, optional): Stochastic depth rate. Default: 0.0 - act_layer (nn.Module, optional): Activation layer. Default: nn.GELU - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - """ - - def __init__( - self, - dim, - input_resolution, - num_heads, - window_size=7, - shift_size=0, - mlp_ratio=4.0, - qkv_bias=True, - qk_scale=None, - drop=0.0, - attn_drop=0.0, - drop_path=0.0, - act_layer=nn.GELU, - norm_layer=nn.LayerNorm, - ): - super().__init__() - self.dim = dim - self.input_resolution = input_resolution - self.num_heads = num_heads - self.window_size = window_size - self.shift_size = shift_size - self.mlp_ratio = mlp_ratio - if min(self.input_resolution) <= self.window_size: - # if window size is larger than input resolution, we don't partition windows - self.shift_size = 0 - self.window_size = min(self.input_resolution) - assert ( - 0 <= self.shift_size < self.window_size - ), "shift_size must in 0-window_size" - - self.norm1 = norm_layer(dim) - self.attn = WindowAttention( - dim, - window_size=to_2tuple(self.window_size), - num_heads=num_heads, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - attn_drop=attn_drop, - proj_drop=drop, - ) - - self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() - self.norm2 = norm_layer(dim) - mlp_hidden_dim = int(dim * mlp_ratio) - self.mlp = Mlp( - in_features=dim, - hidden_features=mlp_hidden_dim, - act_layer=act_layer, - drop=drop, - ) - - if self.shift_size > 0: - attn_mask = self.calculate_mask(self.input_resolution) - else: - attn_mask = None - - self.register_buffer("attn_mask", attn_mask) - - def calculate_mask(self, x_size): - # calculate attention mask for SW-MSA - H, W = x_size - img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1 - h_slices = ( - slice(0, -self.window_size), - slice(-self.window_size, -self.shift_size), - slice(-self.shift_size, None), - ) - w_slices = ( - slice(0, -self.window_size), - slice(-self.window_size, -self.shift_size), - slice(-self.shift_size, None), - ) - cnt = 0 - for h in h_slices: - for w in w_slices: - img_mask[:, h, w, :] = cnt - cnt += 1 - - mask_windows = window_partition( - img_mask, self.window_size - ) # nW, window_size, window_size, 1 - mask_windows = mask_windows.view(-1, self.window_size * self.window_size) - attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) - attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill( - attn_mask == 0, float(0.0) - ) - - return attn_mask - - def forward(self, x, x_size): - H, W = x_size - B, L, C = x.shape - # assert L == H * W, "input feature has wrong size" - - shortcut = x - x = self.norm1(x) - x = x.view(B, H, W, C) - - # cyclic shift - if self.shift_size > 0: - shifted_x = torch.roll( - x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2) - ) - else: - shifted_x = x - - # partition windows - x_windows = window_partition( - shifted_x, self.window_size - ) # nW*B, window_size, window_size, C - x_windows = x_windows.view( - -1, self.window_size * self.window_size, C - ) # nW*B, window_size*window_size, C - - # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size - if self.input_resolution == x_size: - attn_windows = self.attn( - x_windows, mask=self.attn_mask - ) # nW*B, window_size*window_size, C - else: - attn_windows = self.attn( - x_windows, mask=self.calculate_mask(x_size).to(x.device) - ) - - # merge windows - attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) - shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C - - # reverse cyclic shift - if self.shift_size > 0: - x = torch.roll( - shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2) - ) - else: - x = shifted_x - x = x.view(B, H * W, C) - - # FFN - x = shortcut + self.drop_path(x) - x = x + self.drop_path(self.mlp(self.norm2(x))) - - return x - - def extra_repr(self) -> str: - return ( - f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " - f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}" - ) - - def flops(self): - flops = 0 - H, W = self.input_resolution - # norm1 - flops += self.dim * H * W - # W-MSA/SW-MSA - nW = H * W / self.window_size / self.window_size - flops += nW * self.attn.flops(self.window_size * self.window_size) - # mlp - flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio - # norm2 - flops += self.dim * H * W - return flops - - -class PatchMerging(nn.Module): - r"""Patch Merging Layer. - - Args: - input_resolution (tuple[int]): Resolution of input feature. - dim (int): Number of input channels. - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - """ - - def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm): - super().__init__() - self.input_resolution = input_resolution - self.dim = dim - self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) - self.norm = norm_layer(4 * dim) - - def forward(self, x): - """ - x: B, H*W, C - """ - H, W = self.input_resolution - B, L, C = x.shape - assert L == H * W, "input feature has wrong size" - assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even." - - x = x.view(B, H, W, C) - - x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C - x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C - x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C - x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C - x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C - x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C - - x = self.norm(x) - x = self.reduction(x) - - return x - - def extra_repr(self) -> str: - return f"input_resolution={self.input_resolution}, dim={self.dim}" - - def flops(self): - H, W = self.input_resolution - flops = H * W * self.dim - flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim - return flops - - -class BasicLayer(nn.Module): - """A basic Swin Transformer layer for one stage. - - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resolution. - depth (int): Number of blocks. - num_heads (int): Number of attention heads. - window_size (int): Local window size. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. - """ - - def __init__( - self, - dim, - input_resolution, - depth, - num_heads, - window_size, - mlp_ratio=4.0, - qkv_bias=True, - qk_scale=None, - drop=0.0, - attn_drop=0.0, - drop_path=0.0, - norm_layer=nn.LayerNorm, - downsample=None, - use_checkpoint=False, - ): - super().__init__() - self.dim = dim - self.input_resolution = input_resolution - self.depth = depth - self.use_checkpoint = use_checkpoint - - # build blocks - self.blocks = nn.ModuleList( - [ - SwinTransformerBlock( - dim=dim, - input_resolution=input_resolution, - num_heads=num_heads, - window_size=window_size, - shift_size=0 if (i % 2 == 0) else window_size // 2, - mlp_ratio=mlp_ratio, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop, - attn_drop=attn_drop, - drop_path=drop_path[i] - if isinstance(drop_path, list) - else drop_path, - norm_layer=norm_layer, - ) - for i in range(depth) - ] - ) - - # patch merging layer - if downsample is not None: - self.downsample = downsample( - input_resolution, dim=dim, norm_layer=norm_layer - ) - else: - self.downsample = None - - def forward(self, x, x_size): - for blk in self.blocks: - if self.use_checkpoint: - x = checkpoint.checkpoint(blk, x, x_size) - else: - x = blk(x, x_size) - if self.downsample is not None: - x = self.downsample(x) - return x - - def extra_repr(self) -> str: - return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}" - - def flops(self): - flops = 0 - for blk in self.blocks: - flops += blk.flops() # type: ignore - if self.downsample is not None: - flops += self.downsample.flops() - return flops - - -class RSTB(nn.Module): - """Residual Swin Transformer Block (RSTB). - - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resolution. - depth (int): Number of blocks. - num_heads (int): Number of attention heads. - window_size (int): Local window size. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. - img_size: Input image size. - patch_size: Patch size. - resi_connection: The convolutional block before residual connection. - """ - - def __init__( - self, - dim, - input_resolution, - depth, - num_heads, - window_size, - mlp_ratio=4.0, - qkv_bias=True, - qk_scale=None, - drop=0.0, - attn_drop=0.0, - drop_path=0.0, - norm_layer=nn.LayerNorm, - downsample=None, - use_checkpoint=False, - img_size=224, - patch_size=4, - resi_connection="1conv", - ): - super(RSTB, self).__init__() - - self.dim = dim - self.input_resolution = input_resolution - - self.residual_group = BasicLayer( - dim=dim, - input_resolution=input_resolution, - depth=depth, - num_heads=num_heads, - window_size=window_size, - mlp_ratio=mlp_ratio, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop, - attn_drop=attn_drop, - drop_path=drop_path, - norm_layer=norm_layer, - downsample=downsample, - use_checkpoint=use_checkpoint, - ) - - if resi_connection == "1conv": - self.conv = nn.Conv2d(dim, dim, 3, 1, 1) - elif resi_connection == "3conv": - # to save parameters and memory - self.conv = nn.Sequential( - nn.Conv2d(dim, dim // 4, 3, 1, 1), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(dim // 4, dim // 4, 1, 1, 0), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(dim // 4, dim, 3, 1, 1), - ) - - self.patch_embed = PatchEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=0, - embed_dim=dim, - norm_layer=None, - ) - - self.patch_unembed = PatchUnEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=0, - embed_dim=dim, - norm_layer=None, - ) - - def forward(self, x, x_size): - return ( - self.patch_embed( - self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size)) - ) - + x - ) - - def flops(self): - flops = 0 - flops += self.residual_group.flops() - H, W = self.input_resolution - flops += H * W * self.dim * self.dim * 9 - flops += self.patch_embed.flops() - flops += self.patch_unembed.flops() - - return flops - - -class PatchEmbed(nn.Module): - r"""Image to Patch Embedding - - Args: - img_size (int): Image size. Default: 224. - patch_size (int): Patch token size. Default: 4. - in_chans (int): Number of input image channels. Default: 3. - embed_dim (int): Number of linear projection output channels. Default: 96. - norm_layer (nn.Module, optional): Normalization layer. Default: None - """ - - def __init__( - self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None - ): - super().__init__() - img_size = to_2tuple(img_size) - patch_size = to_2tuple(patch_size) - patches_resolution = [ - img_size[0] // patch_size[0], # type: ignore - img_size[1] // patch_size[1], # type: ignore - ] - self.img_size = img_size - self.patch_size = patch_size - self.patches_resolution = patches_resolution - self.num_patches = patches_resolution[0] * patches_resolution[1] - - self.in_chans = in_chans - self.embed_dim = embed_dim - - if norm_layer is not None: - self.norm = norm_layer(embed_dim) - else: - self.norm = None - - def forward(self, x): - x = x.flatten(2).transpose(1, 2) # B Ph*Pw C - if self.norm is not None: - x = self.norm(x) - return x - - def flops(self): - flops = 0 - H, W = self.img_size - if self.norm is not None: - flops += H * W * self.embed_dim # type: ignore - return flops - - -class PatchUnEmbed(nn.Module): - r"""Image to Patch Unembedding - - Args: - img_size (int): Image size. Default: 224. - patch_size (int): Patch token size. Default: 4. - in_chans (int): Number of input image channels. Default: 3. - embed_dim (int): Number of linear projection output channels. Default: 96. - norm_layer (nn.Module, optional): Normalization layer. Default: None - """ - - def __init__( - self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None - ): - super().__init__() - img_size = to_2tuple(img_size) - patch_size = to_2tuple(patch_size) - patches_resolution = [ - img_size[0] // patch_size[0], # type: ignore - img_size[1] // patch_size[1], # type: ignore - ] - self.img_size = img_size - self.patch_size = patch_size - self.patches_resolution = patches_resolution - self.num_patches = patches_resolution[0] * patches_resolution[1] - - self.in_chans = in_chans - self.embed_dim = embed_dim - - def forward(self, x, x_size): - B, HW, C = x.shape - x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C - return x - - def flops(self): - flops = 0 - return flops - - -class Upsample(nn.Sequential): - """Upsample module. - - Args: - scale (int): Scale factor. Supported scales: 2^n and 3. - num_feat (int): Channel number of intermediate features. - """ - - def __init__(self, scale, num_feat): - m = [] - if (scale & (scale - 1)) == 0: # scale = 2^n - for _ in range(int(math.log(scale, 2))): - m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(2)) - elif scale == 3: - m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(3)) - else: - raise ValueError( - f"scale {scale} is not supported. " "Supported scales: 2^n and 3." - ) - super(Upsample, self).__init__(*m) - - -class UpsampleOneStep(nn.Sequential): - """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle) - Used in lightweight SR to save parameters. - - Args: - scale (int): Scale factor. Supported scales: 2^n and 3. - num_feat (int): Channel number of intermediate features. - - """ - - def __init__(self, scale, num_feat, num_out_ch, input_resolution=None): - self.num_feat = num_feat - self.input_resolution = input_resolution - m = [] - m.append(nn.Conv2d(num_feat, (scale**2) * num_out_ch, 3, 1, 1)) - m.append(nn.PixelShuffle(scale)) - super(UpsampleOneStep, self).__init__(*m) - - def flops(self): - H, W = self.input_resolution # type: ignore - flops = H * W * self.num_feat * 3 * 9 - return flops - - -class SwinIR(nn.Module): - r"""SwinIR - A PyTorch impl of : `SwinIR: Image Restoration Using Swin Transformer`, based on Swin Transformer. - - Args: - img_size (int | tuple(int)): Input image size. Default 64 - patch_size (int | tuple(int)): Patch size. Default: 1 - in_chans (int): Number of input image channels. Default: 3 - embed_dim (int): Patch embedding dimension. Default: 96 - depths (tuple(int)): Depth of each Swin Transformer layer. - num_heads (tuple(int)): Number of attention heads in different layers. - window_size (int): Window size. Default: 7 - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4 - qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None - drop_rate (float): Dropout rate. Default: 0 - attn_drop_rate (float): Attention dropout rate. Default: 0 - drop_path_rate (float): Stochastic depth rate. Default: 0.1 - norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. - ape (bool): If True, add absolute position embedding to the patch embedding. Default: False - patch_norm (bool): If True, add normalization after patch embedding. Default: True - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False - upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction - img_range: Image range. 1. or 255. - upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None - resi_connection: The convolutional block before residual connection. '1conv'/'3conv' - """ - - def __init__( - self, - state_dict, - **kwargs, - ): - super(SwinIR, self).__init__() - - # Defaults - img_size = 64 - patch_size = 1 - in_chans = 3 - embed_dim = 96 - depths = [6, 6, 6, 6] - num_heads = [6, 6, 6, 6] - window_size = 7 - mlp_ratio = 4.0 - qkv_bias = True - qk_scale = None - drop_rate = 0.0 - attn_drop_rate = 0.0 - drop_path_rate = 0.1 - norm_layer = nn.LayerNorm - ape = False - patch_norm = True - use_checkpoint = False - upscale = 2 - img_range = 1.0 - upsampler = "" - resi_connection = "1conv" - num_feat = 64 - num_in_ch = in_chans - num_out_ch = in_chans - supports_fp16 = True - self.start_unshuffle = 1 - - self.model_arch = "SwinIR" - self.sub_type = "SR" - self.state = state_dict - if "params_ema" in self.state: - self.state = self.state["params_ema"] - elif "params" in self.state: - self.state = self.state["params"] - - state_keys = self.state.keys() - - if "conv_before_upsample.0.weight" in state_keys: - if "conv_up1.weight" in state_keys: - upsampler = "nearest+conv" - else: - upsampler = "pixelshuffle" - supports_fp16 = False - elif "upsample.0.weight" in state_keys: - upsampler = "pixelshuffledirect" - else: - upsampler = "" - - num_feat = ( - self.state.get("conv_before_upsample.0.weight", None).shape[1] - if self.state.get("conv_before_upsample.weight", None) - else 64 - ) - - if "conv_first.1.weight" in self.state: - self.state["conv_first.weight"] = self.state.pop("conv_first.1.weight") - self.state["conv_first.bias"] = self.state.pop("conv_first.1.bias") - self.start_unshuffle = round(math.sqrt(self.state["conv_first.weight"].shape[1] // 3)) - - num_in_ch = self.state["conv_first.weight"].shape[1] - in_chans = num_in_ch - if "conv_last.weight" in state_keys: - num_out_ch = self.state["conv_last.weight"].shape[0] - else: - num_out_ch = num_in_ch - - upscale = 1 - if upsampler == "nearest+conv": - upsample_keys = [ - x for x in state_keys if "conv_up" in x and "bias" not in x - ] - - for upsample_key in upsample_keys: - upscale *= 2 - elif upsampler == "pixelshuffle": - upsample_keys = [ - x - for x in state_keys - if "upsample" in x and "conv" not in x and "bias" not in x - ] - for upsample_key in upsample_keys: - shape = self.state[upsample_key].shape[0] - upscale *= math.sqrt(shape // num_feat) - upscale = int(upscale) - elif upsampler == "pixelshuffledirect": - upscale = int( - math.sqrt(self.state["upsample.0.bias"].shape[0] // num_out_ch) - ) - - max_layer_num = 0 - max_block_num = 0 - for key in state_keys: - result = re.match( - r"layers.(\d*).residual_group.blocks.(\d*).norm1.weight", key - ) - if result: - layer_num, block_num = result.groups() - max_layer_num = max(max_layer_num, int(layer_num)) - max_block_num = max(max_block_num, int(block_num)) - - depths = [max_block_num + 1 for _ in range(max_layer_num + 1)] - - if ( - "layers.0.residual_group.blocks.0.attn.relative_position_bias_table" - in state_keys - ): - num_heads_num = self.state[ - "layers.0.residual_group.blocks.0.attn.relative_position_bias_table" - ].shape[-1] - num_heads = [num_heads_num for _ in range(max_layer_num + 1)] - else: - num_heads = depths - - embed_dim = self.state["conv_first.weight"].shape[0] - - mlp_ratio = float( - self.state["layers.0.residual_group.blocks.0.mlp.fc1.bias"].shape[0] - / embed_dim - ) - - # TODO: could actually count the layers, but this should do - if "layers.0.conv.4.weight" in state_keys: - resi_connection = "3conv" - else: - resi_connection = "1conv" - - window_size = int( - math.sqrt( - self.state[ - "layers.0.residual_group.blocks.0.attn.relative_position_index" - ].shape[0] - ) - ) - - if "layers.0.residual_group.blocks.1.attn_mask" in state_keys: - img_size = int( - math.sqrt( - self.state["layers.0.residual_group.blocks.1.attn_mask"].shape[0] - ) - * window_size - ) - - # The JPEG models are the only ones with window-size 7, and they also use this range - img_range = 255.0 if window_size == 7 else 1.0 - - self.in_nc = num_in_ch - self.out_nc = num_out_ch - self.num_feat = num_feat - self.embed_dim = embed_dim - self.num_heads = num_heads - self.depths = depths - self.window_size = window_size - self.mlp_ratio = mlp_ratio - self.scale = upscale / self.start_unshuffle - self.upsampler = upsampler - self.img_size = img_size - self.img_range = img_range - self.resi_connection = resi_connection - - self.supports_fp16 = False # Too much weirdness to support this at the moment - self.supports_bfp16 = True - self.min_size_restriction = 16 - - self.img_range = img_range - if in_chans == 3: - rgb_mean = (0.4488, 0.4371, 0.4040) - self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1) - else: - self.mean = torch.zeros(1, 1, 1, 1) - self.upscale = upscale - self.upsampler = upsampler - self.window_size = window_size - - ##################################################################################################### - ################################### 1, shallow feature extraction ################################### - self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1) - - ##################################################################################################### - ################################### 2, deep feature extraction ###################################### - self.num_layers = len(depths) - self.embed_dim = embed_dim - self.ape = ape - self.patch_norm = patch_norm - self.num_features = embed_dim - self.mlp_ratio = mlp_ratio - - # split image into non-overlapping patches - self.patch_embed = PatchEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=embed_dim, - embed_dim=embed_dim, - norm_layer=norm_layer if self.patch_norm else None, - ) - num_patches = self.patch_embed.num_patches - patches_resolution = self.patch_embed.patches_resolution - self.patches_resolution = patches_resolution - - # merge non-overlapping patches into image - self.patch_unembed = PatchUnEmbed( - img_size=img_size, - patch_size=patch_size, - in_chans=embed_dim, - embed_dim=embed_dim, - norm_layer=norm_layer if self.patch_norm else None, - ) - - # absolute position embedding - if self.ape: - self.absolute_pos_embed = nn.Parameter( # type: ignore - torch.zeros(1, num_patches, embed_dim) - ) - trunc_normal_(self.absolute_pos_embed, std=0.02) - - self.pos_drop = nn.Dropout(p=drop_rate) - - # stochastic depth - dpr = [ - x.item() for x in torch.linspace(0, drop_path_rate, sum(depths)) - ] # stochastic depth decay rule - - # build Residual Swin Transformer blocks (RSTB) - self.layers = nn.ModuleList() - for i_layer in range(self.num_layers): - layer = RSTB( - dim=embed_dim, - input_resolution=(patches_resolution[0], patches_resolution[1]), - depth=depths[i_layer], - num_heads=num_heads[i_layer], - window_size=window_size, - mlp_ratio=self.mlp_ratio, - qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop_rate, - attn_drop=attn_drop_rate, - drop_path=dpr[ - sum(depths[:i_layer]) : sum(depths[: i_layer + 1]) # type: ignore - ], # no impact on SR results - norm_layer=norm_layer, - downsample=None, - use_checkpoint=use_checkpoint, - img_size=img_size, - patch_size=patch_size, - resi_connection=resi_connection, - ) - self.layers.append(layer) - self.norm = norm_layer(self.num_features) - - # build the last conv layer in deep feature extraction - if resi_connection == "1conv": - self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1) - elif resi_connection == "3conv": - # to save parameters and memory - self.conv_after_body = nn.Sequential( - nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1), - ) - - ##################################################################################################### - ################################ 3, high quality image reconstruction ################################ - if self.upsampler == "pixelshuffle": - # for classical SR - self.conv_before_upsample = nn.Sequential( - nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.upsample = Upsample(upscale, num_feat) - self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - elif self.upsampler == "pixelshuffledirect": - # for lightweight SR (to save parameters) - self.upsample = UpsampleOneStep( - upscale, - embed_dim, - num_out_ch, - (patches_resolution[0], patches_resolution[1]), - ) - elif self.upsampler == "nearest+conv": - # for real-world SR (less artifacts) - self.conv_before_upsample = nn.Sequential( - nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True) - ) - self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - if self.upscale == 4: - self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - elif self.upscale == 8: - self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - self.conv_up3 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) - else: - # for image denoising and JPEG compression artifact reduction - self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1) - - self.apply(self._init_weights) - self.load_state_dict(self.state, strict=False) - - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=0.02) - if isinstance(m, nn.Linear) and m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.LayerNorm): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) - - @torch.jit.ignore # type: ignore - def no_weight_decay(self): - return {"absolute_pos_embed"} - - @torch.jit.ignore # type: ignore - def no_weight_decay_keywords(self): - return {"relative_position_bias_table"} - - def check_image_size(self, x): - _, _, h, w = x.size() - mod_pad_h = (self.window_size - h % self.window_size) % self.window_size - mod_pad_w = (self.window_size - w % self.window_size) % self.window_size - x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect") - return x - - def forward_features(self, x): - x_size = (x.shape[2], x.shape[3]) - x = self.patch_embed(x) - if self.ape: - x = x + self.absolute_pos_embed - x = self.pos_drop(x) - - for layer in self.layers: - x = layer(x, x_size) - - x = self.norm(x) # B L C - x = self.patch_unembed(x, x_size) - - return x - - def forward(self, x): - H, W = x.shape[2:] - x = self.check_image_size(x) - - self.mean = self.mean.type_as(x) - x = (x - self.mean) * self.img_range - - if self.start_unshuffle > 1: - x = torch.nn.functional.pixel_unshuffle(x, self.start_unshuffle) - - if self.upsampler == "pixelshuffle": - # for classical SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.conv_before_upsample(x) - x = self.conv_last(self.upsample(x)) - elif self.upsampler == "pixelshuffledirect": - # for lightweight SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.upsample(x) - elif self.upsampler == "nearest+conv": - # for real-world SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.conv_before_upsample(x) - x = self.lrelu( - self.conv_up1( - torch.nn.functional.interpolate(x, scale_factor=2, mode="nearest") # type: ignore - ) - ) - if self.upscale == 4: - x = self.lrelu( - self.conv_up2( - torch.nn.functional.interpolate( # type: ignore - x, scale_factor=2, mode="nearest" - ) - ) - ) - elif self.upscale == 8: - x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest'))) - x = self.lrelu(self.conv_up3(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest'))) - x = self.conv_last(self.lrelu(self.conv_hr(x))) - else: - # for image denoising and JPEG compression artifact reduction - x_first = self.conv_first(x) - res = self.conv_after_body(self.forward_features(x_first)) + x_first - x = x + self.conv_last(res) - - x = x / self.img_range + self.mean - - return x[:, :, : H * self.upscale, : W * self.upscale] - - def flops(self): - flops = 0 - H, W = self.patches_resolution - flops += H * W * 3 * self.embed_dim * 9 - flops += self.patch_embed.flops() - for i, layer in enumerate(self.layers): - flops += layer.flops() # type: ignore - flops += H * W * 3 * self.embed_dim * self.embed_dim - flops += self.upsample.flops() # type: ignore - return flops diff --git a/comfy_extras/chainner_models/architecture/__init__.py b/comfy_extras/chainner_models/architecture/__init__.py deleted file mode 100644 index e69de29bb..000000000 diff --git a/comfy_extras/chainner_models/architecture/block.py b/comfy_extras/chainner_models/architecture/block.py deleted file mode 100644 index d7bc5d227..000000000 --- a/comfy_extras/chainner_models/architecture/block.py +++ /dev/null @@ -1,546 +0,0 @@ -#!/usr/bin/env python3 -# -*- coding: utf-8 -*- - -from __future__ import annotations - -from collections import OrderedDict -try: - from typing import Literal -except ImportError: - from typing_extensions import Literal - -import torch -import torch.nn as nn - -#################### -# Basic blocks -#################### - - -def act(act_type: str, inplace=True, neg_slope=0.2, n_prelu=1): - # helper selecting activation - # neg_slope: for leakyrelu and init of prelu - # n_prelu: for p_relu num_parameters - act_type = act_type.lower() - if act_type == "relu": - layer = nn.ReLU(inplace) - elif act_type == "leakyrelu": - layer = nn.LeakyReLU(neg_slope, inplace) - elif act_type == "prelu": - layer = nn.PReLU(num_parameters=n_prelu, init=neg_slope) - else: - raise NotImplementedError( - "activation layer [{:s}] is not found".format(act_type) - ) - return layer - - -def norm(norm_type: str, nc: int): - # helper selecting normalization layer - norm_type = norm_type.lower() - if norm_type == "batch": - layer = nn.BatchNorm2d(nc, affine=True) - elif norm_type == "instance": - layer = nn.InstanceNorm2d(nc, affine=False) - else: - raise NotImplementedError( - "normalization layer [{:s}] is not found".format(norm_type) - ) - return layer - - -def pad(pad_type: str, padding): - # helper selecting padding layer - # if padding is 'zero', do by conv layers - pad_type = pad_type.lower() - if padding == 0: - return None - if pad_type == "reflect": - layer = nn.ReflectionPad2d(padding) - elif pad_type == "replicate": - layer = nn.ReplicationPad2d(padding) - else: - raise NotImplementedError( - "padding layer [{:s}] is not implemented".format(pad_type) - ) - return layer - - -def get_valid_padding(kernel_size, dilation): - kernel_size = kernel_size + (kernel_size - 1) * (dilation - 1) - padding = (kernel_size - 1) // 2 - return padding - - -class ConcatBlock(nn.Module): - # Concat the output of a submodule to its input - def __init__(self, submodule): - super(ConcatBlock, self).__init__() - self.sub = submodule - - def forward(self, x): - output = torch.cat((x, self.sub(x)), dim=1) - return output - - def __repr__(self): - tmpstr = "Identity .. \n|" - modstr = self.sub.__repr__().replace("\n", "\n|") - tmpstr = tmpstr + modstr - return tmpstr - - -class ShortcutBlock(nn.Module): - # Elementwise sum the output of a submodule to its input - def __init__(self, submodule): - super(ShortcutBlock, self).__init__() - self.sub = submodule - - def forward(self, x): - output = x + self.sub(x) - return output - - def __repr__(self): - tmpstr = "Identity + \n|" - modstr = self.sub.__repr__().replace("\n", "\n|") - tmpstr = tmpstr + modstr - return tmpstr - - -class ShortcutBlockSPSR(nn.Module): - # Elementwise sum the output of a submodule to its input - def __init__(self, submodule): - super(ShortcutBlockSPSR, self).__init__() - self.sub = submodule - - def forward(self, x): - return x, self.sub - - def __repr__(self): - tmpstr = "Identity + \n|" - modstr = self.sub.__repr__().replace("\n", "\n|") - tmpstr = tmpstr + modstr - return tmpstr - - -def sequential(*args): - # Flatten Sequential. It unwraps nn.Sequential. - if len(args) == 1: - if isinstance(args[0], OrderedDict): - raise NotImplementedError("sequential does not support OrderedDict input.") - return args[0] # No sequential is needed. - modules = [] - for module in args: - if isinstance(module, nn.Sequential): - for submodule in module.children(): - modules.append(submodule) - elif isinstance(module, nn.Module): - modules.append(module) - return nn.Sequential(*modules) - - -ConvMode = Literal["CNA", "NAC", "CNAC"] - - -# 2x2x2 Conv Block -def conv_block_2c2( - in_nc, - out_nc, - act_type="relu", -): - return sequential( - nn.Conv2d(in_nc, out_nc, kernel_size=2, padding=1), - nn.Conv2d(out_nc, out_nc, kernel_size=2, padding=0), - act(act_type) if act_type else None, - ) - - -def conv_block( - in_nc: int, - out_nc: int, - kernel_size, - stride=1, - dilation=1, - groups=1, - bias=True, - pad_type="zero", - norm_type: str | None = None, - act_type: str | None = "relu", - mode: ConvMode = "CNA", - c2x2=False, -): - """ - Conv layer with padding, normalization, activation - mode: CNA --> Conv -> Norm -> Act - NAC --> Norm -> Act --> Conv (Identity Mappings in Deep Residual Networks, ECCV16) - """ - - if c2x2: - return conv_block_2c2(in_nc, out_nc, act_type=act_type) - - assert mode in ("CNA", "NAC", "CNAC"), "Wrong conv mode [{:s}]".format(mode) - padding = get_valid_padding(kernel_size, dilation) - p = pad(pad_type, padding) if pad_type and pad_type != "zero" else None - padding = padding if pad_type == "zero" else 0 - - c = nn.Conv2d( - in_nc, - out_nc, - kernel_size=kernel_size, - stride=stride, - padding=padding, - dilation=dilation, - bias=bias, - groups=groups, - ) - a = act(act_type) if act_type else None - if mode in ("CNA", "CNAC"): - n = norm(norm_type, out_nc) if norm_type else None - return sequential(p, c, n, a) - elif mode == "NAC": - if norm_type is None and act_type is not None: - a = act(act_type, inplace=False) - # Important! - # input----ReLU(inplace)----Conv--+----output - # |________________________| - # inplace ReLU will modify the input, therefore wrong output - n = norm(norm_type, in_nc) if norm_type else None - return sequential(n, a, p, c) - else: - assert False, f"Invalid conv mode {mode}" - - -#################### -# Useful blocks -#################### - - -class ResNetBlock(nn.Module): - """ - ResNet Block, 3-3 style - with extra residual scaling used in EDSR - (Enhanced Deep Residual Networks for Single Image Super-Resolution, CVPRW 17) - """ - - def __init__( - self, - in_nc, - mid_nc, - out_nc, - kernel_size=3, - stride=1, - dilation=1, - groups=1, - bias=True, - pad_type="zero", - norm_type=None, - act_type="relu", - mode: ConvMode = "CNA", - res_scale=1, - ): - super(ResNetBlock, self).__init__() - conv0 = conv_block( - in_nc, - mid_nc, - kernel_size, - stride, - dilation, - groups, - bias, - pad_type, - norm_type, - act_type, - mode, - ) - if mode == "CNA": - act_type = None - if mode == "CNAC": # Residual path: |-CNAC-| - act_type = None - norm_type = None - conv1 = conv_block( - mid_nc, - out_nc, - kernel_size, - stride, - dilation, - groups, - bias, - pad_type, - norm_type, - act_type, - mode, - ) - # if in_nc != out_nc: - # self.project = conv_block(in_nc, out_nc, 1, stride, dilation, 1, bias, pad_type, \ - # None, None) - # print('Need a projecter in ResNetBlock.') - # else: - # self.project = lambda x:x - self.res = sequential(conv0, conv1) - self.res_scale = res_scale - - def forward(self, x): - res = self.res(x).mul(self.res_scale) - return x + res - - -class RRDB(nn.Module): - """ - Residual in Residual Dense Block - (ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks) - """ - - def __init__( - self, - nf, - kernel_size=3, - gc=32, - stride=1, - bias: bool = True, - pad_type="zero", - norm_type=None, - act_type="leakyrelu", - mode: ConvMode = "CNA", - _convtype="Conv2D", - _spectral_norm=False, - plus=False, - c2x2=False, - ): - super(RRDB, self).__init__() - self.RDB1 = ResidualDenseBlock_5C( - nf, - kernel_size, - gc, - stride, - bias, - pad_type, - norm_type, - act_type, - mode, - plus=plus, - c2x2=c2x2, - ) - self.RDB2 = ResidualDenseBlock_5C( - nf, - kernel_size, - gc, - stride, - bias, - pad_type, - norm_type, - act_type, - mode, - plus=plus, - c2x2=c2x2, - ) - self.RDB3 = ResidualDenseBlock_5C( - nf, - kernel_size, - gc, - stride, - bias, - pad_type, - norm_type, - act_type, - mode, - plus=plus, - c2x2=c2x2, - ) - - def forward(self, x): - out = self.RDB1(x) - out = self.RDB2(out) - out = self.RDB3(out) - return out * 0.2 + x - - -class ResidualDenseBlock_5C(nn.Module): - """ - Residual Dense Block - style: 5 convs - The core module of paper: (Residual Dense Network for Image Super-Resolution, CVPR 18) - Modified options that can be used: - - "Partial Convolution based Padding" arXiv:1811.11718 - - "Spectral normalization" arXiv:1802.05957 - - "ICASSP 2020 - ESRGAN+ : Further Improving ESRGAN" N. C. - {Rakotonirina} and A. {Rasoanaivo} - - Args: - nf (int): Channel number of intermediate features (num_feat). - gc (int): Channels for each growth (num_grow_ch: growth channel, - i.e. intermediate channels). - convtype (str): the type of convolution to use. Default: 'Conv2D' - gaussian_noise (bool): enable the ESRGAN+ gaussian noise (no new - trainable parameters) - plus (bool): enable the additional residual paths from ESRGAN+ - (adds trainable parameters) - """ - - def __init__( - self, - nf=64, - kernel_size=3, - gc=32, - stride=1, - bias: bool = True, - pad_type="zero", - norm_type=None, - act_type="leakyrelu", - mode: ConvMode = "CNA", - plus=False, - c2x2=False, - ): - super(ResidualDenseBlock_5C, self).__init__() - - ## + - self.conv1x1 = conv1x1(nf, gc) if plus else None - ## + - - self.conv1 = conv_block( - nf, - gc, - kernel_size, - stride, - bias=bias, - pad_type=pad_type, - norm_type=norm_type, - act_type=act_type, - mode=mode, - c2x2=c2x2, - ) - self.conv2 = conv_block( - nf + gc, - gc, - kernel_size, - stride, - bias=bias, - pad_type=pad_type, - norm_type=norm_type, - act_type=act_type, - mode=mode, - c2x2=c2x2, - ) - self.conv3 = conv_block( - nf + 2 * gc, - gc, - kernel_size, - stride, - bias=bias, - pad_type=pad_type, - norm_type=norm_type, - act_type=act_type, - mode=mode, - c2x2=c2x2, - ) - self.conv4 = conv_block( - nf + 3 * gc, - gc, - kernel_size, - stride, - bias=bias, - pad_type=pad_type, - norm_type=norm_type, - act_type=act_type, - mode=mode, - c2x2=c2x2, - ) - if mode == "CNA": - last_act = None - else: - last_act = act_type - self.conv5 = conv_block( - nf + 4 * gc, - nf, - 3, - stride, - bias=bias, - pad_type=pad_type, - norm_type=norm_type, - act_type=last_act, - mode=mode, - c2x2=c2x2, - ) - - def forward(self, x): - x1 = self.conv1(x) - x2 = self.conv2(torch.cat((x, x1), 1)) - if self.conv1x1: - # pylint: disable=not-callable - x2 = x2 + self.conv1x1(x) # + - x3 = self.conv3(torch.cat((x, x1, x2), 1)) - x4 = self.conv4(torch.cat((x, x1, x2, x3), 1)) - if self.conv1x1: - x4 = x4 + x2 # + - x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1)) - return x5 * 0.2 + x - - -def conv1x1(in_planes, out_planes, stride=1): - return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) - - -#################### -# Upsampler -#################### - - -def pixelshuffle_block( - in_nc: int, - out_nc: int, - upscale_factor=2, - kernel_size=3, - stride=1, - bias=True, - pad_type="zero", - norm_type: str | None = None, - act_type="relu", -): - """ - Pixel shuffle layer - (Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional - Neural Network, CVPR17) - """ - conv = conv_block( - in_nc, - out_nc * (upscale_factor**2), - kernel_size, - stride, - bias=bias, - pad_type=pad_type, - norm_type=None, - act_type=None, - ) - pixel_shuffle = nn.PixelShuffle(upscale_factor) - - n = norm(norm_type, out_nc) if norm_type else None - a = act(act_type) if act_type else None - return sequential(conv, pixel_shuffle, n, a) - - -def upconv_block( - in_nc: int, - out_nc: int, - upscale_factor=2, - kernel_size=3, - stride=1, - bias=True, - pad_type="zero", - norm_type: str | None = None, - act_type="relu", - mode="nearest", - c2x2=False, -): - # Up conv - # described in https://distill.pub/2016/deconv-checkerboard/ - upsample = nn.Upsample(scale_factor=upscale_factor, mode=mode) - conv = conv_block( - in_nc, - out_nc, - kernel_size, - stride, - bias=bias, - pad_type=pad_type, - norm_type=norm_type, - act_type=act_type, - c2x2=c2x2, - ) - return sequential(upsample, conv) diff --git a/comfy_extras/chainner_models/architecture/face/LICENSE-GFPGAN b/comfy_extras/chainner_models/architecture/face/LICENSE-GFPGAN deleted file mode 100644 index 5ac273fd5..000000000 --- a/comfy_extras/chainner_models/architecture/face/LICENSE-GFPGAN +++ /dev/null @@ -1,351 +0,0 @@ -Tencent is pleased to support the open source community by making GFPGAN available. - -Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved. - -GFPGAN is licensed under the Apache License Version 2.0 except for the third-party components listed below. - - -Terms of the Apache License Version 2.0: ---------------------------------------------- -Apache License - -Version 2.0, January 2004 - -http://www.apache.org/licenses/ - -TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION -1. Definitions. - -“License” shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document. - -“Licensor” shall mean the copyright owner or entity authorized by the copyright owner that is granting the License. - -“Legal Entity” shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, “control” means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity. - -“You” (or “Your”) shall mean an individual or Legal Entity exercising permissions granted by this License. - -“Source” form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files. - -“Object” form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types. - -“Work” shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below). - -“Derivative Works” shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof. - -“Contribution” shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, “submitted” means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as “Not a Contribution.” - -“Contributor” shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work. - -2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form. - -3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed. - -4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions: - -You must give any other recipients of the Work or Derivative Works a copy of this License; and - -You must cause any modified files to carry prominent notices stating that You changed the files; and - -You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and - -If the Work includes a “NOTICE” text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License. - -You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License. - -5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions. - -6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file. - -7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License. - -8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages. - -9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability. - -END OF TERMS AND CONDITIONS - - - -Other dependencies and licenses: - - -Open Source Software licensed under the Apache 2.0 license and Other Licenses of the Third-Party Components therein: ---------------------------------------------- -1. basicsr -Copyright 2018-2020 BasicSR Authors - - -This BasicSR project is released under the Apache 2.0 license. - -A copy of Apache 2.0 is included in this file. - -StyleGAN2 -The codes are modified from the repository stylegan2-pytorch. Many thanks to the author - Kim Seonghyeon 😊 for translating from the official TensorFlow codes to PyTorch ones. Here is the license of stylegan2-pytorch. -The official repository is https://github.com/NVlabs/stylegan2, and here is the NVIDIA license. -DFDNet -The codes are largely modified from the repository DFDNet. Their license is Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. - -Terms of the Nvidia License: ---------------------------------------------- - -1. Definitions - -"Licensor" means any person or entity that distributes its Work. - -"Software" means the original work of authorship made available under -this License. - -"Work" means the Software and any additions to or derivative works of -the Software that are made available under this License. - -"Nvidia Processors" means any central processing unit (CPU), graphics -processing unit (GPU), field-programmable gate array (FPGA), -application-specific integrated circuit (ASIC) or any combination -thereof designed, made, sold, or provided by Nvidia or its affiliates. - -The terms "reproduce," "reproduction," "derivative works," and -"distribution" have the meaning as provided under U.S. copyright law; -provided, however, that for the purposes of this License, derivative -works shall not include works that remain separable from, or merely -link (or bind by name) to the interfaces of, the Work. - -Works, including the Software, are "made available" under this License -by including in or with the Work either (a) a copyright notice -referencing the applicability of this License to the Work, or (b) a -copy of this License. - -2. License Grants - - 2.1 Copyright Grant. Subject to the terms and conditions of this - License, each Licensor grants to you a perpetual, worldwide, - non-exclusive, royalty-free, copyright license to reproduce, - prepare derivative works of, publicly display, publicly perform, - sublicense and distribute its Work and any resulting derivative - works in any form. - -3. Limitations - - 3.1 Redistribution. You may reproduce or distribute the Work only - if (a) you do so under this License, (b) you include a complete - copy of this License with your distribution, and (c) you retain - without modification any copyright, patent, trademark, or - attribution notices that are present in the Work. - - 3.2 Derivative Works. You may specify that additional or different - terms apply to the use, reproduction, and distribution of your - derivative works of the Work ("Your Terms") only if (a) Your Terms - provide that the use limitation in Section 3.3 applies to your - derivative works, and (b) you identify the specific derivative - works that are subject to Your Terms. Notwithstanding Your Terms, - this License (including the redistribution requirements in Section - 3.1) will continue to apply to the Work itself. - - 3.3 Use Limitation. The Work and any derivative works thereof only - may be used or intended for use non-commercially. The Work or - derivative works thereof may be used or intended for use by Nvidia - or its affiliates commercially or non-commercially. As used herein, - "non-commercially" means for research or evaluation purposes only. - - 3.4 Patent Claims. If you bring or threaten to bring a patent claim - against any Licensor (including any claim, cross-claim or - counterclaim in a lawsuit) to enforce any patents that you allege - are infringed by any Work, then your rights under this License from - such Licensor (including the grants in Sections 2.1 and 2.2) will - terminate immediately. - - 3.5 Trademarks. This License does not grant any rights to use any - Licensor's or its affiliates' names, logos, or trademarks, except - as necessary to reproduce the notices described in this License. - - 3.6 Termination. If you violate any term of this License, then your - rights under this License (including the grants in Sections 2.1 and - 2.2) will terminate immediately. - -4. Disclaimer of Warranty. - -THE WORK IS PROVIDED "AS IS" WITHOUT WARRANTIES OR CONDITIONS OF ANY -KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WARRANTIES OR CONDITIONS OF -MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE OR -NON-INFRINGEMENT. YOU BEAR THE RISK OF UNDERTAKING ANY ACTIVITIES UNDER -THIS LICENSE. - -5. Limitation of Liability. - -EXCEPT AS PROHIBITED BY APPLICABLE LAW, IN NO EVENT AND UNDER NO LEGAL -THEORY, WHETHER IN TORT (INCLUDING NEGLIGENCE), CONTRACT, OR OTHERWISE -SHALL ANY LICENSOR BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY DIRECT, -INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES ARISING OUT OF -OR RELATED TO THIS LICENSE, THE USE OR INABILITY TO USE THE WORK -(INCLUDING BUT NOT LIMITED TO LOSS OF GOODWILL, BUSINESS INTERRUPTION, -LOST PROFITS OR DATA, COMPUTER FAILURE OR MALFUNCTION, OR ANY OTHER -COMMERCIAL DAMAGES OR LOSSES), EVEN IF THE LICENSOR HAS BEEN ADVISED OF -THE POSSIBILITY OF SUCH DAMAGES. - -MIT License - -Copyright (c) 2019 Kim Seonghyeon - -Permission is hereby granted, free of charge, to any person obtaining a copy -of this software and associated documentation files (the "Software"), to deal -in the Software without restriction, including without limitation the rights -to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -copies of the Software, and to permit persons to whom the Software is -furnished to do so, subject to the following conditions: - -The above copyright notice and this permission notice shall be included in all -copies or substantial portions of the Software. - -THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -SOFTWARE. - - - -Open Source Software licensed under the BSD 3-Clause license: ---------------------------------------------- -1. torchvision -Copyright (c) Soumith Chintala 2016, -All rights reserved. - -2. torch -Copyright (c) 2016- Facebook, Inc (Adam Paszke) -Copyright (c) 2014- Facebook, Inc (Soumith Chintala) -Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert) -Copyright (c) 2012-2014 Deepmind Technologies (Koray Kavukcuoglu) -Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu) -Copyright (c) 2011-2013 NYU (Clement Farabet) -Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston) -Copyright (c) 2006 Idiap Research Institute (Samy Bengio) -Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz) - - -Terms of the BSD 3-Clause License: ---------------------------------------------- -Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: - -1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. - -2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. - -3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission. - -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - - - -Open Source Software licensed under the BSD 3-Clause License and Other Licenses of the Third-Party Components therein: ---------------------------------------------- -1. numpy -Copyright (c) 2005-2020, NumPy Developers. -All rights reserved. - -A copy of BSD 3-Clause License is included in this file. - -The NumPy repository and source distributions bundle several libraries that are -compatibly licensed. We list these here. - -Name: Numpydoc -Files: doc/sphinxext/numpydoc/* -License: BSD-2-Clause - For details, see doc/sphinxext/LICENSE.txt - -Name: scipy-sphinx-theme -Files: doc/scipy-sphinx-theme/* -License: BSD-3-Clause AND PSF-2.0 AND Apache-2.0 - For details, see doc/scipy-sphinx-theme/LICENSE.txt - -Name: lapack-lite -Files: numpy/linalg/lapack_lite/* -License: BSD-3-Clause - For details, see numpy/linalg/lapack_lite/LICENSE.txt - -Name: tempita -Files: tools/npy_tempita/* -License: MIT - For details, see tools/npy_tempita/license.txt - -Name: dragon4 -Files: numpy/core/src/multiarray/dragon4.c -License: MIT - For license text, see numpy/core/src/multiarray/dragon4.c - - - -Open Source Software licensed under the MIT license: ---------------------------------------------- -1. facexlib -Copyright (c) 2020 Xintao Wang - -2. opencv-python -Copyright (c) Olli-Pekka Heinisuo -Please note that only files in cv2 package are used. - - -Terms of the MIT License: ---------------------------------------------- -Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: - -The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. - -THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. - - - -Open Source Software licensed under the MIT license and Other Licenses of the Third-Party Components therein: ---------------------------------------------- -1. tqdm -Copyright (c) 2013 noamraph - -`tqdm` is a product of collaborative work. -Unless otherwise stated, all authors (see commit logs) retain copyright -for their respective work, and release the work under the MIT licence -(text below). - -Exceptions or notable authors are listed below -in reverse chronological order: - -* files: * - MPLv2.0 2015-2020 (c) Casper da Costa-Luis - [casperdcl](https://github.com/casperdcl). -* files: tqdm/_tqdm.py - MIT 2016 (c) [PR #96] on behalf of Google Inc. -* files: tqdm/_tqdm.py setup.py README.rst MANIFEST.in .gitignore - MIT 2013 (c) Noam Yorav-Raphael, original author. - -[PR #96]: https://github.com/tqdm/tqdm/pull/96 - - -Mozilla Public Licence (MPL) v. 2.0 - Exhibit A ------------------------------------------------ - -This Source Code Form is subject to the terms of the -Mozilla Public License, v. 2.0. -If a copy of the MPL was not distributed with this file, -You can obtain one at https://mozilla.org/MPL/2.0/. - - -MIT License (MIT) ------------------ - -Copyright (c) 2013 noamraph - -Permission is hereby granted, free of charge, to any person obtaining a copy of -this software and associated documentation files (the "Software"), to deal in -the Software without restriction, including without limitation the rights to -use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of -the Software, and to permit persons to whom the Software is furnished to do so, -subject to the following conditions: - -The above copyright notice and this permission notice shall be included in all -copies or substantial portions of the Software. - -THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS -FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR -COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER -IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN -CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/comfy_extras/chainner_models/architecture/face/LICENSE-RestoreFormer b/comfy_extras/chainner_models/architecture/face/LICENSE-RestoreFormer deleted file mode 100644 index 5ac273fd5..000000000 --- a/comfy_extras/chainner_models/architecture/face/LICENSE-RestoreFormer +++ /dev/null @@ -1,351 +0,0 @@ -Tencent is pleased to support the open source community by making GFPGAN available. - -Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved. - -GFPGAN is licensed under the Apache License Version 2.0 except for the third-party components listed below. - - -Terms of the Apache License Version 2.0: ---------------------------------------------- -Apache License - -Version 2.0, January 2004 - -http://www.apache.org/licenses/ - -TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION -1. Definitions. - -“License” shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document. - -“Licensor” shall mean the copyright owner or entity authorized by the copyright owner that is granting the License. - -“Legal Entity” shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, “control” means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity. - -“You” (or “Your”) shall mean an individual or Legal Entity exercising permissions granted by this License. - -“Source” form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files. - -“Object” form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types. - -“Work” shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below). - -“Derivative Works” shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof. - -“Contribution” shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, “submitted” means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as “Not a Contribution.” - -“Contributor” shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work. - -2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form. - -3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed. - -4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions: - -You must give any other recipients of the Work or Derivative Works a copy of this License; and - -You must cause any modified files to carry prominent notices stating that You changed the files; and - -You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and - -If the Work includes a “NOTICE” text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License. - -You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License. - -5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions. - -6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file. - -7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License. - -8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages. - -9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability. - -END OF TERMS AND CONDITIONS - - - -Other dependencies and licenses: - - -Open Source Software licensed under the Apache 2.0 license and Other Licenses of the Third-Party Components therein: ---------------------------------------------- -1. basicsr -Copyright 2018-2020 BasicSR Authors - - -This BasicSR project is released under the Apache 2.0 license. - -A copy of Apache 2.0 is included in this file. - -StyleGAN2 -The codes are modified from the repository stylegan2-pytorch. Many thanks to the author - Kim Seonghyeon 😊 for translating from the official TensorFlow codes to PyTorch ones. Here is the license of stylegan2-pytorch. -The official repository is https://github.com/NVlabs/stylegan2, and here is the NVIDIA license. -DFDNet -The codes are largely modified from the repository DFDNet. Their license is Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. - -Terms of the Nvidia License: ---------------------------------------------- - -1. Definitions - -"Licensor" means any person or entity that distributes its Work. - -"Software" means the original work of authorship made available under -this License. - -"Work" means the Software and any additions to or derivative works of -the Software that are made available under this License. - -"Nvidia Processors" means any central processing unit (CPU), graphics -processing unit (GPU), field-programmable gate array (FPGA), -application-specific integrated circuit (ASIC) or any combination -thereof designed, made, sold, or provided by Nvidia or its affiliates. - -The terms "reproduce," "reproduction," "derivative works," and -"distribution" have the meaning as provided under U.S. copyright law; -provided, however, that for the purposes of this License, derivative -works shall not include works that remain separable from, or merely -link (or bind by name) to the interfaces of, the Work. - -Works, including the Software, are "made available" under this License -by including in or with the Work either (a) a copyright notice -referencing the applicability of this License to the Work, or (b) a -copy of this License. - -2. License Grants - - 2.1 Copyright Grant. Subject to the terms and conditions of this - License, each Licensor grants to you a perpetual, worldwide, - non-exclusive, royalty-free, copyright license to reproduce, - prepare derivative works of, publicly display, publicly perform, - sublicense and distribute its Work and any resulting derivative - works in any form. - -3. Limitations - - 3.1 Redistribution. You may reproduce or distribute the Work only - if (a) you do so under this License, (b) you include a complete - copy of this License with your distribution, and (c) you retain - without modification any copyright, patent, trademark, or - attribution notices that are present in the Work. - - 3.2 Derivative Works. You may specify that additional or different - terms apply to the use, reproduction, and distribution of your - derivative works of the Work ("Your Terms") only if (a) Your Terms - provide that the use limitation in Section 3.3 applies to your - derivative works, and (b) you identify the specific derivative - works that are subject to Your Terms. Notwithstanding Your Terms, - this License (including the redistribution requirements in Section - 3.1) will continue to apply to the Work itself. - - 3.3 Use Limitation. The Work and any derivative works thereof only - may be used or intended for use non-commercially. The Work or - derivative works thereof may be used or intended for use by Nvidia - or its affiliates commercially or non-commercially. As used herein, - "non-commercially" means for research or evaluation purposes only. - - 3.4 Patent Claims. If you bring or threaten to bring a patent claim - against any Licensor (including any claim, cross-claim or - counterclaim in a lawsuit) to enforce any patents that you allege - are infringed by any Work, then your rights under this License from - such Licensor (including the grants in Sections 2.1 and 2.2) will - terminate immediately. - - 3.5 Trademarks. This License does not grant any rights to use any - Licensor's or its affiliates' names, logos, or trademarks, except - as necessary to reproduce the notices described in this License. - - 3.6 Termination. If you violate any term of this License, then your - rights under this License (including the grants in Sections 2.1 and - 2.2) will terminate immediately. - -4. Disclaimer of Warranty. - -THE WORK IS PROVIDED "AS IS" WITHOUT WARRANTIES OR CONDITIONS OF ANY -KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WARRANTIES OR CONDITIONS OF -MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE OR -NON-INFRINGEMENT. YOU BEAR THE RISK OF UNDERTAKING ANY ACTIVITIES UNDER -THIS LICENSE. - -5. Limitation of Liability. - -EXCEPT AS PROHIBITED BY APPLICABLE LAW, IN NO EVENT AND UNDER NO LEGAL -THEORY, WHETHER IN TORT (INCLUDING NEGLIGENCE), CONTRACT, OR OTHERWISE -SHALL ANY LICENSOR BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY DIRECT, -INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES ARISING OUT OF -OR RELATED TO THIS LICENSE, THE USE OR INABILITY TO USE THE WORK -(INCLUDING BUT NOT LIMITED TO LOSS OF GOODWILL, BUSINESS INTERRUPTION, -LOST PROFITS OR DATA, COMPUTER FAILURE OR MALFUNCTION, OR ANY OTHER -COMMERCIAL DAMAGES OR LOSSES), EVEN IF THE LICENSOR HAS BEEN ADVISED OF -THE POSSIBILITY OF SUCH DAMAGES. - -MIT License - -Copyright (c) 2019 Kim Seonghyeon - -Permission is hereby granted, free of charge, to any person obtaining a copy -of this software and associated documentation files (the "Software"), to deal -in the Software without restriction, including without limitation the rights -to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -copies of the Software, and to permit persons to whom the Software is -furnished to do so, subject to the following conditions: - -The above copyright notice and this permission notice shall be included in all -copies or substantial portions of the Software. - -THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -SOFTWARE. - - - -Open Source Software licensed under the BSD 3-Clause license: ---------------------------------------------- -1. torchvision -Copyright (c) Soumith Chintala 2016, -All rights reserved. - -2. torch -Copyright (c) 2016- Facebook, Inc (Adam Paszke) -Copyright (c) 2014- Facebook, Inc (Soumith Chintala) -Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert) -Copyright (c) 2012-2014 Deepmind Technologies (Koray Kavukcuoglu) -Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu) -Copyright (c) 2011-2013 NYU (Clement Farabet) -Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston) -Copyright (c) 2006 Idiap Research Institute (Samy Bengio) -Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz) - - -Terms of the BSD 3-Clause License: ---------------------------------------------- -Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: - -1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. - -2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. - -3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission. - -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - - - -Open Source Software licensed under the BSD 3-Clause License and Other Licenses of the Third-Party Components therein: ---------------------------------------------- -1. numpy -Copyright (c) 2005-2020, NumPy Developers. -All rights reserved. - -A copy of BSD 3-Clause License is included in this file. - -The NumPy repository and source distributions bundle several libraries that are -compatibly licensed. We list these here. - -Name: Numpydoc -Files: doc/sphinxext/numpydoc/* -License: BSD-2-Clause - For details, see doc/sphinxext/LICENSE.txt - -Name: scipy-sphinx-theme -Files: doc/scipy-sphinx-theme/* -License: BSD-3-Clause AND PSF-2.0 AND Apache-2.0 - For details, see doc/scipy-sphinx-theme/LICENSE.txt - -Name: lapack-lite -Files: numpy/linalg/lapack_lite/* -License: BSD-3-Clause - For details, see numpy/linalg/lapack_lite/LICENSE.txt - -Name: tempita -Files: tools/npy_tempita/* -License: MIT - For details, see tools/npy_tempita/license.txt - -Name: dragon4 -Files: numpy/core/src/multiarray/dragon4.c -License: MIT - For license text, see numpy/core/src/multiarray/dragon4.c - - - -Open Source Software licensed under the MIT license: ---------------------------------------------- -1. facexlib -Copyright (c) 2020 Xintao Wang - -2. opencv-python -Copyright (c) Olli-Pekka Heinisuo -Please note that only files in cv2 package are used. - - -Terms of the MIT License: ---------------------------------------------- -Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: - -The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. - -THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. - - - -Open Source Software licensed under the MIT license and Other Licenses of the Third-Party Components therein: ---------------------------------------------- -1. tqdm -Copyright (c) 2013 noamraph - -`tqdm` is a product of collaborative work. -Unless otherwise stated, all authors (see commit logs) retain copyright -for their respective work, and release the work under the MIT licence -(text below). - -Exceptions or notable authors are listed below -in reverse chronological order: - -* files: * - MPLv2.0 2015-2020 (c) Casper da Costa-Luis - [casperdcl](https://github.com/casperdcl). -* files: tqdm/_tqdm.py - MIT 2016 (c) [PR #96] on behalf of Google Inc. -* files: tqdm/_tqdm.py setup.py README.rst MANIFEST.in .gitignore - MIT 2013 (c) Noam Yorav-Raphael, original author. - -[PR #96]: https://github.com/tqdm/tqdm/pull/96 - - -Mozilla Public Licence (MPL) v. 2.0 - Exhibit A ------------------------------------------------ - -This Source Code Form is subject to the terms of the -Mozilla Public License, v. 2.0. -If a copy of the MPL was not distributed with this file, -You can obtain one at https://mozilla.org/MPL/2.0/. - - -MIT License (MIT) ------------------ - -Copyright (c) 2013 noamraph - -Permission is hereby granted, free of charge, to any person obtaining a copy of -this software and associated documentation files (the "Software"), to deal in -the Software without restriction, including without limitation the rights to -use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of -the Software, and to permit persons to whom the Software is furnished to do so, -subject to the following conditions: - -The above copyright notice and this permission notice shall be included in all -copies or substantial portions of the Software. - -THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS -FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR -COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER -IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN -CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. \ No newline at end of file diff --git a/comfy_extras/chainner_models/architecture/face/LICENSE-codeformer b/comfy_extras/chainner_models/architecture/face/LICENSE-codeformer deleted file mode 100644 index be6c4ed80..000000000 --- a/comfy_extras/chainner_models/architecture/face/LICENSE-codeformer +++ /dev/null @@ -1,35 +0,0 @@ -S-Lab License 1.0 - -Copyright 2022 S-Lab - -Redistribution and use for non-commercial purpose in source and -binary forms, with or without modification, are permitted provided -that the following conditions are met: - -1. Redistributions of source code must retain the above copyright - notice, this list of conditions and the following disclaimer. - -2. Redistributions in binary form must reproduce the above copyright - notice, this list of conditions and the following disclaimer in - the documentation and/or other materials provided with the - distribution. - -3. Neither the name of the copyright holder nor the names of its - contributors may be used to endorse or promote products derived - from this software without specific prior written permission. - -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT -LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT -HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, -SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT -LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, -DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY -THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT -(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE -OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - -In the event that redistribution and/or use for commercial purpose in -source or binary forms, with or without modification is required, -please contact the contributor(s) of the work. diff --git a/comfy_extras/chainner_models/architecture/face/arcface_arch.py b/comfy_extras/chainner_models/architecture/face/arcface_arch.py deleted file mode 100644 index b548af059..000000000 --- a/comfy_extras/chainner_models/architecture/face/arcface_arch.py +++ /dev/null @@ -1,265 +0,0 @@ -import torch.nn as nn - - -def conv3x3(inplanes, outplanes, stride=1): - """A simple wrapper for 3x3 convolution with padding. - - Args: - inplanes (int): Channel number of inputs. - outplanes (int): Channel number of outputs. - stride (int): Stride in convolution. Default: 1. - """ - return nn.Conv2d( - inplanes, outplanes, kernel_size=3, stride=stride, padding=1, bias=False - ) - - -class BasicBlock(nn.Module): - """Basic residual block used in the ResNetArcFace architecture. - - Args: - inplanes (int): Channel number of inputs. - planes (int): Channel number of outputs. - stride (int): Stride in convolution. Default: 1. - downsample (nn.Module): The downsample module. Default: None. - """ - - expansion = 1 # output channel expansion ratio - - def __init__(self, inplanes, planes, stride=1, downsample=None): - super(BasicBlock, self).__init__() - self.conv1 = conv3x3(inplanes, planes, stride) - self.bn1 = nn.BatchNorm2d(planes) - self.relu = nn.ReLU(inplace=True) - self.conv2 = conv3x3(planes, planes) - self.bn2 = nn.BatchNorm2d(planes) - self.downsample = downsample - self.stride = stride - - def forward(self, x): - residual = x - - out = self.conv1(x) - out = self.bn1(out) - out = self.relu(out) - - out = self.conv2(out) - out = self.bn2(out) - - if self.downsample is not None: - residual = self.downsample(x) - - out += residual - out = self.relu(out) - - return out - - -class IRBlock(nn.Module): - """Improved residual block (IR Block) used in the ResNetArcFace architecture. - - Args: - inplanes (int): Channel number of inputs. - planes (int): Channel number of outputs. - stride (int): Stride in convolution. Default: 1. - downsample (nn.Module): The downsample module. Default: None. - use_se (bool): Whether use the SEBlock (squeeze and excitation block). Default: True. - """ - - expansion = 1 # output channel expansion ratio - - def __init__(self, inplanes, planes, stride=1, downsample=None, use_se=True): - super(IRBlock, self).__init__() - self.bn0 = nn.BatchNorm2d(inplanes) - self.conv1 = conv3x3(inplanes, inplanes) - self.bn1 = nn.BatchNorm2d(inplanes) - self.prelu = nn.PReLU() - self.conv2 = conv3x3(inplanes, planes, stride) - self.bn2 = nn.BatchNorm2d(planes) - self.downsample = downsample - self.stride = stride - self.use_se = use_se - if self.use_se: - self.se = SEBlock(planes) - - def forward(self, x): - residual = x - out = self.bn0(x) - out = self.conv1(out) - out = self.bn1(out) - out = self.prelu(out) - - out = self.conv2(out) - out = self.bn2(out) - if self.use_se: - out = self.se(out) - - if self.downsample is not None: - residual = self.downsample(x) - - out += residual - out = self.prelu(out) - - return out - - -class Bottleneck(nn.Module): - """Bottleneck block used in the ResNetArcFace architecture. - - Args: - inplanes (int): Channel number of inputs. - planes (int): Channel number of outputs. - stride (int): Stride in convolution. Default: 1. - downsample (nn.Module): The downsample module. Default: None. - """ - - expansion = 4 # output channel expansion ratio - - def __init__(self, inplanes, planes, stride=1, downsample=None): - super(Bottleneck, self).__init__() - self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) - self.bn1 = nn.BatchNorm2d(planes) - self.conv2 = nn.Conv2d( - planes, planes, kernel_size=3, stride=stride, padding=1, bias=False - ) - self.bn2 = nn.BatchNorm2d(planes) - self.conv3 = nn.Conv2d( - planes, planes * self.expansion, kernel_size=1, bias=False - ) - self.bn3 = nn.BatchNorm2d(planes * self.expansion) - self.relu = nn.ReLU(inplace=True) - self.downsample = downsample - self.stride = stride - - def forward(self, x): - residual = x - - out = self.conv1(x) - out = self.bn1(out) - out = self.relu(out) - - out = self.conv2(out) - out = self.bn2(out) - out = self.relu(out) - - out = self.conv3(out) - out = self.bn3(out) - - if self.downsample is not None: - residual = self.downsample(x) - - out += residual - out = self.relu(out) - - return out - - -class SEBlock(nn.Module): - """The squeeze-and-excitation block (SEBlock) used in the IRBlock. - - Args: - channel (int): Channel number of inputs. - reduction (int): Channel reduction ration. Default: 16. - """ - - def __init__(self, channel, reduction=16): - super(SEBlock, self).__init__() - self.avg_pool = nn.AdaptiveAvgPool2d( - 1 - ) # pool to 1x1 without spatial information - self.fc = nn.Sequential( - nn.Linear(channel, channel // reduction), - nn.PReLU(), - nn.Linear(channel // reduction, channel), - nn.Sigmoid(), - ) - - def forward(self, x): - b, c, _, _ = x.size() - y = self.avg_pool(x).view(b, c) - y = self.fc(y).view(b, c, 1, 1) - return x * y - - -class ResNetArcFace(nn.Module): - """ArcFace with ResNet architectures. - - Ref: ArcFace: Additive Angular Margin Loss for Deep Face Recognition. - - Args: - block (str): Block used in the ArcFace architecture. - layers (tuple(int)): Block numbers in each layer. - use_se (bool): Whether use the SEBlock (squeeze and excitation block). Default: True. - """ - - def __init__(self, block, layers, use_se=True): - if block == "IRBlock": - block = IRBlock - self.inplanes = 64 - self.use_se = use_se - super(ResNetArcFace, self).__init__() - - self.conv1 = nn.Conv2d(1, 64, kernel_size=3, padding=1, bias=False) - self.bn1 = nn.BatchNorm2d(64) - self.prelu = nn.PReLU() - self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2) - self.layer1 = self._make_layer(block, 64, layers[0]) - self.layer2 = self._make_layer(block, 128, layers[1], stride=2) - self.layer3 = self._make_layer(block, 256, layers[2], stride=2) - self.layer4 = self._make_layer(block, 512, layers[3], stride=2) - self.bn4 = nn.BatchNorm2d(512) - self.dropout = nn.Dropout() - self.fc5 = nn.Linear(512 * 8 * 8, 512) - self.bn5 = nn.BatchNorm1d(512) - - # initialization - for m in self.modules(): - if isinstance(m, nn.Conv2d): - nn.init.xavier_normal_(m.weight) - elif isinstance(m, nn.BatchNorm2d) or isinstance(m, nn.BatchNorm1d): - nn.init.constant_(m.weight, 1) - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.Linear): - nn.init.xavier_normal_(m.weight) - nn.init.constant_(m.bias, 0) - - def _make_layer(self, block, planes, num_blocks, stride=1): - downsample = None - if stride != 1 or self.inplanes != planes * block.expansion: - downsample = nn.Sequential( - nn.Conv2d( - self.inplanes, - planes * block.expansion, - kernel_size=1, - stride=stride, - bias=False, - ), - nn.BatchNorm2d(planes * block.expansion), - ) - layers = [] - layers.append( - block(self.inplanes, planes, stride, downsample, use_se=self.use_se) - ) - self.inplanes = planes - for _ in range(1, num_blocks): - layers.append(block(self.inplanes, planes, use_se=self.use_se)) - - return nn.Sequential(*layers) - - def forward(self, x): - x = self.conv1(x) - x = self.bn1(x) - x = self.prelu(x) - x = self.maxpool(x) - - x = self.layer1(x) - x = self.layer2(x) - x = self.layer3(x) - x = self.layer4(x) - x = self.bn4(x) - x = self.dropout(x) - x = x.view(x.size(0), -1) - x = self.fc5(x) - x = self.bn5(x) - - return x diff --git a/comfy_extras/chainner_models/architecture/face/codeformer.py b/comfy_extras/chainner_models/architecture/face/codeformer.py deleted file mode 100644 index 066140078..000000000 --- a/comfy_extras/chainner_models/architecture/face/codeformer.py +++ /dev/null @@ -1,790 +0,0 @@ -""" -Modified from https://github.com/sczhou/CodeFormer -VQGAN code, adapted from the original created by the Unleashing Transformers authors: -https://github.com/samb-t/unleashing-transformers/blob/master/models/vqgan.py -This verison of the arch specifically was gathered from an old version of GFPGAN. If this is a problem, please contact me. -""" -import math -from typing import Optional - -import torch -import torch.nn as nn -import torch.nn.functional as F -import logging as logger -from torch import Tensor - - -class VectorQuantizer(nn.Module): - def __init__(self, codebook_size, emb_dim, beta): - super(VectorQuantizer, self).__init__() - self.codebook_size = codebook_size # number of embeddings - self.emb_dim = emb_dim # dimension of embedding - self.beta = beta # commitment cost used in loss term, beta * ||z_e(x)-sg[e]||^2 - self.embedding = nn.Embedding(self.codebook_size, self.emb_dim) - self.embedding.weight.data.uniform_( - -1.0 / self.codebook_size, 1.0 / self.codebook_size - ) - - def forward(self, z): - # reshape z -> (batch, height, width, channel) and flatten - z = z.permute(0, 2, 3, 1).contiguous() - z_flattened = z.view(-1, self.emb_dim) - - # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z - d = ( - (z_flattened**2).sum(dim=1, keepdim=True) - + (self.embedding.weight**2).sum(1) - - 2 * torch.matmul(z_flattened, self.embedding.weight.t()) - ) - - mean_distance = torch.mean(d) - # find closest encodings - # min_encoding_indices = torch.argmin(d, dim=1).unsqueeze(1) - min_encoding_scores, min_encoding_indices = torch.topk( - d, 1, dim=1, largest=False - ) - # [0-1], higher score, higher confidence - min_encoding_scores = torch.exp(-min_encoding_scores / 10) - - min_encodings = torch.zeros( - min_encoding_indices.shape[0], self.codebook_size - ).to(z) - min_encodings.scatter_(1, min_encoding_indices, 1) - - # get quantized latent vectors - z_q = torch.matmul(min_encodings, self.embedding.weight).view(z.shape) - # compute loss for embedding - loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * torch.mean( - (z_q - z.detach()) ** 2 - ) - # preserve gradients - z_q = z + (z_q - z).detach() - - # perplexity - e_mean = torch.mean(min_encodings, dim=0) - perplexity = torch.exp(-torch.sum(e_mean * torch.log(e_mean + 1e-10))) - # reshape back to match original input shape - z_q = z_q.permute(0, 3, 1, 2).contiguous() - - return ( - z_q, - loss, - { - "perplexity": perplexity, - "min_encodings": min_encodings, - "min_encoding_indices": min_encoding_indices, - "min_encoding_scores": min_encoding_scores, - "mean_distance": mean_distance, - }, - ) - - def get_codebook_feat(self, indices, shape): - # input indices: batch*token_num -> (batch*token_num)*1 - # shape: batch, height, width, channel - indices = indices.view(-1, 1) - min_encodings = torch.zeros(indices.shape[0], self.codebook_size).to(indices) - min_encodings.scatter_(1, indices, 1) - # get quantized latent vectors - z_q = torch.matmul(min_encodings.float(), self.embedding.weight) - - if shape is not None: # reshape back to match original input shape - z_q = z_q.view(shape).permute(0, 3, 1, 2).contiguous() - - return z_q - - -class GumbelQuantizer(nn.Module): - def __init__( - self, - codebook_size, - emb_dim, - num_hiddens, - straight_through=False, - kl_weight=5e-4, - temp_init=1.0, - ): - super().__init__() - self.codebook_size = codebook_size # number of embeddings - self.emb_dim = emb_dim # dimension of embedding - self.straight_through = straight_through - self.temperature = temp_init - self.kl_weight = kl_weight - self.proj = nn.Conv2d( - num_hiddens, codebook_size, 1 - ) # projects last encoder layer to quantized logits - self.embed = nn.Embedding(codebook_size, emb_dim) - - def forward(self, z): - hard = self.straight_through if self.training else True - - logits = self.proj(z) - - soft_one_hot = F.gumbel_softmax(logits, tau=self.temperature, dim=1, hard=hard) - - z_q = torch.einsum("b n h w, n d -> b d h w", soft_one_hot, self.embed.weight) - - # + kl divergence to the prior loss - qy = F.softmax(logits, dim=1) - diff = ( - self.kl_weight - * torch.sum(qy * torch.log(qy * self.codebook_size + 1e-10), dim=1).mean() - ) - min_encoding_indices = soft_one_hot.argmax(dim=1) - - return z_q, diff, {"min_encoding_indices": min_encoding_indices} - - -class Downsample(nn.Module): - def __init__(self, in_channels): - super().__init__() - self.conv = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=3, stride=2, padding=0 - ) - - def forward(self, x): - pad = (0, 1, 0, 1) - x = torch.nn.functional.pad(x, pad, mode="constant", value=0) - x = self.conv(x) - return x - - -class Upsample(nn.Module): - def __init__(self, in_channels): - super().__init__() - self.conv = nn.Conv2d( - in_channels, in_channels, kernel_size=3, stride=1, padding=1 - ) - - def forward(self, x): - x = F.interpolate(x, scale_factor=2.0, mode="nearest") - x = self.conv(x) - - return x - - -class AttnBlock(nn.Module): - def __init__(self, in_channels): - super().__init__() - self.in_channels = in_channels - - self.norm = normalize(in_channels) - self.q = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=1, stride=1, padding=0 - ) - self.k = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=1, stride=1, padding=0 - ) - self.v = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=1, stride=1, padding=0 - ) - self.proj_out = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=1, stride=1, padding=0 - ) - - def forward(self, x): - h_ = x - h_ = self.norm(h_) - q = self.q(h_) - k = self.k(h_) - v = self.v(h_) - - # compute attention - b, c, h, w = q.shape - q = q.reshape(b, c, h * w) - q = q.permute(0, 2, 1) - k = k.reshape(b, c, h * w) - w_ = torch.bmm(q, k) - w_ = w_ * (int(c) ** (-0.5)) - w_ = F.softmax(w_, dim=2) - - # attend to values - v = v.reshape(b, c, h * w) - w_ = w_.permute(0, 2, 1) - h_ = torch.bmm(v, w_) - h_ = h_.reshape(b, c, h, w) - - h_ = self.proj_out(h_) - - return x + h_ - - -class Encoder(nn.Module): - def __init__( - self, - in_channels, - nf, - out_channels, - ch_mult, - num_res_blocks, - resolution, - attn_resolutions, - ): - super().__init__() - self.nf = nf - self.num_resolutions = len(ch_mult) - self.num_res_blocks = num_res_blocks - self.resolution = resolution - self.attn_resolutions = attn_resolutions - - curr_res = self.resolution - in_ch_mult = (1,) + tuple(ch_mult) - - blocks = [] - # initial convultion - blocks.append(nn.Conv2d(in_channels, nf, kernel_size=3, stride=1, padding=1)) - - # residual and downsampling blocks, with attention on smaller res (16x16) - for i in range(self.num_resolutions): - block_in_ch = nf * in_ch_mult[i] - block_out_ch = nf * ch_mult[i] - for _ in range(self.num_res_blocks): - blocks.append(ResBlock(block_in_ch, block_out_ch)) - block_in_ch = block_out_ch - if curr_res in attn_resolutions: - blocks.append(AttnBlock(block_in_ch)) - - if i != self.num_resolutions - 1: - blocks.append(Downsample(block_in_ch)) - curr_res = curr_res // 2 - - # non-local attention block - blocks.append(ResBlock(block_in_ch, block_in_ch)) # type: ignore - blocks.append(AttnBlock(block_in_ch)) # type: ignore - blocks.append(ResBlock(block_in_ch, block_in_ch)) # type: ignore - - # normalise and convert to latent size - blocks.append(normalize(block_in_ch)) # type: ignore - blocks.append( - nn.Conv2d(block_in_ch, out_channels, kernel_size=3, stride=1, padding=1) # type: ignore - ) - self.blocks = nn.ModuleList(blocks) - - def forward(self, x): - for block in self.blocks: - x = block(x) - - return x - - -class Generator(nn.Module): - def __init__(self, nf, ch_mult, res_blocks, img_size, attn_resolutions, emb_dim): - super().__init__() - self.nf = nf - self.ch_mult = ch_mult - self.num_resolutions = len(self.ch_mult) - self.num_res_blocks = res_blocks - self.resolution = img_size - self.attn_resolutions = attn_resolutions - self.in_channels = emb_dim - self.out_channels = 3 - block_in_ch = self.nf * self.ch_mult[-1] - curr_res = self.resolution // 2 ** (self.num_resolutions - 1) - - blocks = [] - # initial conv - blocks.append( - nn.Conv2d(self.in_channels, block_in_ch, kernel_size=3, stride=1, padding=1) - ) - - # non-local attention block - blocks.append(ResBlock(block_in_ch, block_in_ch)) - blocks.append(AttnBlock(block_in_ch)) - blocks.append(ResBlock(block_in_ch, block_in_ch)) - - for i in reversed(range(self.num_resolutions)): - block_out_ch = self.nf * self.ch_mult[i] - - for _ in range(self.num_res_blocks): - blocks.append(ResBlock(block_in_ch, block_out_ch)) - block_in_ch = block_out_ch - - if curr_res in self.attn_resolutions: - blocks.append(AttnBlock(block_in_ch)) - - if i != 0: - blocks.append(Upsample(block_in_ch)) - curr_res = curr_res * 2 - - blocks.append(normalize(block_in_ch)) - blocks.append( - nn.Conv2d( - block_in_ch, self.out_channels, kernel_size=3, stride=1, padding=1 - ) - ) - - self.blocks = nn.ModuleList(blocks) - - def forward(self, x): - for block in self.blocks: - x = block(x) - - return x - - -class VQAutoEncoder(nn.Module): - def __init__( - self, - img_size, - nf, - ch_mult, - quantizer="nearest", - res_blocks=2, - attn_resolutions=[16], - codebook_size=1024, - emb_dim=256, - beta=0.25, - gumbel_straight_through=False, - gumbel_kl_weight=1e-8, - model_path=None, - ): - super().__init__() - self.in_channels = 3 - self.nf = nf - self.n_blocks = res_blocks - self.codebook_size = codebook_size - self.embed_dim = emb_dim - self.ch_mult = ch_mult - self.resolution = img_size - self.attn_resolutions = attn_resolutions - self.quantizer_type = quantizer - self.encoder = Encoder( - self.in_channels, - self.nf, - self.embed_dim, - self.ch_mult, - self.n_blocks, - self.resolution, - self.attn_resolutions, - ) - if self.quantizer_type == "nearest": - self.beta = beta # 0.25 - self.quantize = VectorQuantizer( - self.codebook_size, self.embed_dim, self.beta - ) - elif self.quantizer_type == "gumbel": - self.gumbel_num_hiddens = emb_dim - self.straight_through = gumbel_straight_through - self.kl_weight = gumbel_kl_weight - self.quantize = GumbelQuantizer( - self.codebook_size, - self.embed_dim, - self.gumbel_num_hiddens, - self.straight_through, - self.kl_weight, - ) - self.generator = Generator( - nf, ch_mult, res_blocks, img_size, attn_resolutions, emb_dim - ) - - if model_path is not None: - chkpt = torch.load(model_path, map_location="cpu") - if "params_ema" in chkpt: - self.load_state_dict( - torch.load(model_path, map_location="cpu")["params_ema"] - ) - logger.info(f"vqgan is loaded from: {model_path} [params_ema]") - elif "params" in chkpt: - self.load_state_dict( - torch.load(model_path, map_location="cpu")["params"] - ) - logger.info(f"vqgan is loaded from: {model_path} [params]") - else: - raise ValueError("Wrong params!") - - def forward(self, x): - x = self.encoder(x) - quant, codebook_loss, quant_stats = self.quantize(x) - x = self.generator(quant) - return x, codebook_loss, quant_stats - - -def calc_mean_std(feat, eps=1e-5): - """Calculate mean and std for adaptive_instance_normalization. - Args: - feat (Tensor): 4D tensor. - eps (float): A small value added to the variance to avoid - divide-by-zero. Default: 1e-5. - """ - size = feat.size() - assert len(size) == 4, "The input feature should be 4D tensor." - b, c = size[:2] - feat_var = feat.view(b, c, -1).var(dim=2) + eps - feat_std = feat_var.sqrt().view(b, c, 1, 1) - feat_mean = feat.view(b, c, -1).mean(dim=2).view(b, c, 1, 1) - return feat_mean, feat_std - - -def adaptive_instance_normalization(content_feat, style_feat): - """Adaptive instance normalization. - Adjust the reference features to have the similar color and illuminations - as those in the degradate features. - Args: - content_feat (Tensor): The reference feature. - style_feat (Tensor): The degradate features. - """ - size = content_feat.size() - style_mean, style_std = calc_mean_std(style_feat) - content_mean, content_std = calc_mean_std(content_feat) - normalized_feat = (content_feat - content_mean.expand(size)) / content_std.expand( - size - ) - return normalized_feat * style_std.expand(size) + style_mean.expand(size) - - -class PositionEmbeddingSine(nn.Module): - """ - This is a more standard version of the position embedding, very similar to the one - used by the Attention is all you need paper, generalized to work on images. - """ - - def __init__( - self, num_pos_feats=64, temperature=10000, normalize=False, scale=None - ): - super().__init__() - self.num_pos_feats = num_pos_feats - self.temperature = temperature - self.normalize = normalize - if scale is not None and normalize is False: - raise ValueError("normalize should be True if scale is passed") - if scale is None: - scale = 2 * math.pi - self.scale = scale - - def forward(self, x, mask=None): - if mask is None: - mask = torch.zeros( - (x.size(0), x.size(2), x.size(3)), device=x.device, dtype=torch.bool - ) - not_mask = ~mask # pylint: disable=invalid-unary-operand-type - y_embed = not_mask.cumsum(1, dtype=torch.float32) - x_embed = not_mask.cumsum(2, dtype=torch.float32) - if self.normalize: - eps = 1e-6 - y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale - x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale - - dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device) - dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats) - - pos_x = x_embed[:, :, :, None] / dim_t - pos_y = y_embed[:, :, :, None] / dim_t - pos_x = torch.stack( - (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4 - ).flatten(3) - pos_y = torch.stack( - (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4 - ).flatten(3) - pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2) - return pos - - -def _get_activation_fn(activation): - """Return an activation function given a string""" - if activation == "relu": - return F.relu - if activation == "gelu": - return F.gelu - if activation == "glu": - return F.glu - raise RuntimeError(f"activation should be relu/gelu, not {activation}.") - - -class TransformerSALayer(nn.Module): - def __init__( - self, embed_dim, nhead=8, dim_mlp=2048, dropout=0.0, activation="gelu" - ): - super().__init__() - self.self_attn = nn.MultiheadAttention(embed_dim, nhead, dropout=dropout) - # Implementation of Feedforward model - MLP - self.linear1 = nn.Linear(embed_dim, dim_mlp) - self.dropout = nn.Dropout(dropout) - self.linear2 = nn.Linear(dim_mlp, embed_dim) - - self.norm1 = nn.LayerNorm(embed_dim) - self.norm2 = nn.LayerNorm(embed_dim) - self.dropout1 = nn.Dropout(dropout) - self.dropout2 = nn.Dropout(dropout) - - self.activation = _get_activation_fn(activation) - - def with_pos_embed(self, tensor, pos: Optional[Tensor]): - return tensor if pos is None else tensor + pos - - def forward( - self, - tgt, - tgt_mask: Optional[Tensor] = None, - tgt_key_padding_mask: Optional[Tensor] = None, - query_pos: Optional[Tensor] = None, - ): - # self attention - tgt2 = self.norm1(tgt) - q = k = self.with_pos_embed(tgt2, query_pos) - tgt2 = self.self_attn( - q, k, value=tgt2, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask - )[0] - tgt = tgt + self.dropout1(tgt2) - - # ffn - tgt2 = self.norm2(tgt) - tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2)))) - tgt = tgt + self.dropout2(tgt2) - return tgt - - -def normalize(in_channels): - return torch.nn.GroupNorm( - num_groups=32, num_channels=in_channels, eps=1e-6, affine=True - ) - - -@torch.jit.script # type: ignore -def swish(x): - return x * torch.sigmoid(x) - - -class ResBlock(nn.Module): - def __init__(self, in_channels, out_channels=None): - super(ResBlock, self).__init__() - self.in_channels = in_channels - self.out_channels = in_channels if out_channels is None else out_channels - self.norm1 = normalize(in_channels) - self.conv1 = nn.Conv2d( - in_channels, out_channels, kernel_size=3, stride=1, padding=1 # type: ignore - ) - self.norm2 = normalize(out_channels) - self.conv2 = nn.Conv2d( - out_channels, out_channels, kernel_size=3, stride=1, padding=1 # type: ignore - ) - if self.in_channels != self.out_channels: - self.conv_out = nn.Conv2d( - in_channels, out_channels, kernel_size=1, stride=1, padding=0 # type: ignore - ) - - def forward(self, x_in): - x = x_in - x = self.norm1(x) - x = swish(x) - x = self.conv1(x) - x = self.norm2(x) - x = swish(x) - x = self.conv2(x) - if self.in_channels != self.out_channels: - x_in = self.conv_out(x_in) - - return x + x_in - - -class Fuse_sft_block(nn.Module): - def __init__(self, in_ch, out_ch): - super().__init__() - self.encode_enc = ResBlock(2 * in_ch, out_ch) - - self.scale = nn.Sequential( - nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1), - nn.LeakyReLU(0.2, True), - nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1), - ) - - self.shift = nn.Sequential( - nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1), - nn.LeakyReLU(0.2, True), - nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1), - ) - - def forward(self, enc_feat, dec_feat, w=1): - enc_feat = self.encode_enc(torch.cat([enc_feat, dec_feat], dim=1)) - scale = self.scale(enc_feat) - shift = self.shift(enc_feat) - residual = w * (dec_feat * scale + shift) - out = dec_feat + residual - return out - - -class CodeFormer(VQAutoEncoder): - def __init__(self, state_dict): - dim_embd = 512 - n_head = 8 - n_layers = 9 - codebook_size = 1024 - latent_size = 256 - connect_list = ["32", "64", "128", "256"] - fix_modules = ["quantize", "generator"] - - # This is just a guess as I only have one model to look at - position_emb = state_dict["position_emb"] - dim_embd = position_emb.shape[1] - latent_size = position_emb.shape[0] - - try: - n_layers = len( - set([x.split(".")[1] for x in state_dict.keys() if "ft_layers" in x]) - ) - except: - pass - - codebook_size = state_dict["quantize.embedding.weight"].shape[0] - - # This is also just another guess - n_head_exp = ( - state_dict["ft_layers.0.self_attn.in_proj_weight"].shape[0] // dim_embd - ) - n_head = 2**n_head_exp - - in_nc = state_dict["encoder.blocks.0.weight"].shape[1] - - self.model_arch = "CodeFormer" - self.sub_type = "Face SR" - self.scale = 8 - self.in_nc = in_nc - self.out_nc = in_nc - - self.state = state_dict - - self.supports_fp16 = False - self.supports_bf16 = True - self.min_size_restriction = 16 - - super(CodeFormer, self).__init__( - 512, 64, [1, 2, 2, 4, 4, 8], "nearest", 2, [16], codebook_size - ) - - if fix_modules is not None: - for module in fix_modules: - for param in getattr(self, module).parameters(): - param.requires_grad = False - - self.connect_list = connect_list - self.n_layers = n_layers - self.dim_embd = dim_embd - self.dim_mlp = dim_embd * 2 - - self.position_emb = nn.Parameter(torch.zeros(latent_size, self.dim_embd)) # type: ignore - self.feat_emb = nn.Linear(256, self.dim_embd) - - # transformer - self.ft_layers = nn.Sequential( - *[ - TransformerSALayer( - embed_dim=dim_embd, nhead=n_head, dim_mlp=self.dim_mlp, dropout=0.0 - ) - for _ in range(self.n_layers) - ] - ) - - # logits_predict head - self.idx_pred_layer = nn.Sequential( - nn.LayerNorm(dim_embd), nn.Linear(dim_embd, codebook_size, bias=False) - ) - - self.channels = { - "16": 512, - "32": 256, - "64": 256, - "128": 128, - "256": 128, - "512": 64, - } - - # after second residual block for > 16, before attn layer for ==16 - self.fuse_encoder_block = { - "512": 2, - "256": 5, - "128": 8, - "64": 11, - "32": 14, - "16": 18, - } - # after first residual block for > 16, before attn layer for ==16 - self.fuse_generator_block = { - "16": 6, - "32": 9, - "64": 12, - "128": 15, - "256": 18, - "512": 21, - } - - # fuse_convs_dict - self.fuse_convs_dict = nn.ModuleDict() - for f_size in self.connect_list: - in_ch = self.channels[f_size] - self.fuse_convs_dict[f_size] = Fuse_sft_block(in_ch, in_ch) - - self.load_state_dict(state_dict) - - def _init_weights(self, module): - if isinstance(module, (nn.Linear, nn.Embedding)): - module.weight.data.normal_(mean=0.0, std=0.02) - if isinstance(module, nn.Linear) and module.bias is not None: - module.bias.data.zero_() - elif isinstance(module, nn.LayerNorm): - module.bias.data.zero_() - module.weight.data.fill_(1.0) - - def forward(self, x, weight=0.5, **kwargs): - detach_16 = True - code_only = False - adain = True - # ################### Encoder ##################### - enc_feat_dict = {} - out_list = [self.fuse_encoder_block[f_size] for f_size in self.connect_list] - for i, block in enumerate(self.encoder.blocks): - x = block(x) - if i in out_list: - enc_feat_dict[str(x.shape[-1])] = x.clone() - - lq_feat = x - # ################# Transformer ################### - # quant_feat, codebook_loss, quant_stats = self.quantize(lq_feat) - pos_emb = self.position_emb.unsqueeze(1).repeat(1, x.shape[0], 1) - # BCHW -> BC(HW) -> (HW)BC - feat_emb = self.feat_emb(lq_feat.flatten(2).permute(2, 0, 1)) - query_emb = feat_emb - # Transformer encoder - for layer in self.ft_layers: - query_emb = layer(query_emb, query_pos=pos_emb) - - # output logits - logits = self.idx_pred_layer(query_emb) # (hw)bn - logits = logits.permute(1, 0, 2) # (hw)bn -> b(hw)n - - if code_only: # for training stage II - # logits doesn't need softmax before cross_entropy loss - return logits, lq_feat - - # ################# Quantization ################### - # if self.training: - # quant_feat = torch.einsum('btn,nc->btc', [soft_one_hot, self.quantize.embedding.weight]) - # # b(hw)c -> bc(hw) -> bchw - # quant_feat = quant_feat.permute(0,2,1).view(lq_feat.shape) - # ------------ - soft_one_hot = F.softmax(logits, dim=2) - _, top_idx = torch.topk(soft_one_hot, 1, dim=2) - quant_feat = self.quantize.get_codebook_feat( - top_idx, shape=[x.shape[0], 16, 16, 256] # type: ignore - ) - # preserve gradients - # quant_feat = lq_feat + (quant_feat - lq_feat).detach() - - if detach_16: - quant_feat = quant_feat.detach() # for training stage III - if adain: - quant_feat = adaptive_instance_normalization(quant_feat, lq_feat) - - # ################## Generator #################### - x = quant_feat - fuse_list = [self.fuse_generator_block[f_size] for f_size in self.connect_list] - - for i, block in enumerate(self.generator.blocks): - x = block(x) - if i in fuse_list: # fuse after i-th block - f_size = str(x.shape[-1]) - if weight > 0: - x = self.fuse_convs_dict[f_size]( - enc_feat_dict[f_size].detach(), x, weight - ) - out = x - # logits doesn't need softmax before cross_entropy loss - # return out, logits, lq_feat - return out, logits diff --git a/comfy_extras/chainner_models/architecture/face/fused_act.py b/comfy_extras/chainner_models/architecture/face/fused_act.py deleted file mode 100644 index 7ed526547..000000000 --- a/comfy_extras/chainner_models/architecture/face/fused_act.py +++ /dev/null @@ -1,81 +0,0 @@ -# pylint: skip-file -# type: ignore -# modify from https://github.com/rosinality/stylegan2-pytorch/blob/master/op/fused_act.py # noqa:E501 - -import torch -from torch import nn -from torch.autograd import Function - -fused_act_ext = None - - -class FusedLeakyReLUFunctionBackward(Function): - @staticmethod - def forward(ctx, grad_output, out, negative_slope, scale): - ctx.save_for_backward(out) - ctx.negative_slope = negative_slope - ctx.scale = scale - - empty = grad_output.new_empty(0) - - grad_input = fused_act_ext.fused_bias_act( - grad_output, empty, out, 3, 1, negative_slope, scale - ) - - dim = [0] - - if grad_input.ndim > 2: - dim += list(range(2, grad_input.ndim)) - - grad_bias = grad_input.sum(dim).detach() - - return grad_input, grad_bias - - @staticmethod - def backward(ctx, gradgrad_input, gradgrad_bias): - (out,) = ctx.saved_tensors - gradgrad_out = fused_act_ext.fused_bias_act( - gradgrad_input, gradgrad_bias, out, 3, 1, ctx.negative_slope, ctx.scale - ) - - return gradgrad_out, None, None, None - - -class FusedLeakyReLUFunction(Function): - @staticmethod - def forward(ctx, input, bias, negative_slope, scale): - empty = input.new_empty(0) - out = fused_act_ext.fused_bias_act( - input, bias, empty, 3, 0, negative_slope, scale - ) - ctx.save_for_backward(out) - ctx.negative_slope = negative_slope - ctx.scale = scale - - return out - - @staticmethod - def backward(ctx, grad_output): - (out,) = ctx.saved_tensors - - grad_input, grad_bias = FusedLeakyReLUFunctionBackward.apply( - grad_output, out, ctx.negative_slope, ctx.scale - ) - - return grad_input, grad_bias, None, None - - -class FusedLeakyReLU(nn.Module): - def __init__(self, channel, negative_slope=0.2, scale=2**0.5): - super().__init__() - - self.bias = nn.Parameter(torch.zeros(channel)) - self.negative_slope = negative_slope - self.scale = scale - - def forward(self, input): - return fused_leaky_relu(input, self.bias, self.negative_slope, self.scale) - - -def fused_leaky_relu(input, bias, negative_slope=0.2, scale=2**0.5): - return FusedLeakyReLUFunction.apply(input, bias, negative_slope, scale) diff --git a/comfy_extras/chainner_models/architecture/face/gfpgan_bilinear_arch.py b/comfy_extras/chainner_models/architecture/face/gfpgan_bilinear_arch.py deleted file mode 100644 index b6e820e00..000000000 --- a/comfy_extras/chainner_models/architecture/face/gfpgan_bilinear_arch.py +++ /dev/null @@ -1,389 +0,0 @@ -# pylint: skip-file -# type: ignore -import math -import random - -import torch -from torch import nn - -from .gfpganv1_arch import ResUpBlock -from .stylegan2_bilinear_arch import ( - ConvLayer, - EqualConv2d, - EqualLinear, - ResBlock, - ScaledLeakyReLU, - StyleGAN2GeneratorBilinear, -) - - -class StyleGAN2GeneratorBilinearSFT(StyleGAN2GeneratorBilinear): - """StyleGAN2 Generator with SFT modulation (Spatial Feature Transform). - It is the bilinear version. It does not use the complicated UpFirDnSmooth function that is not friendly for - deployment. It can be easily converted to the clean version: StyleGAN2GeneratorCSFT. - Args: - out_size (int): The spatial size of outputs. - num_style_feat (int): Channel number of style features. Default: 512. - num_mlp (int): Layer number of MLP style layers. Default: 8. - channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. - lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01. - narrow (float): The narrow ratio for channels. Default: 1. - sft_half (bool): Whether to apply SFT on half of the input channels. Default: False. - """ - - def __init__( - self, - out_size, - num_style_feat=512, - num_mlp=8, - channel_multiplier=2, - lr_mlp=0.01, - narrow=1, - sft_half=False, - ): - super(StyleGAN2GeneratorBilinearSFT, self).__init__( - out_size, - num_style_feat=num_style_feat, - num_mlp=num_mlp, - channel_multiplier=channel_multiplier, - lr_mlp=lr_mlp, - narrow=narrow, - ) - self.sft_half = sft_half - - def forward( - self, - styles, - conditions, - input_is_latent=False, - noise=None, - randomize_noise=True, - truncation=1, - truncation_latent=None, - inject_index=None, - return_latents=False, - ): - """Forward function for StyleGAN2GeneratorBilinearSFT. - Args: - styles (list[Tensor]): Sample codes of styles. - conditions (list[Tensor]): SFT conditions to generators. - input_is_latent (bool): Whether input is latent style. Default: False. - noise (Tensor | None): Input noise or None. Default: None. - randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True. - truncation (float): The truncation ratio. Default: 1. - truncation_latent (Tensor | None): The truncation latent tensor. Default: None. - inject_index (int | None): The injection index for mixing noise. Default: None. - return_latents (bool): Whether to return style latents. Default: False. - """ - # style codes -> latents with Style MLP layer - if not input_is_latent: - styles = [self.style_mlp(s) for s in styles] - # noises - if noise is None: - if randomize_noise: - noise = [None] * self.num_layers # for each style conv layer - else: # use the stored noise - noise = [ - getattr(self.noises, f"noise{i}") for i in range(self.num_layers) - ] - # style truncation - if truncation < 1: - style_truncation = [] - for style in styles: - style_truncation.append( - truncation_latent + truncation * (style - truncation_latent) - ) - styles = style_truncation - # get style latents with injection - if len(styles) == 1: - inject_index = self.num_latent - - if styles[0].ndim < 3: - # repeat latent code for all the layers - latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - else: # used for encoder with different latent code for each layer - latent = styles[0] - elif len(styles) == 2: # mixing noises - if inject_index is None: - inject_index = random.randint(1, self.num_latent - 1) - latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - latent2 = ( - styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1) - ) - latent = torch.cat([latent1, latent2], 1) - - # main generation - out = self.constant_input(latent.shape[0]) - out = self.style_conv1(out, latent[:, 0], noise=noise[0]) - skip = self.to_rgb1(out, latent[:, 1]) - - i = 1 - for conv1, conv2, noise1, noise2, to_rgb in zip( - self.style_convs[::2], - self.style_convs[1::2], - noise[1::2], - noise[2::2], - self.to_rgbs, - ): - out = conv1(out, latent[:, i], noise=noise1) - - # the conditions may have fewer levels - if i < len(conditions): - # SFT part to combine the conditions - if self.sft_half: # only apply SFT to half of the channels - out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1) - out_sft = out_sft * conditions[i - 1] + conditions[i] - out = torch.cat([out_same, out_sft], dim=1) - else: # apply SFT to all the channels - out = out * conditions[i - 1] + conditions[i] - - out = conv2(out, latent[:, i + 1], noise=noise2) - skip = to_rgb(out, latent[:, i + 2], skip) # feature back to the rgb space - i += 2 - - image = skip - - if return_latents: - return image, latent - else: - return image, None - - -class GFPGANBilinear(nn.Module): - """The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT. - It is the bilinear version and it does not use the complicated UpFirDnSmooth function that is not friendly for - deployment. It can be easily converted to the clean version: GFPGANv1Clean. - Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior. - Args: - out_size (int): The spatial size of outputs. - num_style_feat (int): Channel number of style features. Default: 512. - channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. - decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None. - fix_decoder (bool): Whether to fix the decoder. Default: True. - num_mlp (int): Layer number of MLP style layers. Default: 8. - lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01. - input_is_latent (bool): Whether input is latent style. Default: False. - different_w (bool): Whether to use different latent w for different layers. Default: False. - narrow (float): The narrow ratio for channels. Default: 1. - sft_half (bool): Whether to apply SFT on half of the input channels. Default: False. - """ - - def __init__( - self, - out_size, - num_style_feat=512, - channel_multiplier=1, - decoder_load_path=None, - fix_decoder=True, - # for stylegan decoder - num_mlp=8, - lr_mlp=0.01, - input_is_latent=False, - different_w=False, - narrow=1, - sft_half=False, - ): - super(GFPGANBilinear, self).__init__() - self.input_is_latent = input_is_latent - self.different_w = different_w - self.num_style_feat = num_style_feat - self.min_size_restriction = 512 - - unet_narrow = narrow * 0.5 # by default, use a half of input channels - channels = { - "4": int(512 * unet_narrow), - "8": int(512 * unet_narrow), - "16": int(512 * unet_narrow), - "32": int(512 * unet_narrow), - "64": int(256 * channel_multiplier * unet_narrow), - "128": int(128 * channel_multiplier * unet_narrow), - "256": int(64 * channel_multiplier * unet_narrow), - "512": int(32 * channel_multiplier * unet_narrow), - "1024": int(16 * channel_multiplier * unet_narrow), - } - - self.log_size = int(math.log(out_size, 2)) - first_out_size = 2 ** (int(math.log(out_size, 2))) - - self.conv_body_first = ConvLayer( - 3, channels[f"{first_out_size}"], 1, bias=True, activate=True - ) - - # downsample - in_channels = channels[f"{first_out_size}"] - self.conv_body_down = nn.ModuleList() - for i in range(self.log_size, 2, -1): - out_channels = channels[f"{2**(i - 1)}"] - self.conv_body_down.append(ResBlock(in_channels, out_channels)) - in_channels = out_channels - - self.final_conv = ConvLayer( - in_channels, channels["4"], 3, bias=True, activate=True - ) - - # upsample - in_channels = channels["4"] - self.conv_body_up = nn.ModuleList() - for i in range(3, self.log_size + 1): - out_channels = channels[f"{2**i}"] - self.conv_body_up.append(ResUpBlock(in_channels, out_channels)) - in_channels = out_channels - - # to RGB - self.toRGB = nn.ModuleList() - for i in range(3, self.log_size + 1): - self.toRGB.append( - EqualConv2d( - channels[f"{2**i}"], - 3, - 1, - stride=1, - padding=0, - bias=True, - bias_init_val=0, - ) - ) - - if different_w: - linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat - else: - linear_out_channel = num_style_feat - - self.final_linear = EqualLinear( - channels["4"] * 4 * 4, - linear_out_channel, - bias=True, - bias_init_val=0, - lr_mul=1, - activation=None, - ) - - # the decoder: stylegan2 generator with SFT modulations - self.stylegan_decoder = StyleGAN2GeneratorBilinearSFT( - out_size=out_size, - num_style_feat=num_style_feat, - num_mlp=num_mlp, - channel_multiplier=channel_multiplier, - lr_mlp=lr_mlp, - narrow=narrow, - sft_half=sft_half, - ) - - # load pre-trained stylegan2 model if necessary - if decoder_load_path: - self.stylegan_decoder.load_state_dict( - torch.load( - decoder_load_path, map_location=lambda storage, loc: storage - )["params_ema"] - ) - # fix decoder without updating params - if fix_decoder: - for _, param in self.stylegan_decoder.named_parameters(): - param.requires_grad = False - - # for SFT modulations (scale and shift) - self.condition_scale = nn.ModuleList() - self.condition_shift = nn.ModuleList() - for i in range(3, self.log_size + 1): - out_channels = channels[f"{2**i}"] - if sft_half: - sft_out_channels = out_channels - else: - sft_out_channels = out_channels * 2 - self.condition_scale.append( - nn.Sequential( - EqualConv2d( - out_channels, - out_channels, - 3, - stride=1, - padding=1, - bias=True, - bias_init_val=0, - ), - ScaledLeakyReLU(0.2), - EqualConv2d( - out_channels, - sft_out_channels, - 3, - stride=1, - padding=1, - bias=True, - bias_init_val=1, - ), - ) - ) - self.condition_shift.append( - nn.Sequential( - EqualConv2d( - out_channels, - out_channels, - 3, - stride=1, - padding=1, - bias=True, - bias_init_val=0, - ), - ScaledLeakyReLU(0.2), - EqualConv2d( - out_channels, - sft_out_channels, - 3, - stride=1, - padding=1, - bias=True, - bias_init_val=0, - ), - ) - ) - - def forward(self, x, return_latents=False, return_rgb=True, randomize_noise=True): - """Forward function for GFPGANBilinear. - Args: - x (Tensor): Input images. - return_latents (bool): Whether to return style latents. Default: False. - return_rgb (bool): Whether return intermediate rgb images. Default: True. - randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True. - """ - conditions = [] - unet_skips = [] - out_rgbs = [] - - # encoder - feat = self.conv_body_first(x) - for i in range(self.log_size - 2): - feat = self.conv_body_down[i](feat) - unet_skips.insert(0, feat) - - feat = self.final_conv(feat) - - # style code - style_code = self.final_linear(feat.view(feat.size(0), -1)) - if self.different_w: - style_code = style_code.view(style_code.size(0), -1, self.num_style_feat) - - # decode - for i in range(self.log_size - 2): - # add unet skip - feat = feat + unet_skips[i] - # ResUpLayer - feat = self.conv_body_up[i](feat) - # generate scale and shift for SFT layers - scale = self.condition_scale[i](feat) - conditions.append(scale.clone()) - shift = self.condition_shift[i](feat) - conditions.append(shift.clone()) - # generate rgb images - if return_rgb: - out_rgbs.append(self.toRGB[i](feat)) - - # decoder - image, _ = self.stylegan_decoder( - [style_code], - conditions, - return_latents=return_latents, - input_is_latent=self.input_is_latent, - randomize_noise=randomize_noise, - ) - - return image, out_rgbs diff --git a/comfy_extras/chainner_models/architecture/face/gfpganv1_arch.py b/comfy_extras/chainner_models/architecture/face/gfpganv1_arch.py deleted file mode 100644 index 72d72fc86..000000000 --- a/comfy_extras/chainner_models/architecture/face/gfpganv1_arch.py +++ /dev/null @@ -1,566 +0,0 @@ -# pylint: skip-file -# type: ignore -import math -import random - -import torch -from torch import nn -from torch.nn import functional as F - -from .fused_act import FusedLeakyReLU -from .stylegan2_arch import ( - ConvLayer, - EqualConv2d, - EqualLinear, - ResBlock, - ScaledLeakyReLU, - StyleGAN2Generator, -) - - -class StyleGAN2GeneratorSFT(StyleGAN2Generator): - """StyleGAN2 Generator with SFT modulation (Spatial Feature Transform). - Args: - out_size (int): The spatial size of outputs. - num_style_feat (int): Channel number of style features. Default: 512. - num_mlp (int): Layer number of MLP style layers. Default: 8. - channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. - resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be - applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1). - lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01. - narrow (float): The narrow ratio for channels. Default: 1. - sft_half (bool): Whether to apply SFT on half of the input channels. Default: False. - """ - - def __init__( - self, - out_size, - num_style_feat=512, - num_mlp=8, - channel_multiplier=2, - resample_kernel=(1, 3, 3, 1), - lr_mlp=0.01, - narrow=1, - sft_half=False, - ): - super(StyleGAN2GeneratorSFT, self).__init__( - out_size, - num_style_feat=num_style_feat, - num_mlp=num_mlp, - channel_multiplier=channel_multiplier, - resample_kernel=resample_kernel, - lr_mlp=lr_mlp, - narrow=narrow, - ) - self.sft_half = sft_half - - def forward( - self, - styles, - conditions, - input_is_latent=False, - noise=None, - randomize_noise=True, - truncation=1, - truncation_latent=None, - inject_index=None, - return_latents=False, - ): - """Forward function for StyleGAN2GeneratorSFT. - Args: - styles (list[Tensor]): Sample codes of styles. - conditions (list[Tensor]): SFT conditions to generators. - input_is_latent (bool): Whether input is latent style. Default: False. - noise (Tensor | None): Input noise or None. Default: None. - randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True. - truncation (float): The truncation ratio. Default: 1. - truncation_latent (Tensor | None): The truncation latent tensor. Default: None. - inject_index (int | None): The injection index for mixing noise. Default: None. - return_latents (bool): Whether to return style latents. Default: False. - """ - # style codes -> latents with Style MLP layer - if not input_is_latent: - styles = [self.style_mlp(s) for s in styles] - # noises - if noise is None: - if randomize_noise: - noise = [None] * self.num_layers # for each style conv layer - else: # use the stored noise - noise = [ - getattr(self.noises, f"noise{i}") for i in range(self.num_layers) - ] - # style truncation - if truncation < 1: - style_truncation = [] - for style in styles: - style_truncation.append( - truncation_latent + truncation * (style - truncation_latent) - ) - styles = style_truncation - # get style latents with injection - if len(styles) == 1: - inject_index = self.num_latent - - if styles[0].ndim < 3: - # repeat latent code for all the layers - latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - else: # used for encoder with different latent code for each layer - latent = styles[0] - elif len(styles) == 2: # mixing noises - if inject_index is None: - inject_index = random.randint(1, self.num_latent - 1) - latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - latent2 = ( - styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1) - ) - latent = torch.cat([latent1, latent2], 1) - - # main generation - out = self.constant_input(latent.shape[0]) - out = self.style_conv1(out, latent[:, 0], noise=noise[0]) - skip = self.to_rgb1(out, latent[:, 1]) - - i = 1 - for conv1, conv2, noise1, noise2, to_rgb in zip( - self.style_convs[::2], - self.style_convs[1::2], - noise[1::2], - noise[2::2], - self.to_rgbs, - ): - out = conv1(out, latent[:, i], noise=noise1) - - # the conditions may have fewer levels - if i < len(conditions): - # SFT part to combine the conditions - if self.sft_half: # only apply SFT to half of the channels - out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1) - out_sft = out_sft * conditions[i - 1] + conditions[i] - out = torch.cat([out_same, out_sft], dim=1) - else: # apply SFT to all the channels - out = out * conditions[i - 1] + conditions[i] - - out = conv2(out, latent[:, i + 1], noise=noise2) - skip = to_rgb(out, latent[:, i + 2], skip) # feature back to the rgb space - i += 2 - - image = skip - - if return_latents: - return image, latent - else: - return image, None - - -class ConvUpLayer(nn.Module): - """Convolutional upsampling layer. It uses bilinear upsampler + Conv. - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Size of the convolving kernel. - stride (int): Stride of the convolution. Default: 1 - padding (int): Zero-padding added to both sides of the input. Default: 0. - bias (bool): If ``True``, adds a learnable bias to the output. Default: ``True``. - bias_init_val (float): Bias initialized value. Default: 0. - activate (bool): Whether use activateion. Default: True. - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - stride=1, - padding=0, - bias=True, - bias_init_val=0, - activate=True, - ): - super(ConvUpLayer, self).__init__() - self.in_channels = in_channels - self.out_channels = out_channels - self.kernel_size = kernel_size - self.stride = stride - self.padding = padding - # self.scale is used to scale the convolution weights, which is related to the common initializations. - self.scale = 1 / math.sqrt(in_channels * kernel_size**2) - - self.weight = nn.Parameter( - torch.randn(out_channels, in_channels, kernel_size, kernel_size) - ) - - if bias and not activate: - self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val)) - else: - self.register_parameter("bias", None) - - # activation - if activate: - if bias: - self.activation = FusedLeakyReLU(out_channels) - else: - self.activation = ScaledLeakyReLU(0.2) - else: - self.activation = None - - def forward(self, x): - # bilinear upsample - out = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=False) - # conv - out = F.conv2d( - out, - self.weight * self.scale, - bias=self.bias, - stride=self.stride, - padding=self.padding, - ) - # activation - if self.activation is not None: - out = self.activation(out) - return out - - -class ResUpBlock(nn.Module): - """Residual block with upsampling. - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - """ - - def __init__(self, in_channels, out_channels): - super(ResUpBlock, self).__init__() - - self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True) - self.conv2 = ConvUpLayer( - in_channels, out_channels, 3, stride=1, padding=1, bias=True, activate=True - ) - self.skip = ConvUpLayer( - in_channels, out_channels, 1, bias=False, activate=False - ) - - def forward(self, x): - out = self.conv1(x) - out = self.conv2(out) - skip = self.skip(x) - out = (out + skip) / math.sqrt(2) - return out - - -class GFPGANv1(nn.Module): - """The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT. - Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior. - Args: - out_size (int): The spatial size of outputs. - num_style_feat (int): Channel number of style features. Default: 512. - channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. - resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be - applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1). - decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None. - fix_decoder (bool): Whether to fix the decoder. Default: True. - num_mlp (int): Layer number of MLP style layers. Default: 8. - lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01. - input_is_latent (bool): Whether input is latent style. Default: False. - different_w (bool): Whether to use different latent w for different layers. Default: False. - narrow (float): The narrow ratio for channels. Default: 1. - sft_half (bool): Whether to apply SFT on half of the input channels. Default: False. - """ - - def __init__( - self, - out_size, - num_style_feat=512, - channel_multiplier=1, - resample_kernel=(1, 3, 3, 1), - decoder_load_path=None, - fix_decoder=True, - # for stylegan decoder - num_mlp=8, - lr_mlp=0.01, - input_is_latent=False, - different_w=False, - narrow=1, - sft_half=False, - ): - super(GFPGANv1, self).__init__() - self.input_is_latent = input_is_latent - self.different_w = different_w - self.num_style_feat = num_style_feat - - unet_narrow = narrow * 0.5 # by default, use a half of input channels - channels = { - "4": int(512 * unet_narrow), - "8": int(512 * unet_narrow), - "16": int(512 * unet_narrow), - "32": int(512 * unet_narrow), - "64": int(256 * channel_multiplier * unet_narrow), - "128": int(128 * channel_multiplier * unet_narrow), - "256": int(64 * channel_multiplier * unet_narrow), - "512": int(32 * channel_multiplier * unet_narrow), - "1024": int(16 * channel_multiplier * unet_narrow), - } - - self.log_size = int(math.log(out_size, 2)) - first_out_size = 2 ** (int(math.log(out_size, 2))) - - self.conv_body_first = ConvLayer( - 3, channels[f"{first_out_size}"], 1, bias=True, activate=True - ) - - # downsample - in_channels = channels[f"{first_out_size}"] - self.conv_body_down = nn.ModuleList() - for i in range(self.log_size, 2, -1): - out_channels = channels[f"{2**(i - 1)}"] - self.conv_body_down.append( - ResBlock(in_channels, out_channels, resample_kernel) - ) - in_channels = out_channels - - self.final_conv = ConvLayer( - in_channels, channels["4"], 3, bias=True, activate=True - ) - - # upsample - in_channels = channels["4"] - self.conv_body_up = nn.ModuleList() - for i in range(3, self.log_size + 1): - out_channels = channels[f"{2**i}"] - self.conv_body_up.append(ResUpBlock(in_channels, out_channels)) - in_channels = out_channels - - # to RGB - self.toRGB = nn.ModuleList() - for i in range(3, self.log_size + 1): - self.toRGB.append( - EqualConv2d( - channels[f"{2**i}"], - 3, - 1, - stride=1, - padding=0, - bias=True, - bias_init_val=0, - ) - ) - - if different_w: - linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat - else: - linear_out_channel = num_style_feat - - self.final_linear = EqualLinear( - channels["4"] * 4 * 4, - linear_out_channel, - bias=True, - bias_init_val=0, - lr_mul=1, - activation=None, - ) - - # the decoder: stylegan2 generator with SFT modulations - self.stylegan_decoder = StyleGAN2GeneratorSFT( - out_size=out_size, - num_style_feat=num_style_feat, - num_mlp=num_mlp, - channel_multiplier=channel_multiplier, - resample_kernel=resample_kernel, - lr_mlp=lr_mlp, - narrow=narrow, - sft_half=sft_half, - ) - - # load pre-trained stylegan2 model if necessary - if decoder_load_path: - self.stylegan_decoder.load_state_dict( - torch.load( - decoder_load_path, map_location=lambda storage, loc: storage - )["params_ema"] - ) - # fix decoder without updating params - if fix_decoder: - for _, param in self.stylegan_decoder.named_parameters(): - param.requires_grad = False - - # for SFT modulations (scale and shift) - self.condition_scale = nn.ModuleList() - self.condition_shift = nn.ModuleList() - for i in range(3, self.log_size + 1): - out_channels = channels[f"{2**i}"] - if sft_half: - sft_out_channels = out_channels - else: - sft_out_channels = out_channels * 2 - self.condition_scale.append( - nn.Sequential( - EqualConv2d( - out_channels, - out_channels, - 3, - stride=1, - padding=1, - bias=True, - bias_init_val=0, - ), - ScaledLeakyReLU(0.2), - EqualConv2d( - out_channels, - sft_out_channels, - 3, - stride=1, - padding=1, - bias=True, - bias_init_val=1, - ), - ) - ) - self.condition_shift.append( - nn.Sequential( - EqualConv2d( - out_channels, - out_channels, - 3, - stride=1, - padding=1, - bias=True, - bias_init_val=0, - ), - ScaledLeakyReLU(0.2), - EqualConv2d( - out_channels, - sft_out_channels, - 3, - stride=1, - padding=1, - bias=True, - bias_init_val=0, - ), - ) - ) - - def forward( - self, x, return_latents=False, return_rgb=True, randomize_noise=True, **kwargs - ): - """Forward function for GFPGANv1. - Args: - x (Tensor): Input images. - return_latents (bool): Whether to return style latents. Default: False. - return_rgb (bool): Whether return intermediate rgb images. Default: True. - randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True. - """ - conditions = [] - unet_skips = [] - out_rgbs = [] - - # encoder - feat = self.conv_body_first(x) - for i in range(self.log_size - 2): - feat = self.conv_body_down[i](feat) - unet_skips.insert(0, feat) - - feat = self.final_conv(feat) - - # style code - style_code = self.final_linear(feat.view(feat.size(0), -1)) - if self.different_w: - style_code = style_code.view(style_code.size(0), -1, self.num_style_feat) - - # decode - for i in range(self.log_size - 2): - # add unet skip - feat = feat + unet_skips[i] - # ResUpLayer - feat = self.conv_body_up[i](feat) - # generate scale and shift for SFT layers - scale = self.condition_scale[i](feat) - conditions.append(scale.clone()) - shift = self.condition_shift[i](feat) - conditions.append(shift.clone()) - # generate rgb images - if return_rgb: - out_rgbs.append(self.toRGB[i](feat)) - - # decoder - image, _ = self.stylegan_decoder( - [style_code], - conditions, - return_latents=return_latents, - input_is_latent=self.input_is_latent, - randomize_noise=randomize_noise, - ) - - return image, out_rgbs - - -class FacialComponentDiscriminator(nn.Module): - """Facial component (eyes, mouth, noise) discriminator used in GFPGAN.""" - - def __init__(self): - super(FacialComponentDiscriminator, self).__init__() - # It now uses a VGG-style architectrue with fixed model size - self.conv1 = ConvLayer( - 3, - 64, - 3, - downsample=False, - resample_kernel=(1, 3, 3, 1), - bias=True, - activate=True, - ) - self.conv2 = ConvLayer( - 64, - 128, - 3, - downsample=True, - resample_kernel=(1, 3, 3, 1), - bias=True, - activate=True, - ) - self.conv3 = ConvLayer( - 128, - 128, - 3, - downsample=False, - resample_kernel=(1, 3, 3, 1), - bias=True, - activate=True, - ) - self.conv4 = ConvLayer( - 128, - 256, - 3, - downsample=True, - resample_kernel=(1, 3, 3, 1), - bias=True, - activate=True, - ) - self.conv5 = ConvLayer( - 256, - 256, - 3, - downsample=False, - resample_kernel=(1, 3, 3, 1), - bias=True, - activate=True, - ) - self.final_conv = ConvLayer(256, 1, 3, bias=True, activate=False) - - def forward(self, x, return_feats=False, **kwargs): - """Forward function for FacialComponentDiscriminator. - Args: - x (Tensor): Input images. - return_feats (bool): Whether to return intermediate features. Default: False. - """ - feat = self.conv1(x) - feat = self.conv3(self.conv2(feat)) - rlt_feats = [] - if return_feats: - rlt_feats.append(feat.clone()) - feat = self.conv5(self.conv4(feat)) - if return_feats: - rlt_feats.append(feat.clone()) - out = self.final_conv(feat) - - if return_feats: - return out, rlt_feats - else: - return out, None diff --git a/comfy_extras/chainner_models/architecture/face/gfpganv1_clean_arch.py b/comfy_extras/chainner_models/architecture/face/gfpganv1_clean_arch.py deleted file mode 100644 index 16470d634..000000000 --- a/comfy_extras/chainner_models/architecture/face/gfpganv1_clean_arch.py +++ /dev/null @@ -1,370 +0,0 @@ -# pylint: skip-file -# type: ignore -import math -import random - -import torch -from torch import nn -from torch.nn import functional as F - -from .stylegan2_clean_arch import StyleGAN2GeneratorClean - - -class StyleGAN2GeneratorCSFT(StyleGAN2GeneratorClean): - """StyleGAN2 Generator with SFT modulation (Spatial Feature Transform). - It is the clean version without custom compiled CUDA extensions used in StyleGAN2. - Args: - out_size (int): The spatial size of outputs. - num_style_feat (int): Channel number of style features. Default: 512. - num_mlp (int): Layer number of MLP style layers. Default: 8. - channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. - narrow (float): The narrow ratio for channels. Default: 1. - sft_half (bool): Whether to apply SFT on half of the input channels. Default: False. - """ - - def __init__( - self, - out_size, - num_style_feat=512, - num_mlp=8, - channel_multiplier=2, - narrow=1, - sft_half=False, - ): - super(StyleGAN2GeneratorCSFT, self).__init__( - out_size, - num_style_feat=num_style_feat, - num_mlp=num_mlp, - channel_multiplier=channel_multiplier, - narrow=narrow, - ) - self.sft_half = sft_half - - def forward( - self, - styles, - conditions, - input_is_latent=False, - noise=None, - randomize_noise=True, - truncation=1, - truncation_latent=None, - inject_index=None, - return_latents=False, - ): - """Forward function for StyleGAN2GeneratorCSFT. - Args: - styles (list[Tensor]): Sample codes of styles. - conditions (list[Tensor]): SFT conditions to generators. - input_is_latent (bool): Whether input is latent style. Default: False. - noise (Tensor | None): Input noise or None. Default: None. - randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True. - truncation (float): The truncation ratio. Default: 1. - truncation_latent (Tensor | None): The truncation latent tensor. Default: None. - inject_index (int | None): The injection index for mixing noise. Default: None. - return_latents (bool): Whether to return style latents. Default: False. - """ - # style codes -> latents with Style MLP layer - if not input_is_latent: - styles = [self.style_mlp(s) for s in styles] - # noises - if noise is None: - if randomize_noise: - noise = [None] * self.num_layers # for each style conv layer - else: # use the stored noise - noise = [ - getattr(self.noises, f"noise{i}") for i in range(self.num_layers) - ] - # style truncation - if truncation < 1: - style_truncation = [] - for style in styles: - style_truncation.append( - truncation_latent + truncation * (style - truncation_latent) - ) - styles = style_truncation - # get style latents with injection - if len(styles) == 1: - inject_index = self.num_latent - - if styles[0].ndim < 3: - # repeat latent code for all the layers - latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - else: # used for encoder with different latent code for each layer - latent = styles[0] - elif len(styles) == 2: # mixing noises - if inject_index is None: - inject_index = random.randint(1, self.num_latent - 1) - latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - latent2 = ( - styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1) - ) - latent = torch.cat([latent1, latent2], 1) - - # main generation - out = self.constant_input(latent.shape[0]) - out = self.style_conv1(out, latent[:, 0], noise=noise[0]) - skip = self.to_rgb1(out, latent[:, 1]) - - i = 1 - for conv1, conv2, noise1, noise2, to_rgb in zip( - self.style_convs[::2], - self.style_convs[1::2], - noise[1::2], - noise[2::2], - self.to_rgbs, - ): - out = conv1(out, latent[:, i], noise=noise1) - - # the conditions may have fewer levels - if i < len(conditions): - # SFT part to combine the conditions - if self.sft_half: # only apply SFT to half of the channels - out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1) - out_sft = out_sft * conditions[i - 1] + conditions[i] - out = torch.cat([out_same, out_sft], dim=1) - else: # apply SFT to all the channels - out = out * conditions[i - 1] + conditions[i] - - out = conv2(out, latent[:, i + 1], noise=noise2) - skip = to_rgb(out, latent[:, i + 2], skip) # feature back to the rgb space - i += 2 - - image = skip - - if return_latents: - return image, latent - else: - return image, None - - -class ResBlock(nn.Module): - """Residual block with bilinear upsampling/downsampling. - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - mode (str): Upsampling/downsampling mode. Options: down | up. Default: down. - """ - - def __init__(self, in_channels, out_channels, mode="down"): - super(ResBlock, self).__init__() - - self.conv1 = nn.Conv2d(in_channels, in_channels, 3, 1, 1) - self.conv2 = nn.Conv2d(in_channels, out_channels, 3, 1, 1) - self.skip = nn.Conv2d(in_channels, out_channels, 1, bias=False) - if mode == "down": - self.scale_factor = 0.5 - elif mode == "up": - self.scale_factor = 2 - - def forward(self, x): - out = F.leaky_relu_(self.conv1(x), negative_slope=0.2) - # upsample/downsample - out = F.interpolate( - out, scale_factor=self.scale_factor, mode="bilinear", align_corners=False - ) - out = F.leaky_relu_(self.conv2(out), negative_slope=0.2) - # skip - x = F.interpolate( - x, scale_factor=self.scale_factor, mode="bilinear", align_corners=False - ) - skip = self.skip(x) - out = out + skip - return out - - -class GFPGANv1Clean(nn.Module): - """The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT. - It is the clean version without custom compiled CUDA extensions used in StyleGAN2. - Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior. - Args: - out_size (int): The spatial size of outputs. - num_style_feat (int): Channel number of style features. Default: 512. - channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. - decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None. - fix_decoder (bool): Whether to fix the decoder. Default: True. - num_mlp (int): Layer number of MLP style layers. Default: 8. - input_is_latent (bool): Whether input is latent style. Default: False. - different_w (bool): Whether to use different latent w for different layers. Default: False. - narrow (float): The narrow ratio for channels. Default: 1. - sft_half (bool): Whether to apply SFT on half of the input channels. Default: False. - """ - - def __init__( - self, - state_dict, - ): - super(GFPGANv1Clean, self).__init__() - - out_size = 512 - num_style_feat = 512 - channel_multiplier = 2 - decoder_load_path = None - fix_decoder = False - num_mlp = 8 - input_is_latent = True - different_w = True - narrow = 1 - sft_half = True - - self.model_arch = "GFPGAN" - self.sub_type = "Face SR" - self.scale = 8 - self.in_nc = 3 - self.out_nc = 3 - self.state = state_dict - - self.supports_fp16 = False - self.supports_bf16 = True - self.min_size_restriction = 512 - - self.input_is_latent = input_is_latent - self.different_w = different_w - self.num_style_feat = num_style_feat - - unet_narrow = narrow * 0.5 # by default, use a half of input channels - channels = { - "4": int(512 * unet_narrow), - "8": int(512 * unet_narrow), - "16": int(512 * unet_narrow), - "32": int(512 * unet_narrow), - "64": int(256 * channel_multiplier * unet_narrow), - "128": int(128 * channel_multiplier * unet_narrow), - "256": int(64 * channel_multiplier * unet_narrow), - "512": int(32 * channel_multiplier * unet_narrow), - "1024": int(16 * channel_multiplier * unet_narrow), - } - - self.log_size = int(math.log(out_size, 2)) - first_out_size = 2 ** (int(math.log(out_size, 2))) - - self.conv_body_first = nn.Conv2d(3, channels[f"{first_out_size}"], 1) - - # downsample - in_channels = channels[f"{first_out_size}"] - self.conv_body_down = nn.ModuleList() - for i in range(self.log_size, 2, -1): - out_channels = channels[f"{2**(i - 1)}"] - self.conv_body_down.append(ResBlock(in_channels, out_channels, mode="down")) - in_channels = out_channels - - self.final_conv = nn.Conv2d(in_channels, channels["4"], 3, 1, 1) - - # upsample - in_channels = channels["4"] - self.conv_body_up = nn.ModuleList() - for i in range(3, self.log_size + 1): - out_channels = channels[f"{2**i}"] - self.conv_body_up.append(ResBlock(in_channels, out_channels, mode="up")) - in_channels = out_channels - - # to RGB - self.toRGB = nn.ModuleList() - for i in range(3, self.log_size + 1): - self.toRGB.append(nn.Conv2d(channels[f"{2**i}"], 3, 1)) - - if different_w: - linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat - else: - linear_out_channel = num_style_feat - - self.final_linear = nn.Linear(channels["4"] * 4 * 4, linear_out_channel) - - # the decoder: stylegan2 generator with SFT modulations - self.stylegan_decoder = StyleGAN2GeneratorCSFT( - out_size=out_size, - num_style_feat=num_style_feat, - num_mlp=num_mlp, - channel_multiplier=channel_multiplier, - narrow=narrow, - sft_half=sft_half, - ) - - # load pre-trained stylegan2 model if necessary - if decoder_load_path: - self.stylegan_decoder.load_state_dict( - torch.load( - decoder_load_path, map_location=lambda storage, loc: storage - )["params_ema"] - ) - # fix decoder without updating params - if fix_decoder: - for _, param in self.stylegan_decoder.named_parameters(): - param.requires_grad = False - - # for SFT modulations (scale and shift) - self.condition_scale = nn.ModuleList() - self.condition_shift = nn.ModuleList() - for i in range(3, self.log_size + 1): - out_channels = channels[f"{2**i}"] - if sft_half: - sft_out_channels = out_channels - else: - sft_out_channels = out_channels * 2 - self.condition_scale.append( - nn.Sequential( - nn.Conv2d(out_channels, out_channels, 3, 1, 1), - nn.LeakyReLU(0.2, True), - nn.Conv2d(out_channels, sft_out_channels, 3, 1, 1), - ) - ) - self.condition_shift.append( - nn.Sequential( - nn.Conv2d(out_channels, out_channels, 3, 1, 1), - nn.LeakyReLU(0.2, True), - nn.Conv2d(out_channels, sft_out_channels, 3, 1, 1), - ) - ) - self.load_state_dict(state_dict) - - def forward( - self, x, return_latents=False, return_rgb=True, randomize_noise=True, **kwargs - ): - """Forward function for GFPGANv1Clean. - Args: - x (Tensor): Input images. - return_latents (bool): Whether to return style latents. Default: False. - return_rgb (bool): Whether return intermediate rgb images. Default: True. - randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True. - """ - conditions = [] - unet_skips = [] - out_rgbs = [] - - # encoder - feat = F.leaky_relu_(self.conv_body_first(x), negative_slope=0.2) - for i in range(self.log_size - 2): - feat = self.conv_body_down[i](feat) - unet_skips.insert(0, feat) - feat = F.leaky_relu_(self.final_conv(feat), negative_slope=0.2) - - # style code - style_code = self.final_linear(feat.view(feat.size(0), -1)) - if self.different_w: - style_code = style_code.view(style_code.size(0), -1, self.num_style_feat) - - # decode - for i in range(self.log_size - 2): - # add unet skip - feat = feat + unet_skips[i] - # ResUpLayer - feat = self.conv_body_up[i](feat) - # generate scale and shift for SFT layers - scale = self.condition_scale[i](feat) - conditions.append(scale.clone()) - shift = self.condition_shift[i](feat) - conditions.append(shift.clone()) - # generate rgb images - if return_rgb: - out_rgbs.append(self.toRGB[i](feat)) - - # decoder - image, _ = self.stylegan_decoder( - [style_code], - conditions, - return_latents=return_latents, - input_is_latent=self.input_is_latent, - randomize_noise=randomize_noise, - ) - - return image, out_rgbs diff --git a/comfy_extras/chainner_models/architecture/face/restoreformer_arch.py b/comfy_extras/chainner_models/architecture/face/restoreformer_arch.py deleted file mode 100644 index 449226029..000000000 --- a/comfy_extras/chainner_models/architecture/face/restoreformer_arch.py +++ /dev/null @@ -1,776 +0,0 @@ -# pylint: skip-file -# type: ignore -"""Modified from https://github.com/wzhouxiff/RestoreFormer -""" -import numpy as np -import torch -import torch.nn as nn -import torch.nn.functional as F - - -class VectorQuantizer(nn.Module): - """ - see https://github.com/MishaLaskin/vqvae/blob/d761a999e2267766400dc646d82d3ac3657771d4/models/quantizer.py - ____________________________________________ - Discretization bottleneck part of the VQ-VAE. - Inputs: - - n_e : number of embeddings - - e_dim : dimension of embedding - - beta : commitment cost used in loss term, beta * ||z_e(x)-sg[e]||^2 - _____________________________________________ - """ - - def __init__(self, n_e, e_dim, beta): - super(VectorQuantizer, self).__init__() - self.n_e = n_e - self.e_dim = e_dim - self.beta = beta - - self.embedding = nn.Embedding(self.n_e, self.e_dim) - self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e) - - def forward(self, z): - """ - Inputs the output of the encoder network z and maps it to a discrete - one-hot vector that is the index of the closest embedding vector e_j - z (continuous) -> z_q (discrete) - z.shape = (batch, channel, height, width) - quantization pipeline: - 1. get encoder input (B,C,H,W) - 2. flatten input to (B*H*W,C) - """ - # reshape z -> (batch, height, width, channel) and flatten - z = z.permute(0, 2, 3, 1).contiguous() - z_flattened = z.view(-1, self.e_dim) - # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z - - d = ( - torch.sum(z_flattened**2, dim=1, keepdim=True) - + torch.sum(self.embedding.weight**2, dim=1) - - 2 * torch.matmul(z_flattened, self.embedding.weight.t()) - ) - - # could possible replace this here - # #\start... - # find closest encodings - - min_value, min_encoding_indices = torch.min(d, dim=1) - - min_encoding_indices = min_encoding_indices.unsqueeze(1) - - min_encodings = torch.zeros(min_encoding_indices.shape[0], self.n_e).to(z) - min_encodings.scatter_(1, min_encoding_indices, 1) - - # dtype min encodings: torch.float32 - # min_encodings shape: torch.Size([2048, 512]) - # min_encoding_indices.shape: torch.Size([2048, 1]) - - # get quantized latent vectors - z_q = torch.matmul(min_encodings, self.embedding.weight).view(z.shape) - # .........\end - - # with: - # .........\start - # min_encoding_indices = torch.argmin(d, dim=1) - # z_q = self.embedding(min_encoding_indices) - # ......\end......... (TODO) - - # compute loss for embedding - loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * torch.mean( - (z_q - z.detach()) ** 2 - ) - - # preserve gradients - z_q = z + (z_q - z).detach() - - # perplexity - - e_mean = torch.mean(min_encodings, dim=0) - perplexity = torch.exp(-torch.sum(e_mean * torch.log(e_mean + 1e-10))) - - # reshape back to match original input shape - z_q = z_q.permute(0, 3, 1, 2).contiguous() - - return z_q, loss, (perplexity, min_encodings, min_encoding_indices, d) - - def get_codebook_entry(self, indices, shape): - # shape specifying (batch, height, width, channel) - # TODO: check for more easy handling with nn.Embedding - min_encodings = torch.zeros(indices.shape[0], self.n_e).to(indices) - min_encodings.scatter_(1, indices[:, None], 1) - - # get quantized latent vectors - z_q = torch.matmul(min_encodings.float(), self.embedding.weight) - - if shape is not None: - z_q = z_q.view(shape) - - # reshape back to match original input shape - z_q = z_q.permute(0, 3, 1, 2).contiguous() - - return z_q - - -# pytorch_diffusion + derived encoder decoder -def nonlinearity(x): - # swish - return x * torch.sigmoid(x) - - -def Normalize(in_channels): - return torch.nn.GroupNorm( - num_groups=32, num_channels=in_channels, eps=1e-6, affine=True - ) - - -class Upsample(nn.Module): - def __init__(self, in_channels, with_conv): - super().__init__() - self.with_conv = with_conv - if self.with_conv: - self.conv = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=3, stride=1, padding=1 - ) - - def forward(self, x): - x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") - if self.with_conv: - x = self.conv(x) - return x - - -class Downsample(nn.Module): - def __init__(self, in_channels, with_conv): - super().__init__() - self.with_conv = with_conv - if self.with_conv: - # no asymmetric padding in torch conv, must do it ourselves - self.conv = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=3, stride=2, padding=0 - ) - - def forward(self, x): - if self.with_conv: - pad = (0, 1, 0, 1) - x = torch.nn.functional.pad(x, pad, mode="constant", value=0) - x = self.conv(x) - else: - x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) - return x - - -class ResnetBlock(nn.Module): - def __init__( - self, - *, - in_channels, - out_channels=None, - conv_shortcut=False, - dropout, - temb_channels=512 - ): - super().__init__() - self.in_channels = in_channels - out_channels = in_channels if out_channels is None else out_channels - self.out_channels = out_channels - self.use_conv_shortcut = conv_shortcut - - self.norm1 = Normalize(in_channels) - self.conv1 = torch.nn.Conv2d( - in_channels, out_channels, kernel_size=3, stride=1, padding=1 - ) - if temb_channels > 0: - self.temb_proj = torch.nn.Linear(temb_channels, out_channels) - self.norm2 = Normalize(out_channels) - self.dropout = torch.nn.Dropout(dropout) - self.conv2 = torch.nn.Conv2d( - out_channels, out_channels, kernel_size=3, stride=1, padding=1 - ) - if self.in_channels != self.out_channels: - if self.use_conv_shortcut: - self.conv_shortcut = torch.nn.Conv2d( - in_channels, out_channels, kernel_size=3, stride=1, padding=1 - ) - else: - self.nin_shortcut = torch.nn.Conv2d( - in_channels, out_channels, kernel_size=1, stride=1, padding=0 - ) - - def forward(self, x, temb): - h = x - h = self.norm1(h) - h = nonlinearity(h) - h = self.conv1(h) - - if temb is not None: - h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None] - - h = self.norm2(h) - h = nonlinearity(h) - h = self.dropout(h) - h = self.conv2(h) - - if self.in_channels != self.out_channels: - if self.use_conv_shortcut: - x = self.conv_shortcut(x) - else: - x = self.nin_shortcut(x) - - return x + h - - -class MultiHeadAttnBlock(nn.Module): - def __init__(self, in_channels, head_size=1): - super().__init__() - self.in_channels = in_channels - self.head_size = head_size - self.att_size = in_channels // head_size - assert ( - in_channels % head_size == 0 - ), "The size of head should be divided by the number of channels." - - self.norm1 = Normalize(in_channels) - self.norm2 = Normalize(in_channels) - - self.q = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=1, stride=1, padding=0 - ) - self.k = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=1, stride=1, padding=0 - ) - self.v = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=1, stride=1, padding=0 - ) - self.proj_out = torch.nn.Conv2d( - in_channels, in_channels, kernel_size=1, stride=1, padding=0 - ) - self.num = 0 - - def forward(self, x, y=None): - h_ = x - h_ = self.norm1(h_) - if y is None: - y = h_ - else: - y = self.norm2(y) - - q = self.q(y) - k = self.k(h_) - v = self.v(h_) - - # compute attention - b, c, h, w = q.shape - q = q.reshape(b, self.head_size, self.att_size, h * w) - q = q.permute(0, 3, 1, 2) # b, hw, head, att - - k = k.reshape(b, self.head_size, self.att_size, h * w) - k = k.permute(0, 3, 1, 2) - - v = v.reshape(b, self.head_size, self.att_size, h * w) - v = v.permute(0, 3, 1, 2) - - q = q.transpose(1, 2) - v = v.transpose(1, 2) - k = k.transpose(1, 2).transpose(2, 3) - - scale = int(self.att_size) ** (-0.5) - q.mul_(scale) - w_ = torch.matmul(q, k) - w_ = F.softmax(w_, dim=3) - - w_ = w_.matmul(v) - - w_ = w_.transpose(1, 2).contiguous() # [b, h*w, head, att] - w_ = w_.view(b, h, w, -1) - w_ = w_.permute(0, 3, 1, 2) - - w_ = self.proj_out(w_) - - return x + w_ - - -class MultiHeadEncoder(nn.Module): - def __init__( - self, - ch, - out_ch, - ch_mult=(1, 2, 4, 8), - num_res_blocks=2, - attn_resolutions=(16,), - dropout=0.0, - resamp_with_conv=True, - in_channels=3, - resolution=512, - z_channels=256, - double_z=True, - enable_mid=True, - head_size=1, - **ignore_kwargs - ): - super().__init__() - self.ch = ch - self.temb_ch = 0 - self.num_resolutions = len(ch_mult) - self.num_res_blocks = num_res_blocks - self.resolution = resolution - self.in_channels = in_channels - self.enable_mid = enable_mid - - # downsampling - self.conv_in = torch.nn.Conv2d( - in_channels, self.ch, kernel_size=3, stride=1, padding=1 - ) - - curr_res = resolution - in_ch_mult = (1,) + tuple(ch_mult) - self.down = nn.ModuleList() - for i_level in range(self.num_resolutions): - block = nn.ModuleList() - attn = nn.ModuleList() - block_in = ch * in_ch_mult[i_level] - block_out = ch * ch_mult[i_level] - for i_block in range(self.num_res_blocks): - block.append( - ResnetBlock( - in_channels=block_in, - out_channels=block_out, - temb_channels=self.temb_ch, - dropout=dropout, - ) - ) - block_in = block_out - if curr_res in attn_resolutions: - attn.append(MultiHeadAttnBlock(block_in, head_size)) - down = nn.Module() - down.block = block - down.attn = attn - if i_level != self.num_resolutions - 1: - down.downsample = Downsample(block_in, resamp_with_conv) - curr_res = curr_res // 2 - self.down.append(down) - - # middle - if self.enable_mid: - self.mid = nn.Module() - self.mid.block_1 = ResnetBlock( - in_channels=block_in, - out_channels=block_in, - temb_channels=self.temb_ch, - dropout=dropout, - ) - self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size) - self.mid.block_2 = ResnetBlock( - in_channels=block_in, - out_channels=block_in, - temb_channels=self.temb_ch, - dropout=dropout, - ) - - # end - self.norm_out = Normalize(block_in) - self.conv_out = torch.nn.Conv2d( - block_in, - 2 * z_channels if double_z else z_channels, - kernel_size=3, - stride=1, - padding=1, - ) - - def forward(self, x): - hs = {} - # timestep embedding - temb = None - - # downsampling - h = self.conv_in(x) - hs["in"] = h - for i_level in range(self.num_resolutions): - for i_block in range(self.num_res_blocks): - h = self.down[i_level].block[i_block](h, temb) - if len(self.down[i_level].attn) > 0: - h = self.down[i_level].attn[i_block](h) - - if i_level != self.num_resolutions - 1: - # hs.append(h) - hs["block_" + str(i_level)] = h - h = self.down[i_level].downsample(h) - - # middle - # h = hs[-1] - if self.enable_mid: - h = self.mid.block_1(h, temb) - hs["block_" + str(i_level) + "_atten"] = h - h = self.mid.attn_1(h) - h = self.mid.block_2(h, temb) - hs["mid_atten"] = h - - # end - h = self.norm_out(h) - h = nonlinearity(h) - h = self.conv_out(h) - # hs.append(h) - hs["out"] = h - - return hs - - -class MultiHeadDecoder(nn.Module): - def __init__( - self, - ch, - out_ch, - ch_mult=(1, 2, 4, 8), - num_res_blocks=2, - attn_resolutions=(16,), - dropout=0.0, - resamp_with_conv=True, - in_channels=3, - resolution=512, - z_channels=256, - give_pre_end=False, - enable_mid=True, - head_size=1, - **ignorekwargs - ): - super().__init__() - self.ch = ch - self.temb_ch = 0 - self.num_resolutions = len(ch_mult) - self.num_res_blocks = num_res_blocks - self.resolution = resolution - self.in_channels = in_channels - self.give_pre_end = give_pre_end - self.enable_mid = enable_mid - - # compute in_ch_mult, block_in and curr_res at lowest res - block_in = ch * ch_mult[self.num_resolutions - 1] - curr_res = resolution // 2 ** (self.num_resolutions - 1) - self.z_shape = (1, z_channels, curr_res, curr_res) - print( - "Working with z of shape {} = {} dimensions.".format( - self.z_shape, np.prod(self.z_shape) - ) - ) - - # z to block_in - self.conv_in = torch.nn.Conv2d( - z_channels, block_in, kernel_size=3, stride=1, padding=1 - ) - - # middle - if self.enable_mid: - self.mid = nn.Module() - self.mid.block_1 = ResnetBlock( - in_channels=block_in, - out_channels=block_in, - temb_channels=self.temb_ch, - dropout=dropout, - ) - self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size) - self.mid.block_2 = ResnetBlock( - in_channels=block_in, - out_channels=block_in, - temb_channels=self.temb_ch, - dropout=dropout, - ) - - # upsampling - self.up = nn.ModuleList() - for i_level in reversed(range(self.num_resolutions)): - block = nn.ModuleList() - attn = nn.ModuleList() - block_out = ch * ch_mult[i_level] - for i_block in range(self.num_res_blocks + 1): - block.append( - ResnetBlock( - in_channels=block_in, - out_channels=block_out, - temb_channels=self.temb_ch, - dropout=dropout, - ) - ) - block_in = block_out - if curr_res in attn_resolutions: - attn.append(MultiHeadAttnBlock(block_in, head_size)) - up = nn.Module() - up.block = block - up.attn = attn - if i_level != 0: - up.upsample = Upsample(block_in, resamp_with_conv) - curr_res = curr_res * 2 - self.up.insert(0, up) # prepend to get consistent order - - # end - self.norm_out = Normalize(block_in) - self.conv_out = torch.nn.Conv2d( - block_in, out_ch, kernel_size=3, stride=1, padding=1 - ) - - def forward(self, z): - # assert z.shape[1:] == self.z_shape[1:] - self.last_z_shape = z.shape - - # timestep embedding - temb = None - - # z to block_in - h = self.conv_in(z) - - # middle - if self.enable_mid: - h = self.mid.block_1(h, temb) - h = self.mid.attn_1(h) - h = self.mid.block_2(h, temb) - - # upsampling - for i_level in reversed(range(self.num_resolutions)): - for i_block in range(self.num_res_blocks + 1): - h = self.up[i_level].block[i_block](h, temb) - if len(self.up[i_level].attn) > 0: - h = self.up[i_level].attn[i_block](h) - if i_level != 0: - h = self.up[i_level].upsample(h) - - # end - if self.give_pre_end: - return h - - h = self.norm_out(h) - h = nonlinearity(h) - h = self.conv_out(h) - return h - - -class MultiHeadDecoderTransformer(nn.Module): - def __init__( - self, - ch, - out_ch, - ch_mult=(1, 2, 4, 8), - num_res_blocks=2, - attn_resolutions=(16,), - dropout=0.0, - resamp_with_conv=True, - in_channels=3, - resolution=512, - z_channels=256, - give_pre_end=False, - enable_mid=True, - head_size=1, - **ignorekwargs - ): - super().__init__() - self.ch = ch - self.temb_ch = 0 - self.num_resolutions = len(ch_mult) - self.num_res_blocks = num_res_blocks - self.resolution = resolution - self.in_channels = in_channels - self.give_pre_end = give_pre_end - self.enable_mid = enable_mid - - # compute in_ch_mult, block_in and curr_res at lowest res - block_in = ch * ch_mult[self.num_resolutions - 1] - curr_res = resolution // 2 ** (self.num_resolutions - 1) - self.z_shape = (1, z_channels, curr_res, curr_res) - print( - "Working with z of shape {} = {} dimensions.".format( - self.z_shape, np.prod(self.z_shape) - ) - ) - - # z to block_in - self.conv_in = torch.nn.Conv2d( - z_channels, block_in, kernel_size=3, stride=1, padding=1 - ) - - # middle - if self.enable_mid: - self.mid = nn.Module() - self.mid.block_1 = ResnetBlock( - in_channels=block_in, - out_channels=block_in, - temb_channels=self.temb_ch, - dropout=dropout, - ) - self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size) - self.mid.block_2 = ResnetBlock( - in_channels=block_in, - out_channels=block_in, - temb_channels=self.temb_ch, - dropout=dropout, - ) - - # upsampling - self.up = nn.ModuleList() - for i_level in reversed(range(self.num_resolutions)): - block = nn.ModuleList() - attn = nn.ModuleList() - block_out = ch * ch_mult[i_level] - for i_block in range(self.num_res_blocks + 1): - block.append( - ResnetBlock( - in_channels=block_in, - out_channels=block_out, - temb_channels=self.temb_ch, - dropout=dropout, - ) - ) - block_in = block_out - if curr_res in attn_resolutions: - attn.append(MultiHeadAttnBlock(block_in, head_size)) - up = nn.Module() - up.block = block - up.attn = attn - if i_level != 0: - up.upsample = Upsample(block_in, resamp_with_conv) - curr_res = curr_res * 2 - self.up.insert(0, up) # prepend to get consistent order - - # end - self.norm_out = Normalize(block_in) - self.conv_out = torch.nn.Conv2d( - block_in, out_ch, kernel_size=3, stride=1, padding=1 - ) - - def forward(self, z, hs): - # assert z.shape[1:] == self.z_shape[1:] - # self.last_z_shape = z.shape - - # timestep embedding - temb = None - - # z to block_in - h = self.conv_in(z) - - # middle - if self.enable_mid: - h = self.mid.block_1(h, temb) - h = self.mid.attn_1(h, hs["mid_atten"]) - h = self.mid.block_2(h, temb) - - # upsampling - for i_level in reversed(range(self.num_resolutions)): - for i_block in range(self.num_res_blocks + 1): - h = self.up[i_level].block[i_block](h, temb) - if len(self.up[i_level].attn) > 0: - h = self.up[i_level].attn[i_block]( - h, hs["block_" + str(i_level) + "_atten"] - ) - # hfeature = h.clone() - if i_level != 0: - h = self.up[i_level].upsample(h) - - # end - if self.give_pre_end: - return h - - h = self.norm_out(h) - h = nonlinearity(h) - h = self.conv_out(h) - return h - - -class RestoreFormer(nn.Module): - def __init__( - self, - state_dict, - ): - super(RestoreFormer, self).__init__() - - n_embed = 1024 - embed_dim = 256 - ch = 64 - out_ch = 3 - ch_mult = (1, 2, 2, 4, 4, 8) - num_res_blocks = 2 - attn_resolutions = (16,) - dropout = 0.0 - in_channels = 3 - resolution = 512 - z_channels = 256 - double_z = False - enable_mid = True - fix_decoder = False - fix_codebook = True - fix_encoder = False - head_size = 8 - - self.model_arch = "RestoreFormer" - self.sub_type = "Face SR" - self.scale = 8 - self.in_nc = 3 - self.out_nc = out_ch - self.state = state_dict - - self.supports_fp16 = False - self.supports_bf16 = True - self.min_size_restriction = 16 - - self.encoder = MultiHeadEncoder( - ch=ch, - out_ch=out_ch, - ch_mult=ch_mult, - num_res_blocks=num_res_blocks, - attn_resolutions=attn_resolutions, - dropout=dropout, - in_channels=in_channels, - resolution=resolution, - z_channels=z_channels, - double_z=double_z, - enable_mid=enable_mid, - head_size=head_size, - ) - self.decoder = MultiHeadDecoderTransformer( - ch=ch, - out_ch=out_ch, - ch_mult=ch_mult, - num_res_blocks=num_res_blocks, - attn_resolutions=attn_resolutions, - dropout=dropout, - in_channels=in_channels, - resolution=resolution, - z_channels=z_channels, - enable_mid=enable_mid, - head_size=head_size, - ) - - self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25) - - self.quant_conv = torch.nn.Conv2d(z_channels, embed_dim, 1) - self.post_quant_conv = torch.nn.Conv2d(embed_dim, z_channels, 1) - - if fix_decoder: - for _, param in self.decoder.named_parameters(): - param.requires_grad = False - for _, param in self.post_quant_conv.named_parameters(): - param.requires_grad = False - for _, param in self.quantize.named_parameters(): - param.requires_grad = False - elif fix_codebook: - for _, param in self.quantize.named_parameters(): - param.requires_grad = False - - if fix_encoder: - for _, param in self.encoder.named_parameters(): - param.requires_grad = False - - self.load_state_dict(state_dict) - - def encode(self, x): - hs = self.encoder(x) - h = self.quant_conv(hs["out"]) - quant, emb_loss, info = self.quantize(h) - return quant, emb_loss, info, hs - - def decode(self, quant, hs): - quant = self.post_quant_conv(quant) - dec = self.decoder(quant, hs) - - return dec - - def forward(self, input, **kwargs): - quant, diff, info, hs = self.encode(input) - dec = self.decode(quant, hs) - - return dec, None diff --git a/comfy_extras/chainner_models/architecture/face/stylegan2_arch.py b/comfy_extras/chainner_models/architecture/face/stylegan2_arch.py deleted file mode 100644 index 1eb0e9f15..000000000 --- a/comfy_extras/chainner_models/architecture/face/stylegan2_arch.py +++ /dev/null @@ -1,865 +0,0 @@ -# pylint: skip-file -# type: ignore -import math -import random - -import torch -from torch import nn -from torch.nn import functional as F - -from .fused_act import FusedLeakyReLU, fused_leaky_relu -from .upfirdn2d import upfirdn2d - - -class NormStyleCode(nn.Module): - def forward(self, x): - """Normalize the style codes. - - Args: - x (Tensor): Style codes with shape (b, c). - - Returns: - Tensor: Normalized tensor. - """ - return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8) - - -def make_resample_kernel(k): - """Make resampling kernel for UpFirDn. - - Args: - k (list[int]): A list indicating the 1D resample kernel magnitude. - - Returns: - Tensor: 2D resampled kernel. - """ - k = torch.tensor(k, dtype=torch.float32) - if k.ndim == 1: - k = k[None, :] * k[:, None] # to 2D kernel, outer product - # normalize - k /= k.sum() - return k - - -class UpFirDnUpsample(nn.Module): - """Upsample, FIR filter, and downsample (upsampole version). - - References: - 1. https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.upfirdn.html # noqa: E501 - 2. http://www.ece.northwestern.edu/local-apps/matlabhelp/toolbox/signal/upfirdn.html # noqa: E501 - - Args: - resample_kernel (list[int]): A list indicating the 1D resample kernel - magnitude. - factor (int): Upsampling scale factor. Default: 2. - """ - - def __init__(self, resample_kernel, factor=2): - super(UpFirDnUpsample, self).__init__() - self.kernel = make_resample_kernel(resample_kernel) * (factor**2) - self.factor = factor - - pad = self.kernel.shape[0] - factor - self.pad = ((pad + 1) // 2 + factor - 1, pad // 2) - - def forward(self, x): - out = upfirdn2d(x, self.kernel.type_as(x), up=self.factor, down=1, pad=self.pad) - return out - - def __repr__(self): - return f"{self.__class__.__name__}(factor={self.factor})" - - -class UpFirDnDownsample(nn.Module): - """Upsample, FIR filter, and downsample (downsampole version). - - Args: - resample_kernel (list[int]): A list indicating the 1D resample kernel - magnitude. - factor (int): Downsampling scale factor. Default: 2. - """ - - def __init__(self, resample_kernel, factor=2): - super(UpFirDnDownsample, self).__init__() - self.kernel = make_resample_kernel(resample_kernel) - self.factor = factor - - pad = self.kernel.shape[0] - factor - self.pad = ((pad + 1) // 2, pad // 2) - - def forward(self, x): - out = upfirdn2d(x, self.kernel.type_as(x), up=1, down=self.factor, pad=self.pad) - return out - - def __repr__(self): - return f"{self.__class__.__name__}(factor={self.factor})" - - -class UpFirDnSmooth(nn.Module): - """Upsample, FIR filter, and downsample (smooth version). - - Args: - resample_kernel (list[int]): A list indicating the 1D resample kernel - magnitude. - upsample_factor (int): Upsampling scale factor. Default: 1. - downsample_factor (int): Downsampling scale factor. Default: 1. - kernel_size (int): Kernel size: Default: 1. - """ - - def __init__( - self, resample_kernel, upsample_factor=1, downsample_factor=1, kernel_size=1 - ): - super(UpFirDnSmooth, self).__init__() - self.upsample_factor = upsample_factor - self.downsample_factor = downsample_factor - self.kernel = make_resample_kernel(resample_kernel) - if upsample_factor > 1: - self.kernel = self.kernel * (upsample_factor**2) - - if upsample_factor > 1: - pad = (self.kernel.shape[0] - upsample_factor) - (kernel_size - 1) - self.pad = ((pad + 1) // 2 + upsample_factor - 1, pad // 2 + 1) - elif downsample_factor > 1: - pad = (self.kernel.shape[0] - downsample_factor) + (kernel_size - 1) - self.pad = ((pad + 1) // 2, pad // 2) - else: - raise NotImplementedError - - def forward(self, x): - out = upfirdn2d(x, self.kernel.type_as(x), up=1, down=1, pad=self.pad) - return out - - def __repr__(self): - return ( - f"{self.__class__.__name__}(upsample_factor={self.upsample_factor}" - f", downsample_factor={self.downsample_factor})" - ) - - -class EqualLinear(nn.Module): - """Equalized Linear as StyleGAN2. - - Args: - in_channels (int): Size of each sample. - out_channels (int): Size of each output sample. - bias (bool): If set to ``False``, the layer will not learn an additive - bias. Default: ``True``. - bias_init_val (float): Bias initialized value. Default: 0. - lr_mul (float): Learning rate multiplier. Default: 1. - activation (None | str): The activation after ``linear`` operation. - Supported: 'fused_lrelu', None. Default: None. - """ - - def __init__( - self, - in_channels, - out_channels, - bias=True, - bias_init_val=0, - lr_mul=1, - activation=None, - ): - super(EqualLinear, self).__init__() - self.in_channels = in_channels - self.out_channels = out_channels - self.lr_mul = lr_mul - self.activation = activation - if self.activation not in ["fused_lrelu", None]: - raise ValueError( - f"Wrong activation value in EqualLinear: {activation}" - "Supported ones are: ['fused_lrelu', None]." - ) - self.scale = (1 / math.sqrt(in_channels)) * lr_mul - - self.weight = nn.Parameter(torch.randn(out_channels, in_channels).div_(lr_mul)) - if bias: - self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val)) - else: - self.register_parameter("bias", None) - - def forward(self, x): - if self.bias is None: - bias = None - else: - bias = self.bias * self.lr_mul - if self.activation == "fused_lrelu": - out = F.linear(x, self.weight * self.scale) - out = fused_leaky_relu(out, bias) - else: - out = F.linear(x, self.weight * self.scale, bias=bias) - return out - - def __repr__(self): - return ( - f"{self.__class__.__name__}(in_channels={self.in_channels}, " - f"out_channels={self.out_channels}, bias={self.bias is not None})" - ) - - -class ModulatedConv2d(nn.Module): - """Modulated Conv2d used in StyleGAN2. - - There is no bias in ModulatedConv2d. - - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Size of the convolving kernel. - num_style_feat (int): Channel number of style features. - demodulate (bool): Whether to demodulate in the conv layer. - Default: True. - sample_mode (str | None): Indicating 'upsample', 'downsample' or None. - Default: None. - resample_kernel (list[int]): A list indicating the 1D resample kernel - magnitude. Default: (1, 3, 3, 1). - eps (float): A value added to the denominator for numerical stability. - Default: 1e-8. - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - num_style_feat, - demodulate=True, - sample_mode=None, - resample_kernel=(1, 3, 3, 1), - eps=1e-8, - ): - super(ModulatedConv2d, self).__init__() - self.in_channels = in_channels - self.out_channels = out_channels - self.kernel_size = kernel_size - self.demodulate = demodulate - self.sample_mode = sample_mode - self.eps = eps - - if self.sample_mode == "upsample": - self.smooth = UpFirDnSmooth( - resample_kernel, - upsample_factor=2, - downsample_factor=1, - kernel_size=kernel_size, - ) - elif self.sample_mode == "downsample": - self.smooth = UpFirDnSmooth( - resample_kernel, - upsample_factor=1, - downsample_factor=2, - kernel_size=kernel_size, - ) - elif self.sample_mode is None: - pass - else: - raise ValueError( - f"Wrong sample mode {self.sample_mode}, " - "supported ones are ['upsample', 'downsample', None]." - ) - - self.scale = 1 / math.sqrt(in_channels * kernel_size**2) - # modulation inside each modulated conv - self.modulation = EqualLinear( - num_style_feat, - in_channels, - bias=True, - bias_init_val=1, - lr_mul=1, - activation=None, - ) - - self.weight = nn.Parameter( - torch.randn(1, out_channels, in_channels, kernel_size, kernel_size) - ) - self.padding = kernel_size // 2 - - def forward(self, x, style): - """Forward function. - - Args: - x (Tensor): Tensor with shape (b, c, h, w). - style (Tensor): Tensor with shape (b, num_style_feat). - - Returns: - Tensor: Modulated tensor after convolution. - """ - b, c, h, w = x.shape # c = c_in - # weight modulation - style = self.modulation(style).view(b, 1, c, 1, 1) - # self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1) - weight = self.scale * self.weight * style # (b, c_out, c_in, k, k) - - if self.demodulate: - demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps) - weight = weight * demod.view(b, self.out_channels, 1, 1, 1) - - weight = weight.view( - b * self.out_channels, c, self.kernel_size, self.kernel_size - ) - - if self.sample_mode == "upsample": - x = x.view(1, b * c, h, w) - weight = weight.view( - b, self.out_channels, c, self.kernel_size, self.kernel_size - ) - weight = weight.transpose(1, 2).reshape( - b * c, self.out_channels, self.kernel_size, self.kernel_size - ) - out = F.conv_transpose2d(x, weight, padding=0, stride=2, groups=b) - out = out.view(b, self.out_channels, *out.shape[2:4]) - out = self.smooth(out) - elif self.sample_mode == "downsample": - x = self.smooth(x) - x = x.view(1, b * c, *x.shape[2:4]) - out = F.conv2d(x, weight, padding=0, stride=2, groups=b) - out = out.view(b, self.out_channels, *out.shape[2:4]) - else: - x = x.view(1, b * c, h, w) - # weight: (b*c_out, c_in, k, k), groups=b - out = F.conv2d(x, weight, padding=self.padding, groups=b) - out = out.view(b, self.out_channels, *out.shape[2:4]) - - return out - - def __repr__(self): - return ( - f"{self.__class__.__name__}(in_channels={self.in_channels}, " - f"out_channels={self.out_channels}, " - f"kernel_size={self.kernel_size}, " - f"demodulate={self.demodulate}, sample_mode={self.sample_mode})" - ) - - -class StyleConv(nn.Module): - """Style conv. - - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Size of the convolving kernel. - num_style_feat (int): Channel number of style features. - demodulate (bool): Whether demodulate in the conv layer. Default: True. - sample_mode (str | None): Indicating 'upsample', 'downsample' or None. - Default: None. - resample_kernel (list[int]): A list indicating the 1D resample kernel - magnitude. Default: (1, 3, 3, 1). - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - num_style_feat, - demodulate=True, - sample_mode=None, - resample_kernel=(1, 3, 3, 1), - ): - super(StyleConv, self).__init__() - self.modulated_conv = ModulatedConv2d( - in_channels, - out_channels, - kernel_size, - num_style_feat, - demodulate=demodulate, - sample_mode=sample_mode, - resample_kernel=resample_kernel, - ) - self.weight = nn.Parameter(torch.zeros(1)) # for noise injection - self.activate = FusedLeakyReLU(out_channels) - - def forward(self, x, style, noise=None): - # modulate - out = self.modulated_conv(x, style) - # noise injection - if noise is None: - b, _, h, w = out.shape - noise = out.new_empty(b, 1, h, w).normal_() - out = out + self.weight * noise - # activation (with bias) - out = self.activate(out) - return out - - -class ToRGB(nn.Module): - """To RGB from features. - - Args: - in_channels (int): Channel number of input. - num_style_feat (int): Channel number of style features. - upsample (bool): Whether to upsample. Default: True. - resample_kernel (list[int]): A list indicating the 1D resample kernel - magnitude. Default: (1, 3, 3, 1). - """ - - def __init__( - self, in_channels, num_style_feat, upsample=True, resample_kernel=(1, 3, 3, 1) - ): - super(ToRGB, self).__init__() - if upsample: - self.upsample = UpFirDnUpsample(resample_kernel, factor=2) - else: - self.upsample = None - self.modulated_conv = ModulatedConv2d( - in_channels, - 3, - kernel_size=1, - num_style_feat=num_style_feat, - demodulate=False, - sample_mode=None, - ) - self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1)) - - def forward(self, x, style, skip=None): - """Forward function. - - Args: - x (Tensor): Feature tensor with shape (b, c, h, w). - style (Tensor): Tensor with shape (b, num_style_feat). - skip (Tensor): Base/skip tensor. Default: None. - - Returns: - Tensor: RGB images. - """ - out = self.modulated_conv(x, style) - out = out + self.bias - if skip is not None: - if self.upsample: - skip = self.upsample(skip) - out = out + skip - return out - - -class ConstantInput(nn.Module): - """Constant input. - - Args: - num_channel (int): Channel number of constant input. - size (int): Spatial size of constant input. - """ - - def __init__(self, num_channel, size): - super(ConstantInput, self).__init__() - self.weight = nn.Parameter(torch.randn(1, num_channel, size, size)) - - def forward(self, batch): - out = self.weight.repeat(batch, 1, 1, 1) - return out - - -class StyleGAN2Generator(nn.Module): - """StyleGAN2 Generator. - - Args: - out_size (int): The spatial size of outputs. - num_style_feat (int): Channel number of style features. Default: 512. - num_mlp (int): Layer number of MLP style layers. Default: 8. - channel_multiplier (int): Channel multiplier for large networks of - StyleGAN2. Default: 2. - resample_kernel (list[int]): A list indicating the 1D resample kernel - magnitude. A cross production will be applied to extent 1D resample - kernel to 2D resample kernel. Default: (1, 3, 3, 1). - lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01. - narrow (float): Narrow ratio for channels. Default: 1.0. - """ - - def __init__( - self, - out_size, - num_style_feat=512, - num_mlp=8, - channel_multiplier=2, - resample_kernel=(1, 3, 3, 1), - lr_mlp=0.01, - narrow=1, - ): - super(StyleGAN2Generator, self).__init__() - # Style MLP layers - self.num_style_feat = num_style_feat - style_mlp_layers = [NormStyleCode()] - for i in range(num_mlp): - style_mlp_layers.append( - EqualLinear( - num_style_feat, - num_style_feat, - bias=True, - bias_init_val=0, - lr_mul=lr_mlp, - activation="fused_lrelu", - ) - ) - self.style_mlp = nn.Sequential(*style_mlp_layers) - - channels = { - "4": int(512 * narrow), - "8": int(512 * narrow), - "16": int(512 * narrow), - "32": int(512 * narrow), - "64": int(256 * channel_multiplier * narrow), - "128": int(128 * channel_multiplier * narrow), - "256": int(64 * channel_multiplier * narrow), - "512": int(32 * channel_multiplier * narrow), - "1024": int(16 * channel_multiplier * narrow), - } - self.channels = channels - - self.constant_input = ConstantInput(channels["4"], size=4) - self.style_conv1 = StyleConv( - channels["4"], - channels["4"], - kernel_size=3, - num_style_feat=num_style_feat, - demodulate=True, - sample_mode=None, - resample_kernel=resample_kernel, - ) - self.to_rgb1 = ToRGB( - channels["4"], - num_style_feat, - upsample=False, - resample_kernel=resample_kernel, - ) - - self.log_size = int(math.log(out_size, 2)) - self.num_layers = (self.log_size - 2) * 2 + 1 - self.num_latent = self.log_size * 2 - 2 - - self.style_convs = nn.ModuleList() - self.to_rgbs = nn.ModuleList() - self.noises = nn.Module() - - in_channels = channels["4"] - # noise - for layer_idx in range(self.num_layers): - resolution = 2 ** ((layer_idx + 5) // 2) - shape = [1, 1, resolution, resolution] - self.noises.register_buffer(f"noise{layer_idx}", torch.randn(*shape)) - # style convs and to_rgbs - for i in range(3, self.log_size + 1): - out_channels = channels[f"{2**i}"] - self.style_convs.append( - StyleConv( - in_channels, - out_channels, - kernel_size=3, - num_style_feat=num_style_feat, - demodulate=True, - sample_mode="upsample", - resample_kernel=resample_kernel, - ) - ) - self.style_convs.append( - StyleConv( - out_channels, - out_channels, - kernel_size=3, - num_style_feat=num_style_feat, - demodulate=True, - sample_mode=None, - resample_kernel=resample_kernel, - ) - ) - self.to_rgbs.append( - ToRGB( - out_channels, - num_style_feat, - upsample=True, - resample_kernel=resample_kernel, - ) - ) - in_channels = out_channels - - def make_noise(self): - """Make noise for noise injection.""" - device = self.constant_input.weight.device - noises = [torch.randn(1, 1, 4, 4, device=device)] - - for i in range(3, self.log_size + 1): - for _ in range(2): - noises.append(torch.randn(1, 1, 2**i, 2**i, device=device)) - - return noises - - def get_latent(self, x): - return self.style_mlp(x) - - def mean_latent(self, num_latent): - latent_in = torch.randn( - num_latent, self.num_style_feat, device=self.constant_input.weight.device - ) - latent = self.style_mlp(latent_in).mean(0, keepdim=True) - return latent - - def forward( - self, - styles, - input_is_latent=False, - noise=None, - randomize_noise=True, - truncation=1, - truncation_latent=None, - inject_index=None, - return_latents=False, - ): - """Forward function for StyleGAN2Generator. - - Args: - styles (list[Tensor]): Sample codes of styles. - input_is_latent (bool): Whether input is latent style. - Default: False. - noise (Tensor | None): Input noise or None. Default: None. - randomize_noise (bool): Randomize noise, used when 'noise' is - False. Default: True. - truncation (float): TODO. Default: 1. - truncation_latent (Tensor | None): TODO. Default: None. - inject_index (int | None): The injection index for mixing noise. - Default: None. - return_latents (bool): Whether to return style latents. - Default: False. - """ - # style codes -> latents with Style MLP layer - if not input_is_latent: - styles = [self.style_mlp(s) for s in styles] - # noises - if noise is None: - if randomize_noise: - noise = [None] * self.num_layers # for each style conv layer - else: # use the stored noise - noise = [ - getattr(self.noises, f"noise{i}") for i in range(self.num_layers) - ] - # style truncation - if truncation < 1: - style_truncation = [] - for style in styles: - style_truncation.append( - truncation_latent + truncation * (style - truncation_latent) - ) - styles = style_truncation - # get style latent with injection - if len(styles) == 1: - inject_index = self.num_latent - - if styles[0].ndim < 3: - # repeat latent code for all the layers - latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - else: # used for encoder with different latent code for each layer - latent = styles[0] - elif len(styles) == 2: # mixing noises - if inject_index is None: - inject_index = random.randint(1, self.num_latent - 1) - latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - latent2 = ( - styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1) - ) - latent = torch.cat([latent1, latent2], 1) - - # main generation - out = self.constant_input(latent.shape[0]) - out = self.style_conv1(out, latent[:, 0], noise=noise[0]) - skip = self.to_rgb1(out, latent[:, 1]) - - i = 1 - for conv1, conv2, noise1, noise2, to_rgb in zip( - self.style_convs[::2], - self.style_convs[1::2], - noise[1::2], - noise[2::2], - self.to_rgbs, - ): - out = conv1(out, latent[:, i], noise=noise1) - out = conv2(out, latent[:, i + 1], noise=noise2) - skip = to_rgb(out, latent[:, i + 2], skip) - i += 2 - - image = skip - - if return_latents: - return image, latent - else: - return image, None - - -class ScaledLeakyReLU(nn.Module): - """Scaled LeakyReLU. - - Args: - negative_slope (float): Negative slope. Default: 0.2. - """ - - def __init__(self, negative_slope=0.2): - super(ScaledLeakyReLU, self).__init__() - self.negative_slope = negative_slope - - def forward(self, x): - out = F.leaky_relu(x, negative_slope=self.negative_slope) - return out * math.sqrt(2) - - -class EqualConv2d(nn.Module): - """Equalized Linear as StyleGAN2. - - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Size of the convolving kernel. - stride (int): Stride of the convolution. Default: 1 - padding (int): Zero-padding added to both sides of the input. - Default: 0. - bias (bool): If ``True``, adds a learnable bias to the output. - Default: ``True``. - bias_init_val (float): Bias initialized value. Default: 0. - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - stride=1, - padding=0, - bias=True, - bias_init_val=0, - ): - super(EqualConv2d, self).__init__() - self.in_channels = in_channels - self.out_channels = out_channels - self.kernel_size = kernel_size - self.stride = stride - self.padding = padding - self.scale = 1 / math.sqrt(in_channels * kernel_size**2) - - self.weight = nn.Parameter( - torch.randn(out_channels, in_channels, kernel_size, kernel_size) - ) - if bias: - self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val)) - else: - self.register_parameter("bias", None) - - def forward(self, x): - out = F.conv2d( - x, - self.weight * self.scale, - bias=self.bias, - stride=self.stride, - padding=self.padding, - ) - - return out - - def __repr__(self): - return ( - f"{self.__class__.__name__}(in_channels={self.in_channels}, " - f"out_channels={self.out_channels}, " - f"kernel_size={self.kernel_size}," - f" stride={self.stride}, padding={self.padding}, " - f"bias={self.bias is not None})" - ) - - -class ConvLayer(nn.Sequential): - """Conv Layer used in StyleGAN2 Discriminator. - - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Kernel size. - downsample (bool): Whether downsample by a factor of 2. - Default: False. - resample_kernel (list[int]): A list indicating the 1D resample - kernel magnitude. A cross production will be applied to - extent 1D resample kernel to 2D resample kernel. - Default: (1, 3, 3, 1). - bias (bool): Whether with bias. Default: True. - activate (bool): Whether use activateion. Default: True. - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - downsample=False, - resample_kernel=(1, 3, 3, 1), - bias=True, - activate=True, - ): - layers = [] - # downsample - if downsample: - layers.append( - UpFirDnSmooth( - resample_kernel, - upsample_factor=1, - downsample_factor=2, - kernel_size=kernel_size, - ) - ) - stride = 2 - self.padding = 0 - else: - stride = 1 - self.padding = kernel_size // 2 - # conv - layers.append( - EqualConv2d( - in_channels, - out_channels, - kernel_size, - stride=stride, - padding=self.padding, - bias=bias and not activate, - ) - ) - # activation - if activate: - if bias: - layers.append(FusedLeakyReLU(out_channels)) - else: - layers.append(ScaledLeakyReLU(0.2)) - - super(ConvLayer, self).__init__(*layers) - - -class ResBlock(nn.Module): - """Residual block used in StyleGAN2 Discriminator. - - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - resample_kernel (list[int]): A list indicating the 1D resample - kernel magnitude. A cross production will be applied to - extent 1D resample kernel to 2D resample kernel. - Default: (1, 3, 3, 1). - """ - - def __init__(self, in_channels, out_channels, resample_kernel=(1, 3, 3, 1)): - super(ResBlock, self).__init__() - - self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True) - self.conv2 = ConvLayer( - in_channels, - out_channels, - 3, - downsample=True, - resample_kernel=resample_kernel, - bias=True, - activate=True, - ) - self.skip = ConvLayer( - in_channels, - out_channels, - 1, - downsample=True, - resample_kernel=resample_kernel, - bias=False, - activate=False, - ) - - def forward(self, x): - out = self.conv1(x) - out = self.conv2(out) - skip = self.skip(x) - out = (out + skip) / math.sqrt(2) - return out diff --git a/comfy_extras/chainner_models/architecture/face/stylegan2_bilinear_arch.py b/comfy_extras/chainner_models/architecture/face/stylegan2_bilinear_arch.py deleted file mode 100644 index 601f8cc4b..000000000 --- a/comfy_extras/chainner_models/architecture/face/stylegan2_bilinear_arch.py +++ /dev/null @@ -1,709 +0,0 @@ -# pylint: skip-file -# type: ignore -import math -import random - -import torch -from torch import nn -from torch.nn import functional as F - -from .fused_act import FusedLeakyReLU, fused_leaky_relu - - -class NormStyleCode(nn.Module): - def forward(self, x): - """Normalize the style codes. - Args: - x (Tensor): Style codes with shape (b, c). - Returns: - Tensor: Normalized tensor. - """ - return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8) - - -class EqualLinear(nn.Module): - """Equalized Linear as StyleGAN2. - Args: - in_channels (int): Size of each sample. - out_channels (int): Size of each output sample. - bias (bool): If set to ``False``, the layer will not learn an additive - bias. Default: ``True``. - bias_init_val (float): Bias initialized value. Default: 0. - lr_mul (float): Learning rate multiplier. Default: 1. - activation (None | str): The activation after ``linear`` operation. - Supported: 'fused_lrelu', None. Default: None. - """ - - def __init__( - self, - in_channels, - out_channels, - bias=True, - bias_init_val=0, - lr_mul=1, - activation=None, - ): - super(EqualLinear, self).__init__() - self.in_channels = in_channels - self.out_channels = out_channels - self.lr_mul = lr_mul - self.activation = activation - if self.activation not in ["fused_lrelu", None]: - raise ValueError( - f"Wrong activation value in EqualLinear: {activation}" - "Supported ones are: ['fused_lrelu', None]." - ) - self.scale = (1 / math.sqrt(in_channels)) * lr_mul - - self.weight = nn.Parameter(torch.randn(out_channels, in_channels).div_(lr_mul)) - if bias: - self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val)) - else: - self.register_parameter("bias", None) - - def forward(self, x): - if self.bias is None: - bias = None - else: - bias = self.bias * self.lr_mul - if self.activation == "fused_lrelu": - out = F.linear(x, self.weight * self.scale) - out = fused_leaky_relu(out, bias) - else: - out = F.linear(x, self.weight * self.scale, bias=bias) - return out - - def __repr__(self): - return ( - f"{self.__class__.__name__}(in_channels={self.in_channels}, " - f"out_channels={self.out_channels}, bias={self.bias is not None})" - ) - - -class ModulatedConv2d(nn.Module): - """Modulated Conv2d used in StyleGAN2. - There is no bias in ModulatedConv2d. - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Size of the convolving kernel. - num_style_feat (int): Channel number of style features. - demodulate (bool): Whether to demodulate in the conv layer. - Default: True. - sample_mode (str | None): Indicating 'upsample', 'downsample' or None. - Default: None. - eps (float): A value added to the denominator for numerical stability. - Default: 1e-8. - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - num_style_feat, - demodulate=True, - sample_mode=None, - eps=1e-8, - interpolation_mode="bilinear", - ): - super(ModulatedConv2d, self).__init__() - self.in_channels = in_channels - self.out_channels = out_channels - self.kernel_size = kernel_size - self.demodulate = demodulate - self.sample_mode = sample_mode - self.eps = eps - self.interpolation_mode = interpolation_mode - if self.interpolation_mode == "nearest": - self.align_corners = None - else: - self.align_corners = False - - self.scale = 1 / math.sqrt(in_channels * kernel_size**2) - # modulation inside each modulated conv - self.modulation = EqualLinear( - num_style_feat, - in_channels, - bias=True, - bias_init_val=1, - lr_mul=1, - activation=None, - ) - - self.weight = nn.Parameter( - torch.randn(1, out_channels, in_channels, kernel_size, kernel_size) - ) - self.padding = kernel_size // 2 - - def forward(self, x, style): - """Forward function. - Args: - x (Tensor): Tensor with shape (b, c, h, w). - style (Tensor): Tensor with shape (b, num_style_feat). - Returns: - Tensor: Modulated tensor after convolution. - """ - b, c, h, w = x.shape # c = c_in - # weight modulation - style = self.modulation(style).view(b, 1, c, 1, 1) - # self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1) - weight = self.scale * self.weight * style # (b, c_out, c_in, k, k) - - if self.demodulate: - demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps) - weight = weight * demod.view(b, self.out_channels, 1, 1, 1) - - weight = weight.view( - b * self.out_channels, c, self.kernel_size, self.kernel_size - ) - - if self.sample_mode == "upsample": - x = F.interpolate( - x, - scale_factor=2, - mode=self.interpolation_mode, - align_corners=self.align_corners, - ) - elif self.sample_mode == "downsample": - x = F.interpolate( - x, - scale_factor=0.5, - mode=self.interpolation_mode, - align_corners=self.align_corners, - ) - - b, c, h, w = x.shape - x = x.view(1, b * c, h, w) - # weight: (b*c_out, c_in, k, k), groups=b - out = F.conv2d(x, weight, padding=self.padding, groups=b) - out = out.view(b, self.out_channels, *out.shape[2:4]) - - return out - - def __repr__(self): - return ( - f"{self.__class__.__name__}(in_channels={self.in_channels}, " - f"out_channels={self.out_channels}, " - f"kernel_size={self.kernel_size}, " - f"demodulate={self.demodulate}, sample_mode={self.sample_mode})" - ) - - -class StyleConv(nn.Module): - """Style conv. - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Size of the convolving kernel. - num_style_feat (int): Channel number of style features. - demodulate (bool): Whether demodulate in the conv layer. Default: True. - sample_mode (str | None): Indicating 'upsample', 'downsample' or None. - Default: None. - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - num_style_feat, - demodulate=True, - sample_mode=None, - interpolation_mode="bilinear", - ): - super(StyleConv, self).__init__() - self.modulated_conv = ModulatedConv2d( - in_channels, - out_channels, - kernel_size, - num_style_feat, - demodulate=demodulate, - sample_mode=sample_mode, - interpolation_mode=interpolation_mode, - ) - self.weight = nn.Parameter(torch.zeros(1)) # for noise injection - self.activate = FusedLeakyReLU(out_channels) - - def forward(self, x, style, noise=None): - # modulate - out = self.modulated_conv(x, style) - # noise injection - if noise is None: - b, _, h, w = out.shape - noise = out.new_empty(b, 1, h, w).normal_() - out = out + self.weight * noise - # activation (with bias) - out = self.activate(out) - return out - - -class ToRGB(nn.Module): - """To RGB from features. - Args: - in_channels (int): Channel number of input. - num_style_feat (int): Channel number of style features. - upsample (bool): Whether to upsample. Default: True. - """ - - def __init__( - self, in_channels, num_style_feat, upsample=True, interpolation_mode="bilinear" - ): - super(ToRGB, self).__init__() - self.upsample = upsample - self.interpolation_mode = interpolation_mode - if self.interpolation_mode == "nearest": - self.align_corners = None - else: - self.align_corners = False - self.modulated_conv = ModulatedConv2d( - in_channels, - 3, - kernel_size=1, - num_style_feat=num_style_feat, - demodulate=False, - sample_mode=None, - interpolation_mode=interpolation_mode, - ) - self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1)) - - def forward(self, x, style, skip=None): - """Forward function. - Args: - x (Tensor): Feature tensor with shape (b, c, h, w). - style (Tensor): Tensor with shape (b, num_style_feat). - skip (Tensor): Base/skip tensor. Default: None. - Returns: - Tensor: RGB images. - """ - out = self.modulated_conv(x, style) - out = out + self.bias - if skip is not None: - if self.upsample: - skip = F.interpolate( - skip, - scale_factor=2, - mode=self.interpolation_mode, - align_corners=self.align_corners, - ) - out = out + skip - return out - - -class ConstantInput(nn.Module): - """Constant input. - Args: - num_channel (int): Channel number of constant input. - size (int): Spatial size of constant input. - """ - - def __init__(self, num_channel, size): - super(ConstantInput, self).__init__() - self.weight = nn.Parameter(torch.randn(1, num_channel, size, size)) - - def forward(self, batch): - out = self.weight.repeat(batch, 1, 1, 1) - return out - - -class StyleGAN2GeneratorBilinear(nn.Module): - """StyleGAN2 Generator. - Args: - out_size (int): The spatial size of outputs. - num_style_feat (int): Channel number of style features. Default: 512. - num_mlp (int): Layer number of MLP style layers. Default: 8. - channel_multiplier (int): Channel multiplier for large networks of - StyleGAN2. Default: 2. - lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01. - narrow (float): Narrow ratio for channels. Default: 1.0. - """ - - def __init__( - self, - out_size, - num_style_feat=512, - num_mlp=8, - channel_multiplier=2, - lr_mlp=0.01, - narrow=1, - interpolation_mode="bilinear", - ): - super(StyleGAN2GeneratorBilinear, self).__init__() - # Style MLP layers - self.num_style_feat = num_style_feat - style_mlp_layers = [NormStyleCode()] - for i in range(num_mlp): - style_mlp_layers.append( - EqualLinear( - num_style_feat, - num_style_feat, - bias=True, - bias_init_val=0, - lr_mul=lr_mlp, - activation="fused_lrelu", - ) - ) - self.style_mlp = nn.Sequential(*style_mlp_layers) - - channels = { - "4": int(512 * narrow), - "8": int(512 * narrow), - "16": int(512 * narrow), - "32": int(512 * narrow), - "64": int(256 * channel_multiplier * narrow), - "128": int(128 * channel_multiplier * narrow), - "256": int(64 * channel_multiplier * narrow), - "512": int(32 * channel_multiplier * narrow), - "1024": int(16 * channel_multiplier * narrow), - } - self.channels = channels - - self.constant_input = ConstantInput(channels["4"], size=4) - self.style_conv1 = StyleConv( - channels["4"], - channels["4"], - kernel_size=3, - num_style_feat=num_style_feat, - demodulate=True, - sample_mode=None, - interpolation_mode=interpolation_mode, - ) - self.to_rgb1 = ToRGB( - channels["4"], - num_style_feat, - upsample=False, - interpolation_mode=interpolation_mode, - ) - - self.log_size = int(math.log(out_size, 2)) - self.num_layers = (self.log_size - 2) * 2 + 1 - self.num_latent = self.log_size * 2 - 2 - - self.style_convs = nn.ModuleList() - self.to_rgbs = nn.ModuleList() - self.noises = nn.Module() - - in_channels = channels["4"] - # noise - for layer_idx in range(self.num_layers): - resolution = 2 ** ((layer_idx + 5) // 2) - shape = [1, 1, resolution, resolution] - self.noises.register_buffer(f"noise{layer_idx}", torch.randn(*shape)) - # style convs and to_rgbs - for i in range(3, self.log_size + 1): - out_channels = channels[f"{2**i}"] - self.style_convs.append( - StyleConv( - in_channels, - out_channels, - kernel_size=3, - num_style_feat=num_style_feat, - demodulate=True, - sample_mode="upsample", - interpolation_mode=interpolation_mode, - ) - ) - self.style_convs.append( - StyleConv( - out_channels, - out_channels, - kernel_size=3, - num_style_feat=num_style_feat, - demodulate=True, - sample_mode=None, - interpolation_mode=interpolation_mode, - ) - ) - self.to_rgbs.append( - ToRGB( - out_channels, - num_style_feat, - upsample=True, - interpolation_mode=interpolation_mode, - ) - ) - in_channels = out_channels - - def make_noise(self): - """Make noise for noise injection.""" - device = self.constant_input.weight.device - noises = [torch.randn(1, 1, 4, 4, device=device)] - - for i in range(3, self.log_size + 1): - for _ in range(2): - noises.append(torch.randn(1, 1, 2**i, 2**i, device=device)) - - return noises - - def get_latent(self, x): - return self.style_mlp(x) - - def mean_latent(self, num_latent): - latent_in = torch.randn( - num_latent, self.num_style_feat, device=self.constant_input.weight.device - ) - latent = self.style_mlp(latent_in).mean(0, keepdim=True) - return latent - - def forward( - self, - styles, - input_is_latent=False, - noise=None, - randomize_noise=True, - truncation=1, - truncation_latent=None, - inject_index=None, - return_latents=False, - ): - """Forward function for StyleGAN2Generator. - Args: - styles (list[Tensor]): Sample codes of styles. - input_is_latent (bool): Whether input is latent style. - Default: False. - noise (Tensor | None): Input noise or None. Default: None. - randomize_noise (bool): Randomize noise, used when 'noise' is - False. Default: True. - truncation (float): TODO. Default: 1. - truncation_latent (Tensor | None): TODO. Default: None. - inject_index (int | None): The injection index for mixing noise. - Default: None. - return_latents (bool): Whether to return style latents. - Default: False. - """ - # style codes -> latents with Style MLP layer - if not input_is_latent: - styles = [self.style_mlp(s) for s in styles] - # noises - if noise is None: - if randomize_noise: - noise = [None] * self.num_layers # for each style conv layer - else: # use the stored noise - noise = [ - getattr(self.noises, f"noise{i}") for i in range(self.num_layers) - ] - # style truncation - if truncation < 1: - style_truncation = [] - for style in styles: - style_truncation.append( - truncation_latent + truncation * (style - truncation_latent) - ) - styles = style_truncation - # get style latent with injection - if len(styles) == 1: - inject_index = self.num_latent - - if styles[0].ndim < 3: - # repeat latent code for all the layers - latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - else: # used for encoder with different latent code for each layer - latent = styles[0] - elif len(styles) == 2: # mixing noises - if inject_index is None: - inject_index = random.randint(1, self.num_latent - 1) - latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - latent2 = ( - styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1) - ) - latent = torch.cat([latent1, latent2], 1) - - # main generation - out = self.constant_input(latent.shape[0]) - out = self.style_conv1(out, latent[:, 0], noise=noise[0]) - skip = self.to_rgb1(out, latent[:, 1]) - - i = 1 - for conv1, conv2, noise1, noise2, to_rgb in zip( - self.style_convs[::2], - self.style_convs[1::2], - noise[1::2], - noise[2::2], - self.to_rgbs, - ): - out = conv1(out, latent[:, i], noise=noise1) - out = conv2(out, latent[:, i + 1], noise=noise2) - skip = to_rgb(out, latent[:, i + 2], skip) - i += 2 - - image = skip - - if return_latents: - return image, latent - else: - return image, None - - -class ScaledLeakyReLU(nn.Module): - """Scaled LeakyReLU. - Args: - negative_slope (float): Negative slope. Default: 0.2. - """ - - def __init__(self, negative_slope=0.2): - super(ScaledLeakyReLU, self).__init__() - self.negative_slope = negative_slope - - def forward(self, x): - out = F.leaky_relu(x, negative_slope=self.negative_slope) - return out * math.sqrt(2) - - -class EqualConv2d(nn.Module): - """Equalized Linear as StyleGAN2. - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Size of the convolving kernel. - stride (int): Stride of the convolution. Default: 1 - padding (int): Zero-padding added to both sides of the input. - Default: 0. - bias (bool): If ``True``, adds a learnable bias to the output. - Default: ``True``. - bias_init_val (float): Bias initialized value. Default: 0. - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - stride=1, - padding=0, - bias=True, - bias_init_val=0, - ): - super(EqualConv2d, self).__init__() - self.in_channels = in_channels - self.out_channels = out_channels - self.kernel_size = kernel_size - self.stride = stride - self.padding = padding - self.scale = 1 / math.sqrt(in_channels * kernel_size**2) - - self.weight = nn.Parameter( - torch.randn(out_channels, in_channels, kernel_size, kernel_size) - ) - if bias: - self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val)) - else: - self.register_parameter("bias", None) - - def forward(self, x): - out = F.conv2d( - x, - self.weight * self.scale, - bias=self.bias, - stride=self.stride, - padding=self.padding, - ) - - return out - - def __repr__(self): - return ( - f"{self.__class__.__name__}(in_channels={self.in_channels}, " - f"out_channels={self.out_channels}, " - f"kernel_size={self.kernel_size}," - f" stride={self.stride}, padding={self.padding}, " - f"bias={self.bias is not None})" - ) - - -class ConvLayer(nn.Sequential): - """Conv Layer used in StyleGAN2 Discriminator. - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Kernel size. - downsample (bool): Whether downsample by a factor of 2. - Default: False. - bias (bool): Whether with bias. Default: True. - activate (bool): Whether use activateion. Default: True. - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - downsample=False, - bias=True, - activate=True, - interpolation_mode="bilinear", - ): - layers = [] - self.interpolation_mode = interpolation_mode - # downsample - if downsample: - if self.interpolation_mode == "nearest": - self.align_corners = None - else: - self.align_corners = False - - layers.append( - torch.nn.Upsample( - scale_factor=0.5, - mode=interpolation_mode, - align_corners=self.align_corners, - ) - ) - stride = 1 - self.padding = kernel_size // 2 - # conv - layers.append( - EqualConv2d( - in_channels, - out_channels, - kernel_size, - stride=stride, - padding=self.padding, - bias=bias and not activate, - ) - ) - # activation - if activate: - if bias: - layers.append(FusedLeakyReLU(out_channels)) - else: - layers.append(ScaledLeakyReLU(0.2)) - - super(ConvLayer, self).__init__(*layers) - - -class ResBlock(nn.Module): - """Residual block used in StyleGAN2 Discriminator. - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - """ - - def __init__(self, in_channels, out_channels, interpolation_mode="bilinear"): - super(ResBlock, self).__init__() - - self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True) - self.conv2 = ConvLayer( - in_channels, - out_channels, - 3, - downsample=True, - interpolation_mode=interpolation_mode, - bias=True, - activate=True, - ) - self.skip = ConvLayer( - in_channels, - out_channels, - 1, - downsample=True, - interpolation_mode=interpolation_mode, - bias=False, - activate=False, - ) - - def forward(self, x): - out = self.conv1(x) - out = self.conv2(out) - skip = self.skip(x) - out = (out + skip) / math.sqrt(2) - return out diff --git a/comfy_extras/chainner_models/architecture/face/stylegan2_clean_arch.py b/comfy_extras/chainner_models/architecture/face/stylegan2_clean_arch.py deleted file mode 100644 index c48de9af6..000000000 --- a/comfy_extras/chainner_models/architecture/face/stylegan2_clean_arch.py +++ /dev/null @@ -1,453 +0,0 @@ -# pylint: skip-file -# type: ignore -import math - -import torch -from torch import nn -from torch.nn import functional as F -from torch.nn import init -from torch.nn.modules.batchnorm import _BatchNorm - - -@torch.no_grad() -def default_init_weights(module_list, scale=1, bias_fill=0, **kwargs): - """Initialize network weights. - Args: - module_list (list[nn.Module] | nn.Module): Modules to be initialized. - scale (float): Scale initialized weights, especially for residual - blocks. Default: 1. - bias_fill (float): The value to fill bias. Default: 0 - kwargs (dict): Other arguments for initialization function. - """ - if not isinstance(module_list, list): - module_list = [module_list] - for module in module_list: - for m in module.modules(): - if isinstance(m, nn.Conv2d): - init.kaiming_normal_(m.weight, **kwargs) - m.weight.data *= scale - if m.bias is not None: - m.bias.data.fill_(bias_fill) - elif isinstance(m, nn.Linear): - init.kaiming_normal_(m.weight, **kwargs) - m.weight.data *= scale - if m.bias is not None: - m.bias.data.fill_(bias_fill) - elif isinstance(m, _BatchNorm): - init.constant_(m.weight, 1) - if m.bias is not None: - m.bias.data.fill_(bias_fill) - - -class NormStyleCode(nn.Module): - def forward(self, x): - """Normalize the style codes. - Args: - x (Tensor): Style codes with shape (b, c). - Returns: - Tensor: Normalized tensor. - """ - return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8) - - -class ModulatedConv2d(nn.Module): - """Modulated Conv2d used in StyleGAN2. - There is no bias in ModulatedConv2d. - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Size of the convolving kernel. - num_style_feat (int): Channel number of style features. - demodulate (bool): Whether to demodulate in the conv layer. Default: True. - sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None. - eps (float): A value added to the denominator for numerical stability. Default: 1e-8. - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - num_style_feat, - demodulate=True, - sample_mode=None, - eps=1e-8, - ): - super(ModulatedConv2d, self).__init__() - self.in_channels = in_channels - self.out_channels = out_channels - self.kernel_size = kernel_size - self.demodulate = demodulate - self.sample_mode = sample_mode - self.eps = eps - - # modulation inside each modulated conv - self.modulation = nn.Linear(num_style_feat, in_channels, bias=True) - # initialization - default_init_weights( - self.modulation, - scale=1, - bias_fill=1, - a=0, - mode="fan_in", - nonlinearity="linear", - ) - - self.weight = nn.Parameter( - torch.randn(1, out_channels, in_channels, kernel_size, kernel_size) - / math.sqrt(in_channels * kernel_size**2) - ) - self.padding = kernel_size // 2 - - def forward(self, x, style): - """Forward function. - Args: - x (Tensor): Tensor with shape (b, c, h, w). - style (Tensor): Tensor with shape (b, num_style_feat). - Returns: - Tensor: Modulated tensor after convolution. - """ - b, c, h, w = x.shape # c = c_in - # weight modulation - style = self.modulation(style).view(b, 1, c, 1, 1) - # self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1) - weight = self.weight * style # (b, c_out, c_in, k, k) - - if self.demodulate: - demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps) - weight = weight * demod.view(b, self.out_channels, 1, 1, 1) - - weight = weight.view( - b * self.out_channels, c, self.kernel_size, self.kernel_size - ) - - # upsample or downsample if necessary - if self.sample_mode == "upsample": - x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=False) - elif self.sample_mode == "downsample": - x = F.interpolate(x, scale_factor=0.5, mode="bilinear", align_corners=False) - - b, c, h, w = x.shape - x = x.view(1, b * c, h, w) - # weight: (b*c_out, c_in, k, k), groups=b - out = F.conv2d(x, weight, padding=self.padding, groups=b) - out = out.view(b, self.out_channels, *out.shape[2:4]) - - return out - - def __repr__(self): - return ( - f"{self.__class__.__name__}(in_channels={self.in_channels}, out_channels={self.out_channels}, " - f"kernel_size={self.kernel_size}, demodulate={self.demodulate}, sample_mode={self.sample_mode})" - ) - - -class StyleConv(nn.Module): - """Style conv used in StyleGAN2. - Args: - in_channels (int): Channel number of the input. - out_channels (int): Channel number of the output. - kernel_size (int): Size of the convolving kernel. - num_style_feat (int): Channel number of style features. - demodulate (bool): Whether demodulate in the conv layer. Default: True. - sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None. - """ - - def __init__( - self, - in_channels, - out_channels, - kernel_size, - num_style_feat, - demodulate=True, - sample_mode=None, - ): - super(StyleConv, self).__init__() - self.modulated_conv = ModulatedConv2d( - in_channels, - out_channels, - kernel_size, - num_style_feat, - demodulate=demodulate, - sample_mode=sample_mode, - ) - self.weight = nn.Parameter(torch.zeros(1)) # for noise injection - self.bias = nn.Parameter(torch.zeros(1, out_channels, 1, 1)) - self.activate = nn.LeakyReLU(negative_slope=0.2, inplace=True) - - def forward(self, x, style, noise=None): - # modulate - out = self.modulated_conv(x, style) * 2**0.5 # for conversion - # noise injection - if noise is None: - b, _, h, w = out.shape - noise = out.new_empty(b, 1, h, w).normal_() - out = out + self.weight * noise - # add bias - out = out + self.bias - # activation - out = self.activate(out) - return out - - -class ToRGB(nn.Module): - """To RGB (image space) from features. - Args: - in_channels (int): Channel number of input. - num_style_feat (int): Channel number of style features. - upsample (bool): Whether to upsample. Default: True. - """ - - def __init__(self, in_channels, num_style_feat, upsample=True): - super(ToRGB, self).__init__() - self.upsample = upsample - self.modulated_conv = ModulatedConv2d( - in_channels, - 3, - kernel_size=1, - num_style_feat=num_style_feat, - demodulate=False, - sample_mode=None, - ) - self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1)) - - def forward(self, x, style, skip=None): - """Forward function. - Args: - x (Tensor): Feature tensor with shape (b, c, h, w). - style (Tensor): Tensor with shape (b, num_style_feat). - skip (Tensor): Base/skip tensor. Default: None. - Returns: - Tensor: RGB images. - """ - out = self.modulated_conv(x, style) - out = out + self.bias - if skip is not None: - if self.upsample: - skip = F.interpolate( - skip, scale_factor=2, mode="bilinear", align_corners=False - ) - out = out + skip - return out - - -class ConstantInput(nn.Module): - """Constant input. - Args: - num_channel (int): Channel number of constant input. - size (int): Spatial size of constant input. - """ - - def __init__(self, num_channel, size): - super(ConstantInput, self).__init__() - self.weight = nn.Parameter(torch.randn(1, num_channel, size, size)) - - def forward(self, batch): - out = self.weight.repeat(batch, 1, 1, 1) - return out - - -class StyleGAN2GeneratorClean(nn.Module): - """Clean version of StyleGAN2 Generator. - Args: - out_size (int): The spatial size of outputs. - num_style_feat (int): Channel number of style features. Default: 512. - num_mlp (int): Layer number of MLP style layers. Default: 8. - channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. - narrow (float): Narrow ratio for channels. Default: 1.0. - """ - - def __init__( - self, out_size, num_style_feat=512, num_mlp=8, channel_multiplier=2, narrow=1 - ): - super(StyleGAN2GeneratorClean, self).__init__() - # Style MLP layers - self.num_style_feat = num_style_feat - style_mlp_layers = [NormStyleCode()] - for i in range(num_mlp): - style_mlp_layers.extend( - [ - nn.Linear(num_style_feat, num_style_feat, bias=True), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - ] - ) - self.style_mlp = nn.Sequential(*style_mlp_layers) - # initialization - default_init_weights( - self.style_mlp, - scale=1, - bias_fill=0, - a=0.2, - mode="fan_in", - nonlinearity="leaky_relu", - ) - - # channel list - channels = { - "4": int(512 * narrow), - "8": int(512 * narrow), - "16": int(512 * narrow), - "32": int(512 * narrow), - "64": int(256 * channel_multiplier * narrow), - "128": int(128 * channel_multiplier * narrow), - "256": int(64 * channel_multiplier * narrow), - "512": int(32 * channel_multiplier * narrow), - "1024": int(16 * channel_multiplier * narrow), - } - self.channels = channels - - self.constant_input = ConstantInput(channels["4"], size=4) - self.style_conv1 = StyleConv( - channels["4"], - channels["4"], - kernel_size=3, - num_style_feat=num_style_feat, - demodulate=True, - sample_mode=None, - ) - self.to_rgb1 = ToRGB(channels["4"], num_style_feat, upsample=False) - - self.log_size = int(math.log(out_size, 2)) - self.num_layers = (self.log_size - 2) * 2 + 1 - self.num_latent = self.log_size * 2 - 2 - - self.style_convs = nn.ModuleList() - self.to_rgbs = nn.ModuleList() - self.noises = nn.Module() - - in_channels = channels["4"] - # noise - for layer_idx in range(self.num_layers): - resolution = 2 ** ((layer_idx + 5) // 2) - shape = [1, 1, resolution, resolution] - self.noises.register_buffer(f"noise{layer_idx}", torch.randn(*shape)) - # style convs and to_rgbs - for i in range(3, self.log_size + 1): - out_channels = channels[f"{2**i}"] - self.style_convs.append( - StyleConv( - in_channels, - out_channels, - kernel_size=3, - num_style_feat=num_style_feat, - demodulate=True, - sample_mode="upsample", - ) - ) - self.style_convs.append( - StyleConv( - out_channels, - out_channels, - kernel_size=3, - num_style_feat=num_style_feat, - demodulate=True, - sample_mode=None, - ) - ) - self.to_rgbs.append(ToRGB(out_channels, num_style_feat, upsample=True)) - in_channels = out_channels - - def make_noise(self): - """Make noise for noise injection.""" - device = self.constant_input.weight.device - noises = [torch.randn(1, 1, 4, 4, device=device)] - - for i in range(3, self.log_size + 1): - for _ in range(2): - noises.append(torch.randn(1, 1, 2**i, 2**i, device=device)) - - return noises - - def get_latent(self, x): - return self.style_mlp(x) - - def mean_latent(self, num_latent): - latent_in = torch.randn( - num_latent, self.num_style_feat, device=self.constant_input.weight.device - ) - latent = self.style_mlp(latent_in).mean(0, keepdim=True) - return latent - - def forward( - self, - styles, - input_is_latent=False, - noise=None, - randomize_noise=True, - truncation=1, - truncation_latent=None, - inject_index=None, - return_latents=False, - ): - """Forward function for StyleGAN2GeneratorClean. - Args: - styles (list[Tensor]): Sample codes of styles. - input_is_latent (bool): Whether input is latent style. Default: False. - noise (Tensor | None): Input noise or None. Default: None. - randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True. - truncation (float): The truncation ratio. Default: 1. - truncation_latent (Tensor | None): The truncation latent tensor. Default: None. - inject_index (int | None): The injection index for mixing noise. Default: None. - return_latents (bool): Whether to return style latents. Default: False. - """ - # style codes -> latents with Style MLP layer - if not input_is_latent: - styles = [self.style_mlp(s) for s in styles] - # noises - if noise is None: - if randomize_noise: - noise = [None] * self.num_layers # for each style conv layer - else: # use the stored noise - noise = [ - getattr(self.noises, f"noise{i}") for i in range(self.num_layers) - ] - # style truncation - if truncation < 1: - style_truncation = [] - for style in styles: - style_truncation.append( - truncation_latent + truncation * (style - truncation_latent) - ) - styles = style_truncation - # get style latents with injection - if len(styles) == 1: - inject_index = self.num_latent - - if styles[0].ndim < 3: - # repeat latent code for all the layers - latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - else: # used for encoder with different latent code for each layer - latent = styles[0] - elif len(styles) == 2: # mixing noises - if inject_index is None: - inject_index = random.randint(1, self.num_latent - 1) - latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1) - latent2 = ( - styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1) - ) - latent = torch.cat([latent1, latent2], 1) - - # main generation - out = self.constant_input(latent.shape[0]) - out = self.style_conv1(out, latent[:, 0], noise=noise[0]) - skip = self.to_rgb1(out, latent[:, 1]) - - i = 1 - for conv1, conv2, noise1, noise2, to_rgb in zip( - self.style_convs[::2], - self.style_convs[1::2], - noise[1::2], - noise[2::2], - self.to_rgbs, - ): - out = conv1(out, latent[:, i], noise=noise1) - out = conv2(out, latent[:, i + 1], noise=noise2) - skip = to_rgb(out, latent[:, i + 2], skip) # feature back to the rgb space - i += 2 - - image = skip - - if return_latents: - return image, latent - else: - return image, None diff --git a/comfy_extras/chainner_models/architecture/face/upfirdn2d.py b/comfy_extras/chainner_models/architecture/face/upfirdn2d.py deleted file mode 100644 index 4ea454151..000000000 --- a/comfy_extras/chainner_models/architecture/face/upfirdn2d.py +++ /dev/null @@ -1,194 +0,0 @@ -# pylint: skip-file -# type: ignore -# modify from https://github.com/rosinality/stylegan2-pytorch/blob/master/op/upfirdn2d.py # noqa:E501 - -import os - -import torch -from torch.autograd import Function -from torch.nn import functional as F - -upfirdn2d_ext = None - - -class UpFirDn2dBackward(Function): - @staticmethod - def forward( - ctx, grad_output, kernel, grad_kernel, up, down, pad, g_pad, in_size, out_size - ): - up_x, up_y = up - down_x, down_y = down - g_pad_x0, g_pad_x1, g_pad_y0, g_pad_y1 = g_pad - - grad_output = grad_output.reshape(-1, out_size[0], out_size[1], 1) - - grad_input = upfirdn2d_ext.upfirdn2d( - grad_output, - grad_kernel, - down_x, - down_y, - up_x, - up_y, - g_pad_x0, - g_pad_x1, - g_pad_y0, - g_pad_y1, - ) - grad_input = grad_input.view(in_size[0], in_size[1], in_size[2], in_size[3]) - - ctx.save_for_backward(kernel) - - pad_x0, pad_x1, pad_y0, pad_y1 = pad - - ctx.up_x = up_x - ctx.up_y = up_y - ctx.down_x = down_x - ctx.down_y = down_y - ctx.pad_x0 = pad_x0 - ctx.pad_x1 = pad_x1 - ctx.pad_y0 = pad_y0 - ctx.pad_y1 = pad_y1 - ctx.in_size = in_size - ctx.out_size = out_size - - return grad_input - - @staticmethod - def backward(ctx, gradgrad_input): - (kernel,) = ctx.saved_tensors - - gradgrad_input = gradgrad_input.reshape(-1, ctx.in_size[2], ctx.in_size[3], 1) - - gradgrad_out = upfirdn2d_ext.upfirdn2d( - gradgrad_input, - kernel, - ctx.up_x, - ctx.up_y, - ctx.down_x, - ctx.down_y, - ctx.pad_x0, - ctx.pad_x1, - ctx.pad_y0, - ctx.pad_y1, - ) - # gradgrad_out = gradgrad_out.view(ctx.in_size[0], ctx.out_size[0], - # ctx.out_size[1], ctx.in_size[3]) - gradgrad_out = gradgrad_out.view( - ctx.in_size[0], ctx.in_size[1], ctx.out_size[0], ctx.out_size[1] - ) - - return gradgrad_out, None, None, None, None, None, None, None, None - - -class UpFirDn2d(Function): - @staticmethod - def forward(ctx, input, kernel, up, down, pad): - up_x, up_y = up - down_x, down_y = down - pad_x0, pad_x1, pad_y0, pad_y1 = pad - - kernel_h, kernel_w = kernel.shape - _, channel, in_h, in_w = input.shape - ctx.in_size = input.shape - - input = input.reshape(-1, in_h, in_w, 1) - - ctx.save_for_backward(kernel, torch.flip(kernel, [0, 1])) - - out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1 - out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1 - ctx.out_size = (out_h, out_w) - - ctx.up = (up_x, up_y) - ctx.down = (down_x, down_y) - ctx.pad = (pad_x0, pad_x1, pad_y0, pad_y1) - - g_pad_x0 = kernel_w - pad_x0 - 1 - g_pad_y0 = kernel_h - pad_y0 - 1 - g_pad_x1 = in_w * up_x - out_w * down_x + pad_x0 - up_x + 1 - g_pad_y1 = in_h * up_y - out_h * down_y + pad_y0 - up_y + 1 - - ctx.g_pad = (g_pad_x0, g_pad_x1, g_pad_y0, g_pad_y1) - - out = upfirdn2d_ext.upfirdn2d( - input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1 - ) - # out = out.view(major, out_h, out_w, minor) - out = out.view(-1, channel, out_h, out_w) - - return out - - @staticmethod - def backward(ctx, grad_output): - kernel, grad_kernel = ctx.saved_tensors - - grad_input = UpFirDn2dBackward.apply( - grad_output, - kernel, - grad_kernel, - ctx.up, - ctx.down, - ctx.pad, - ctx.g_pad, - ctx.in_size, - ctx.out_size, - ) - - return grad_input, None, None, None, None - - -def upfirdn2d(input, kernel, up=1, down=1, pad=(0, 0)): - if input.device.type == "cpu": - out = upfirdn2d_native( - input, kernel, up, up, down, down, pad[0], pad[1], pad[0], pad[1] - ) - else: - out = UpFirDn2d.apply( - input, kernel, (up, up), (down, down), (pad[0], pad[1], pad[0], pad[1]) - ) - - return out - - -def upfirdn2d_native( - input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1 -): - _, channel, in_h, in_w = input.shape - input = input.reshape(-1, in_h, in_w, 1) - - _, in_h, in_w, minor = input.shape - kernel_h, kernel_w = kernel.shape - - out = input.view(-1, in_h, 1, in_w, 1, minor) - out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1]) - out = out.view(-1, in_h * up_y, in_w * up_x, minor) - - out = F.pad( - out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)] - ) - out = out[ - :, - max(-pad_y0, 0) : out.shape[1] - max(-pad_y1, 0), - max(-pad_x0, 0) : out.shape[2] - max(-pad_x1, 0), - :, - ] - - out = out.permute(0, 3, 1, 2) - out = out.reshape( - [-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1] - ) - w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w) - out = F.conv2d(out, w) - out = out.reshape( - -1, - minor, - in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1, - in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1, - ) - out = out.permute(0, 2, 3, 1) - out = out[:, ::down_y, ::down_x, :] - - out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1 - out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1 - - return out.view(-1, channel, out_h, out_w) diff --git a/comfy_extras/chainner_models/architecture/timm/LICENSE b/comfy_extras/chainner_models/architecture/timm/LICENSE deleted file mode 100644 index b4e9438bd..000000000 --- a/comfy_extras/chainner_models/architecture/timm/LICENSE +++ /dev/null @@ -1,201 +0,0 @@ - Apache License - Version 2.0, January 2004 - http://www.apache.org/licenses/ - - TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION - - 1. Definitions. - - "License" shall mean the terms and conditions for use, reproduction, - and distribution as defined by Sections 1 through 9 of this document. - - "Licensor" shall mean the copyright owner or entity authorized by - the copyright owner that is granting the License. - - "Legal Entity" shall mean the union of the acting entity and all - other entities that control, are controlled by, or are under common - control with that entity. For the purposes of this definition, - "control" means (i) the power, direct or indirect, to cause the - direction or management of such entity, whether by contract or - otherwise, or (ii) ownership of fifty percent (50%) or more of the - outstanding shares, or (iii) beneficial ownership of such entity. - - "You" (or "Your") shall mean an individual or Legal Entity - exercising permissions granted by this License. - - "Source" form shall mean the preferred form for making modifications, - including but not limited to software source code, documentation - source, and configuration files. - - "Object" form shall mean any form resulting from mechanical - transformation or translation of a Source form, including but - not limited to compiled object code, generated documentation, - and conversions to other media types. - - "Work" shall mean the work of authorship, whether in Source or - Object form, made available under the License, as indicated by a - copyright notice that is included in or attached to the work - (an example is provided in the Appendix below). - - "Derivative Works" shall mean any work, whether in Source or Object - form, that is based on (or derived from) the Work and for which the - editorial revisions, annotations, elaborations, or other modifications - represent, as a whole, an original work of authorship. For the purposes - of this License, Derivative Works shall not include works that remain - separable from, or merely link (or bind by name) to the interfaces of, - the Work and Derivative Works thereof. - - "Contribution" shall mean any work of authorship, including - the original version of the Work and any modifications or additions - to that Work or Derivative Works thereof, that is intentionally - submitted to Licensor for inclusion in the Work by the copyright owner - or by an individual or Legal Entity authorized to submit on behalf of - the copyright owner. For the purposes of this definition, "submitted" - means any form of electronic, verbal, or written communication sent - to the Licensor or its representatives, including but not limited to - communication on electronic mailing lists, source code control systems, - and issue tracking systems that are managed by, or on behalf of, the - Licensor for the purpose of discussing and improving the Work, but - excluding communication that is conspicuously marked or otherwise - designated in writing by the copyright owner as "Not a Contribution." - - "Contributor" shall mean Licensor and any individual or Legal Entity - on behalf of whom a Contribution has been received by Licensor and - subsequently incorporated within the Work. - - 2. Grant of Copyright License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - copyright license to reproduce, prepare Derivative Works of, - publicly display, publicly perform, sublicense, and distribute the - Work and such Derivative Works in Source or Object form. - - 3. Grant of Patent License. Subject to the terms and conditions of - this License, each Contributor hereby grants to You a perpetual, - worldwide, non-exclusive, no-charge, royalty-free, irrevocable - (except as stated in this section) patent license to make, have made, - use, offer to sell, sell, import, and otherwise transfer the Work, - where such license applies only to those patent claims licensable - by such Contributor that are necessarily infringed by their - Contribution(s) alone or by combination of their Contribution(s) - with the Work to which such Contribution(s) was submitted. If You - institute patent litigation against any entity (including a - cross-claim or counterclaim in a lawsuit) alleging that the Work - or a Contribution incorporated within the Work constitutes direct - or contributory patent infringement, then any patent licenses - granted to You under this License for that Work shall terminate - as of the date such litigation is filed. - - 4. Redistribution. You may reproduce and distribute copies of the - Work or Derivative Works thereof in any medium, with or without - modifications, and in Source or Object form, provided that You - meet the following conditions: - - (a) You must give any other recipients of the Work or - Derivative Works a copy of this License; and - - (b) You must cause any modified files to carry prominent notices - stating that You changed the files; and - - (c) You must retain, in the Source form of any Derivative Works - that You distribute, all copyright, patent, trademark, and - attribution notices from the Source form of the Work, - excluding those notices that do not pertain to any part of - the Derivative Works; and - - (d) If the Work includes a "NOTICE" text file as part of its - distribution, then any Derivative Works that You distribute must - include a readable copy of the attribution notices contained - within such NOTICE file, excluding those notices that do not - pertain to any part of the Derivative Works, in at least one - of the following places: within a NOTICE text file distributed - as part of the Derivative Works; within the Source form or - documentation, if provided along with the Derivative Works; or, - within a display generated by the Derivative Works, if and - wherever such third-party notices normally appear. The contents - of the NOTICE file are for informational purposes only and - do not modify the License. You may add Your own attribution - notices within Derivative Works that You distribute, alongside - or as an addendum to the NOTICE text from the Work, provided - that such additional attribution notices cannot be construed - as modifying the License. - - You may add Your own copyright statement to Your modifications and - may provide additional or different license terms and conditions - for use, reproduction, or distribution of Your modifications, or - for any such Derivative Works as a whole, provided Your use, - reproduction, and distribution of the Work otherwise complies with - the conditions stated in this License. - - 5. Submission of Contributions. Unless You explicitly state otherwise, - any Contribution intentionally submitted for inclusion in the Work - by You to the Licensor shall be under the terms and conditions of - this License, without any additional terms or conditions. - Notwithstanding the above, nothing herein shall supersede or modify - the terms of any separate license agreement you may have executed - with Licensor regarding such Contributions. - - 6. Trademarks. This License does not grant permission to use the trade - names, trademarks, service marks, or product names of the Licensor, - except as required for reasonable and customary use in describing the - origin of the Work and reproducing the content of the NOTICE file. - - 7. Disclaimer of Warranty. Unless required by applicable law or - agreed to in writing, Licensor provides the Work (and each - Contributor provides its Contributions) on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or - implied, including, without limitation, any warranties or conditions - of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A - PARTICULAR PURPOSE. You are solely responsible for determining the - appropriateness of using or redistributing the Work and assume any - risks associated with Your exercise of permissions under this License. - - 8. Limitation of Liability. In no event and under no legal theory, - whether in tort (including negligence), contract, or otherwise, - unless required by applicable law (such as deliberate and grossly - negligent acts) or agreed to in writing, shall any Contributor be - liable to You for damages, including any direct, indirect, special, - incidental, or consequential damages of any character arising as a - result of this License or out of the use or inability to use the - Work (including but not limited to damages for loss of goodwill, - work stoppage, computer failure or malfunction, or any and all - other commercial damages or losses), even if such Contributor - has been advised of the possibility of such damages. - - 9. Accepting Warranty or Additional Liability. While redistributing - the Work or Derivative Works thereof, You may choose to offer, - and charge a fee for, acceptance of support, warranty, indemnity, - or other liability obligations and/or rights consistent with this - License. However, in accepting such obligations, You may act only - on Your own behalf and on Your sole responsibility, not on behalf - of any other Contributor, and only if You agree to indemnify, - defend, and hold each Contributor harmless for any liability - incurred by, or claims asserted against, such Contributor by reason - of your accepting any such warranty or additional liability. - - END OF TERMS AND CONDITIONS - - APPENDIX: How to apply the Apache License to your work. - - To apply the Apache License to your work, attach the following - boilerplate notice, with the fields enclosed by brackets "{}" - replaced with your own identifying information. (Don't include - the brackets!) The text should be enclosed in the appropriate - comment syntax for the file format. We also recommend that a - file or class name and description of purpose be included on the - same "printed page" as the copyright notice for easier - identification within third-party archives. - - Copyright 2019 Ross Wightman - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. \ No newline at end of file diff --git a/comfy_extras/chainner_models/architecture/timm/drop.py b/comfy_extras/chainner_models/architecture/timm/drop.py deleted file mode 100644 index 14f0da914..000000000 --- a/comfy_extras/chainner_models/architecture/timm/drop.py +++ /dev/null @@ -1,223 +0,0 @@ -""" DropBlock, DropPath - -PyTorch implementations of DropBlock and DropPath (Stochastic Depth) regularization layers. - -Papers: -DropBlock: A regularization method for convolutional networks (https://arxiv.org/abs/1810.12890) - -Deep Networks with Stochastic Depth (https://arxiv.org/abs/1603.09382) - -Code: -DropBlock impl inspired by two Tensorflow impl that I liked: - - https://github.com/tensorflow/tpu/blob/master/models/official/resnet/resnet_model.py#L74 - - https://github.com/clovaai/assembled-cnn/blob/master/nets/blocks.py - -Hacked together by / Copyright 2020 Ross Wightman -""" -import torch -import torch.nn as nn -import torch.nn.functional as F - - -def drop_block_2d( - x, - drop_prob: float = 0.1, - block_size: int = 7, - gamma_scale: float = 1.0, - with_noise: bool = False, - inplace: bool = False, - batchwise: bool = False, -): - """DropBlock. See https://arxiv.org/pdf/1810.12890.pdf - - DropBlock with an experimental gaussian noise option. This layer has been tested on a few training - runs with success, but needs further validation and possibly optimization for lower runtime impact. - """ - _, C, H, W = x.shape - total_size = W * H - clipped_block_size = min(block_size, min(W, H)) - # seed_drop_rate, the gamma parameter - gamma = ( - gamma_scale - * drop_prob - * total_size - / clipped_block_size**2 - / ((W - block_size + 1) * (H - block_size + 1)) - ) - - # Forces the block to be inside the feature map. - w_i, h_i = torch.meshgrid( - torch.arange(W).to(x.device), torch.arange(H).to(x.device) - ) - valid_block = ( - (w_i >= clipped_block_size // 2) & (w_i < W - (clipped_block_size - 1) // 2) - ) & ((h_i >= clipped_block_size // 2) & (h_i < H - (clipped_block_size - 1) // 2)) - valid_block = torch.reshape(valid_block, (1, 1, H, W)).to(dtype=x.dtype) - - if batchwise: - # one mask for whole batch, quite a bit faster - uniform_noise = torch.rand((1, C, H, W), dtype=x.dtype, device=x.device) - else: - uniform_noise = torch.rand_like(x) - block_mask = ((2 - gamma - valid_block + uniform_noise) >= 1).to(dtype=x.dtype) - block_mask = -F.max_pool2d( - -block_mask, - kernel_size=clipped_block_size, # block_size, - stride=1, - padding=clipped_block_size // 2, - ) - - if with_noise: - normal_noise = ( - torch.randn((1, C, H, W), dtype=x.dtype, device=x.device) - if batchwise - else torch.randn_like(x) - ) - if inplace: - x.mul_(block_mask).add_(normal_noise * (1 - block_mask)) - else: - x = x * block_mask + normal_noise * (1 - block_mask) - else: - normalize_scale = ( - block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-7) - ).to(x.dtype) - if inplace: - x.mul_(block_mask * normalize_scale) - else: - x = x * block_mask * normalize_scale - return x - - -def drop_block_fast_2d( - x: torch.Tensor, - drop_prob: float = 0.1, - block_size: int = 7, - gamma_scale: float = 1.0, - with_noise: bool = False, - inplace: bool = False, -): - """DropBlock. See https://arxiv.org/pdf/1810.12890.pdf - - DropBlock with an experimental gaussian noise option. Simplied from above without concern for valid - block mask at edges. - """ - _, _, H, W = x.shape - total_size = W * H - clipped_block_size = min(block_size, min(W, H)) - gamma = ( - gamma_scale - * drop_prob - * total_size - / clipped_block_size**2 - / ((W - block_size + 1) * (H - block_size + 1)) - ) - - block_mask = torch.empty_like(x).bernoulli_(gamma) - block_mask = F.max_pool2d( - block_mask.to(x.dtype), - kernel_size=clipped_block_size, - stride=1, - padding=clipped_block_size // 2, - ) - - if with_noise: - normal_noise = torch.empty_like(x).normal_() - if inplace: - x.mul_(1.0 - block_mask).add_(normal_noise * block_mask) - else: - x = x * (1.0 - block_mask) + normal_noise * block_mask - else: - block_mask = 1 - block_mask - normalize_scale = ( - block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-6) - ).to(dtype=x.dtype) - if inplace: - x.mul_(block_mask * normalize_scale) - else: - x = x * block_mask * normalize_scale - return x - - -class DropBlock2d(nn.Module): - """DropBlock. See https://arxiv.org/pdf/1810.12890.pdf""" - - def __init__( - self, - drop_prob: float = 0.1, - block_size: int = 7, - gamma_scale: float = 1.0, - with_noise: bool = False, - inplace: bool = False, - batchwise: bool = False, - fast: bool = True, - ): - super(DropBlock2d, self).__init__() - self.drop_prob = drop_prob - self.gamma_scale = gamma_scale - self.block_size = block_size - self.with_noise = with_noise - self.inplace = inplace - self.batchwise = batchwise - self.fast = fast # FIXME finish comparisons of fast vs not - - def forward(self, x): - if not self.training or not self.drop_prob: - return x - if self.fast: - return drop_block_fast_2d( - x, - self.drop_prob, - self.block_size, - self.gamma_scale, - self.with_noise, - self.inplace, - ) - else: - return drop_block_2d( - x, - self.drop_prob, - self.block_size, - self.gamma_scale, - self.with_noise, - self.inplace, - self.batchwise, - ) - - -def drop_path( - x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True -): - """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). - - This is the same as the DropConnect impl I created for EfficientNet, etc networks, however, - the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper... - See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for - changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use - 'survival rate' as the argument. - - """ - if drop_prob == 0.0 or not training: - return x - keep_prob = 1 - drop_prob - shape = (x.shape[0],) + (1,) * ( - x.ndim - 1 - ) # work with diff dim tensors, not just 2D ConvNets - random_tensor = x.new_empty(shape).bernoulli_(keep_prob) - if keep_prob > 0.0 and scale_by_keep: - random_tensor.div_(keep_prob) - return x * random_tensor - - -class DropPath(nn.Module): - """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).""" - - def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True): - super(DropPath, self).__init__() - self.drop_prob = drop_prob - self.scale_by_keep = scale_by_keep - - def forward(self, x): - return drop_path(x, self.drop_prob, self.training, self.scale_by_keep) - - def extra_repr(self): - return f"drop_prob={round(self.drop_prob,3):0.3f}" diff --git a/comfy_extras/chainner_models/architecture/timm/helpers.py b/comfy_extras/chainner_models/architecture/timm/helpers.py deleted file mode 100644 index cdafee070..000000000 --- a/comfy_extras/chainner_models/architecture/timm/helpers.py +++ /dev/null @@ -1,31 +0,0 @@ -""" Layer/Module Helpers -Hacked together by / Copyright 2020 Ross Wightman -""" -import collections.abc -from itertools import repeat - - -# From PyTorch internals -def _ntuple(n): - def parse(x): - if isinstance(x, collections.abc.Iterable) and not isinstance(x, str): - return x - return tuple(repeat(x, n)) - - return parse - - -to_1tuple = _ntuple(1) -to_2tuple = _ntuple(2) -to_3tuple = _ntuple(3) -to_4tuple = _ntuple(4) -to_ntuple = _ntuple - - -def make_divisible(v, divisor=8, min_value=None, round_limit=0.9): - min_value = min_value or divisor - new_v = max(min_value, int(v + divisor / 2) // divisor * divisor) - # Make sure that round down does not go down by more than 10%. - if new_v < round_limit * v: - new_v += divisor - return new_v diff --git a/comfy_extras/chainner_models/architecture/timm/weight_init.py b/comfy_extras/chainner_models/architecture/timm/weight_init.py deleted file mode 100644 index b01697746..000000000 --- a/comfy_extras/chainner_models/architecture/timm/weight_init.py +++ /dev/null @@ -1,128 +0,0 @@ -import math -import warnings - -import torch -from torch.nn.init import _calculate_fan_in_and_fan_out - - -def _no_grad_trunc_normal_(tensor, mean, std, a, b): - # Cut & paste from PyTorch official master until it's in a few official releases - RW - # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf - def norm_cdf(x): - # Computes standard normal cumulative distribution function - return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0 - - if (mean < a - 2 * std) or (mean > b + 2 * std): - warnings.warn( - "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. " - "The distribution of values may be incorrect.", - stacklevel=2, - ) - - with torch.no_grad(): - # Values are generated by using a truncated uniform distribution and - # then using the inverse CDF for the normal distribution. - # Get upper and lower cdf values - l = norm_cdf((a - mean) / std) - u = norm_cdf((b - mean) / std) - - # Uniformly fill tensor with values from [l, u], then translate to - # [2l-1, 2u-1]. - tensor.uniform_(2 * l - 1, 2 * u - 1) - - # Use inverse cdf transform for normal distribution to get truncated - # standard normal - tensor.erfinv_() - - # Transform to proper mean, std - tensor.mul_(std * math.sqrt(2.0)) - tensor.add_(mean) - - # Clamp to ensure it's in the proper range - tensor.clamp_(min=a, max=b) - return tensor - - -def trunc_normal_( - tensor: torch.Tensor, mean=0.0, std=1.0, a=-2.0, b=2.0 -) -> torch.Tensor: - r"""Fills the input Tensor with values drawn from a truncated - normal distribution. The values are effectively drawn from the - normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)` - with values outside :math:`[a, b]` redrawn until they are within - the bounds. The method used for generating the random values works - best when :math:`a \leq \text{mean} \leq b`. - - NOTE: this impl is similar to the PyTorch trunc_normal_, the bounds [a, b] are - applied while sampling the normal with mean/std applied, therefore a, b args - should be adjusted to match the range of mean, std args. - - Args: - tensor: an n-dimensional `torch.Tensor` - mean: the mean of the normal distribution - std: the standard deviation of the normal distribution - a: the minimum cutoff value - b: the maximum cutoff value - Examples: - >>> w = torch.empty(3, 5) - >>> nn.init.trunc_normal_(w) - """ - return _no_grad_trunc_normal_(tensor, mean, std, a, b) - - -def trunc_normal_tf_( - tensor: torch.Tensor, mean=0.0, std=1.0, a=-2.0, b=2.0 -) -> torch.Tensor: - r"""Fills the input Tensor with values drawn from a truncated - normal distribution. The values are effectively drawn from the - normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)` - with values outside :math:`[a, b]` redrawn until they are within - the bounds. The method used for generating the random values works - best when :math:`a \leq \text{mean} \leq b`. - - NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the - bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0 - and the result is subsquently scaled and shifted by the mean and std args. - - Args: - tensor: an n-dimensional `torch.Tensor` - mean: the mean of the normal distribution - std: the standard deviation of the normal distribution - a: the minimum cutoff value - b: the maximum cutoff value - Examples: - >>> w = torch.empty(3, 5) - >>> nn.init.trunc_normal_(w) - """ - _no_grad_trunc_normal_(tensor, 0, 1.0, a, b) - with torch.no_grad(): - tensor.mul_(std).add_(mean) - return tensor - - -def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"): - fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor) - if mode == "fan_in": - denom = fan_in - elif mode == "fan_out": - denom = fan_out - elif mode == "fan_avg": - denom = (fan_in + fan_out) / 2 - - variance = scale / denom # type: ignore - - if distribution == "truncated_normal": - # constant is stddev of standard normal truncated to (-2, 2) - trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978) - elif distribution == "normal": - tensor.normal_(std=math.sqrt(variance)) - elif distribution == "uniform": - bound = math.sqrt(3 * variance) - # pylint: disable=invalid-unary-operand-type - tensor.uniform_(-bound, bound) - else: - raise ValueError(f"invalid distribution {distribution}") - - -def lecun_normal_(tensor): - variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal") diff --git a/comfy_extras/chainner_models/model_loading.py b/comfy_extras/chainner_models/model_loading.py index e000871c1..d48bc238c 100644 --- a/comfy_extras/chainner_models/model_loading.py +++ b/comfy_extras/chainner_models/model_loading.py @@ -1,99 +1,5 @@ -import logging as logger +from spandrel import ModelLoader -from .architecture.DAT import DAT -from .architecture.face.codeformer import CodeFormer -from .architecture.face.gfpganv1_clean_arch import GFPGANv1Clean -from .architecture.face.restoreformer_arch import RestoreFormer -from .architecture.HAT import HAT -from .architecture.LaMa import LaMa -from .architecture.OmniSR.OmniSR import OmniSR -from .architecture.RRDB import RRDBNet as ESRGAN -from .architecture.SCUNet import SCUNet -from .architecture.SPSR import SPSRNet as SPSR -from .architecture.SRVGG import SRVGGNetCompact as RealESRGANv2 -from .architecture.SwiftSRGAN import Generator as SwiftSRGAN -from .architecture.Swin2SR import Swin2SR -from .architecture.SwinIR import SwinIR -from .types import PyTorchModel - - -class UnsupportedModel(Exception): - pass - - -def load_state_dict(state_dict) -> PyTorchModel: - logger.debug(f"Loading state dict into pytorch model arch") - - state_dict_keys = list(state_dict.keys()) - - if "params_ema" in state_dict_keys: - state_dict = state_dict["params_ema"] - elif "params-ema" in state_dict_keys: - state_dict = state_dict["params-ema"] - elif "params" in state_dict_keys: - state_dict = state_dict["params"] - - state_dict_keys = list(state_dict.keys()) - # SRVGGNet Real-ESRGAN (v2) - if "body.0.weight" in state_dict_keys and "body.1.weight" in state_dict_keys: - model = RealESRGANv2(state_dict) - # SPSR (ESRGAN with lots of extra layers) - elif "f_HR_conv1.0.weight" in state_dict: - model = SPSR(state_dict) - # Swift-SRGAN - elif ( - "model" in state_dict_keys - and "initial.cnn.depthwise.weight" in state_dict["model"].keys() - ): - model = SwiftSRGAN(state_dict) - # SwinIR, Swin2SR, HAT - elif "layers.0.residual_group.blocks.0.norm1.weight" in state_dict_keys: - if ( - "layers.0.residual_group.blocks.0.conv_block.cab.0.weight" - in state_dict_keys - ): - model = HAT(state_dict) - elif "patch_embed.proj.weight" in state_dict_keys: - model = Swin2SR(state_dict) - else: - model = SwinIR(state_dict) - # GFPGAN - elif ( - "toRGB.0.weight" in state_dict_keys - and "stylegan_decoder.style_mlp.1.weight" in state_dict_keys - ): - model = GFPGANv1Clean(state_dict) - # RestoreFormer - elif ( - "encoder.conv_in.weight" in state_dict_keys - and "encoder.down.0.block.0.norm1.weight" in state_dict_keys - ): - model = RestoreFormer(state_dict) - elif ( - "encoder.blocks.0.weight" in state_dict_keys - and "quantize.embedding.weight" in state_dict_keys - ): - model = CodeFormer(state_dict) - # LaMa - elif ( - "model.model.1.bn_l.running_mean" in state_dict_keys - or "generator.model.1.bn_l.running_mean" in state_dict_keys - ): - model = LaMa(state_dict) - # Omni-SR - elif "residual_layer.0.residual_layer.0.layer.0.fn.0.weight" in state_dict_keys: - model = OmniSR(state_dict) - # SCUNet - elif "m_head.0.weight" in state_dict_keys and "m_tail.0.weight" in state_dict_keys: - model = SCUNet(state_dict) - # DAT - elif "layers.0.blocks.2.attn.attn_mask_0" in state_dict_keys: - model = DAT(state_dict) - # Regular ESRGAN, "new-arch" ESRGAN, Real-ESRGAN v1 - else: - try: - model = ESRGAN(state_dict) - except: - # pylint: disable=raise-missing-from - raise UnsupportedModel - return model +def load_state_dict(state_dict): + print("WARNING: comfy_extras.chainner_models is deprecated and has been replaced by the spandrel library.") + return ModelLoader().load_from_state_dict(state_dict).eval() diff --git a/comfy_extras/chainner_models/types.py b/comfy_extras/chainner_models/types.py deleted file mode 100644 index 193333b9e..000000000 --- a/comfy_extras/chainner_models/types.py +++ /dev/null @@ -1,69 +0,0 @@ -from typing import Union - -from .architecture.DAT import DAT -from .architecture.face.codeformer import CodeFormer -from .architecture.face.gfpganv1_clean_arch import GFPGANv1Clean -from .architecture.face.restoreformer_arch import RestoreFormer -from .architecture.HAT import HAT -from .architecture.LaMa import LaMa -from .architecture.OmniSR.OmniSR import OmniSR -from .architecture.RRDB import RRDBNet as ESRGAN -from .architecture.SCUNet import SCUNet -from .architecture.SPSR import SPSRNet as SPSR -from .architecture.SRVGG import SRVGGNetCompact as RealESRGANv2 -from .architecture.SwiftSRGAN import Generator as SwiftSRGAN -from .architecture.Swin2SR import Swin2SR -from .architecture.SwinIR import SwinIR - -PyTorchSRModels = ( - RealESRGANv2, - SPSR, - SwiftSRGAN, - ESRGAN, - SwinIR, - Swin2SR, - HAT, - OmniSR, - SCUNet, - DAT, -) -PyTorchSRModel = Union[ - RealESRGANv2, - SPSR, - SwiftSRGAN, - ESRGAN, - SwinIR, - Swin2SR, - HAT, - OmniSR, - SCUNet, - DAT, -] - - -def is_pytorch_sr_model(model: object): - return isinstance(model, PyTorchSRModels) - - -PyTorchFaceModels = (GFPGANv1Clean, RestoreFormer, CodeFormer) -PyTorchFaceModel = Union[GFPGANv1Clean, RestoreFormer, CodeFormer] - - -def is_pytorch_face_model(model: object): - return isinstance(model, PyTorchFaceModels) - - -PyTorchInpaintModels = (LaMa,) -PyTorchInpaintModel = Union[LaMa] - - -def is_pytorch_inpaint_model(model: object): - return isinstance(model, PyTorchInpaintModels) - - -PyTorchModels = (*PyTorchSRModels, *PyTorchFaceModels, *PyTorchInpaintModels) -PyTorchModel = Union[PyTorchSRModel, PyTorchFaceModel, PyTorchInpaintModel] - - -def is_pytorch_model(model: object): - return isinstance(model, PyTorchModels) diff --git a/comfy_extras/nodes_advanced_samplers.py b/comfy_extras/nodes_advanced_samplers.py new file mode 100644 index 000000000..d973def81 --- /dev/null +++ b/comfy_extras/nodes_advanced_samplers.py @@ -0,0 +1,61 @@ +import comfy.samplers +import comfy.utils +import torch +import numpy as np +from tqdm.auto import trange, tqdm +import math + + +@torch.no_grad() +def sample_lcm_upscale(model, x, sigmas, extra_args=None, callback=None, disable=None, total_upscale=2.0, upscale_method="bislerp", upscale_steps=None): + extra_args = {} if extra_args is None else extra_args + + if upscale_steps is None: + upscale_steps = max(len(sigmas) // 2 + 1, 2) + else: + upscale_steps += 1 + upscale_steps = min(upscale_steps, len(sigmas) + 1) + + upscales = np.linspace(1.0, total_upscale, upscale_steps)[1:] + + orig_shape = x.size() + s_in = x.new_ones([x.shape[0]]) + for i in trange(len(sigmas) - 1, disable=disable): + denoised = model(x, sigmas[i] * s_in, **extra_args) + if callback is not None: + callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) + + x = denoised + if i < len(upscales): + x = comfy.utils.common_upscale(x, round(orig_shape[-1] * upscales[i]), round(orig_shape[-2] * upscales[i]), upscale_method, "disabled") + + if sigmas[i + 1] > 0: + x += sigmas[i + 1] * torch.randn_like(x) + return x + + +class SamplerLCMUpscale: + upscale_methods = ["bislerp", "nearest-exact", "bilinear", "area", "bicubic"] + + @classmethod + def INPUT_TYPES(s): + return {"required": + {"scale_ratio": ("FLOAT", {"default": 1.0, "min": 0.1, "max": 20.0, "step": 0.01}), + "scale_steps": ("INT", {"default": -1, "min": -1, "max": 1000, "step": 1}), + "upscale_method": (s.upscale_methods,), + } + } + RETURN_TYPES = ("SAMPLER",) + CATEGORY = "sampling/custom_sampling/samplers" + + FUNCTION = "get_sampler" + + def get_sampler(self, scale_ratio, scale_steps, upscale_method): + if scale_steps < 0: + scale_steps = None + sampler = comfy.samplers.KSAMPLER(sample_lcm_upscale, extra_options={"total_upscale": scale_ratio, "upscale_steps": scale_steps, "upscale_method": upscale_method}) + return (sampler, ) + +NODE_CLASS_MAPPINGS = { + "SamplerLCMUpscale": SamplerLCMUpscale, +} diff --git a/comfy_extras/nodes_align_your_steps.py b/comfy_extras/nodes_align_your_steps.py new file mode 100644 index 000000000..3ffe53187 --- /dev/null +++ b/comfy_extras/nodes_align_your_steps.py @@ -0,0 +1,53 @@ +#from: https://research.nvidia.com/labs/toronto-ai/AlignYourSteps/howto.html +import numpy as np +import torch + +def loglinear_interp(t_steps, num_steps): + """ + Performs log-linear interpolation of a given array of decreasing numbers. + """ + xs = np.linspace(0, 1, len(t_steps)) + ys = np.log(t_steps[::-1]) + + new_xs = np.linspace(0, 1, num_steps) + new_ys = np.interp(new_xs, xs, ys) + + interped_ys = np.exp(new_ys)[::-1].copy() + return interped_ys + +NOISE_LEVELS = {"SD1": [14.6146412293, 6.4745760956, 3.8636745985, 2.6946151520, 1.8841921177, 1.3943805092, 0.9642583904, 0.6523686016, 0.3977456272, 0.1515232662, 0.0291671582], + "SDXL":[14.6146412293, 6.3184485287, 3.7681790315, 2.1811480769, 1.3405244945, 0.8620721141, 0.5550693289, 0.3798540708, 0.2332364134, 0.1114188177, 0.0291671582], + "SVD": [700.00, 54.5, 15.886, 7.977, 4.248, 1.789, 0.981, 0.403, 0.173, 0.034, 0.002]} + +class AlignYourStepsScheduler: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"model_type": (["SD1", "SDXL", "SVD"], ), + "steps": ("INT", {"default": 10, "min": 10, "max": 10000}), + "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), + } + } + RETURN_TYPES = ("SIGMAS",) + CATEGORY = "sampling/custom_sampling/schedulers" + + FUNCTION = "get_sigmas" + + def get_sigmas(self, model_type, steps, denoise): + total_steps = steps + if denoise < 1.0: + if denoise <= 0.0: + return (torch.FloatTensor([]),) + total_steps = round(steps * denoise) + + sigmas = NOISE_LEVELS[model_type][:] + if (steps + 1) != len(sigmas): + sigmas = loglinear_interp(sigmas, steps + 1) + + sigmas = sigmas[-(total_steps + 1):] + sigmas[-1] = 0 + return (torch.FloatTensor(sigmas), ) + +NODE_CLASS_MAPPINGS = { + "AlignYourStepsScheduler": AlignYourStepsScheduler, +} diff --git a/comfy_extras/nodes_attention_multiply.py b/comfy_extras/nodes_attention_multiply.py new file mode 100644 index 000000000..4747eb395 --- /dev/null +++ b/comfy_extras/nodes_attention_multiply.py @@ -0,0 +1,120 @@ + +def attention_multiply(attn, model, q, k, v, out): + m = model.clone() + sd = model.model_state_dict() + + for key in sd: + if key.endswith("{}.to_q.bias".format(attn)) or key.endswith("{}.to_q.weight".format(attn)): + m.add_patches({key: (None,)}, 0.0, q) + if key.endswith("{}.to_k.bias".format(attn)) or key.endswith("{}.to_k.weight".format(attn)): + m.add_patches({key: (None,)}, 0.0, k) + if key.endswith("{}.to_v.bias".format(attn)) or key.endswith("{}.to_v.weight".format(attn)): + m.add_patches({key: (None,)}, 0.0, v) + if key.endswith("{}.to_out.0.bias".format(attn)) or key.endswith("{}.to_out.0.weight".format(attn)): + m.add_patches({key: (None,)}, 0.0, out) + + return m + + +class UNetSelfAttentionMultiply: + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "q": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "k": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "v": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "out": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + }} + RETURN_TYPES = ("MODEL",) + FUNCTION = "patch" + + CATEGORY = "_for_testing/attention_experiments" + + def patch(self, model, q, k, v, out): + m = attention_multiply("attn1", model, q, k, v, out) + return (m, ) + +class UNetCrossAttentionMultiply: + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "q": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "k": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "v": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "out": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + }} + RETURN_TYPES = ("MODEL",) + FUNCTION = "patch" + + CATEGORY = "_for_testing/attention_experiments" + + def patch(self, model, q, k, v, out): + m = attention_multiply("attn2", model, q, k, v, out) + return (m, ) + +class CLIPAttentionMultiply: + @classmethod + def INPUT_TYPES(s): + return {"required": { "clip": ("CLIP",), + "q": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "k": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "v": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "out": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + }} + RETURN_TYPES = ("CLIP",) + FUNCTION = "patch" + + CATEGORY = "_for_testing/attention_experiments" + + def patch(self, clip, q, k, v, out): + m = clip.clone() + sd = m.patcher.model_state_dict() + + for key in sd: + if key.endswith("self_attn.q_proj.weight") or key.endswith("self_attn.q_proj.bias"): + m.add_patches({key: (None,)}, 0.0, q) + if key.endswith("self_attn.k_proj.weight") or key.endswith("self_attn.k_proj.bias"): + m.add_patches({key: (None,)}, 0.0, k) + if key.endswith("self_attn.v_proj.weight") or key.endswith("self_attn.v_proj.bias"): + m.add_patches({key: (None,)}, 0.0, v) + if key.endswith("self_attn.out_proj.weight") or key.endswith("self_attn.out_proj.bias"): + m.add_patches({key: (None,)}, 0.0, out) + return (m, ) + +class UNetTemporalAttentionMultiply: + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "self_structural": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "self_temporal": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "cross_structural": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + "cross_temporal": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + }} + RETURN_TYPES = ("MODEL",) + FUNCTION = "patch" + + CATEGORY = "_for_testing/attention_experiments" + + def patch(self, model, self_structural, self_temporal, cross_structural, cross_temporal): + m = model.clone() + sd = model.model_state_dict() + + for k in sd: + if (k.endswith("attn1.to_out.0.bias") or k.endswith("attn1.to_out.0.weight")): + if '.time_stack.' in k: + m.add_patches({k: (None,)}, 0.0, self_temporal) + else: + m.add_patches({k: (None,)}, 0.0, self_structural) + elif (k.endswith("attn2.to_out.0.bias") or k.endswith("attn2.to_out.0.weight")): + if '.time_stack.' in k: + m.add_patches({k: (None,)}, 0.0, cross_temporal) + else: + m.add_patches({k: (None,)}, 0.0, cross_structural) + return (m, ) + +NODE_CLASS_MAPPINGS = { + "UNetSelfAttentionMultiply": UNetSelfAttentionMultiply, + "UNetCrossAttentionMultiply": UNetCrossAttentionMultiply, + "CLIPAttentionMultiply": CLIPAttentionMultiply, + "UNetTemporalAttentionMultiply": UNetTemporalAttentionMultiply, +} diff --git a/comfy_extras/nodes_canny.py b/comfy_extras/nodes_canny.py index 730dded5f..d85e6b856 100644 --- a/comfy_extras/nodes_canny.py +++ b/comfy_extras/nodes_canny.py @@ -1,280 +1,6 @@ -#From https://github.com/kornia/kornia -import math - -import torch -import torch.nn.functional as F +from kornia.filters import canny import comfy.model_management -def get_canny_nms_kernel(device=None, dtype=None): - """Utility function that returns 3x3 kernels for the Canny Non-maximal suppression.""" - return torch.tensor( - [ - [[[0.0, 0.0, 0.0], [0.0, 1.0, -1.0], [0.0, 0.0, 0.0]]], - [[[0.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, -1.0]]], - [[[0.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, -1.0, 0.0]]], - [[[0.0, 0.0, 0.0], [0.0, 1.0, 0.0], [-1.0, 0.0, 0.0]]], - [[[0.0, 0.0, 0.0], [-1.0, 1.0, 0.0], [0.0, 0.0, 0.0]]], - [[[-1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 0.0]]], - [[[0.0, -1.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 0.0]]], - [[[0.0, 0.0, -1.0], [0.0, 1.0, 0.0], [0.0, 0.0, 0.0]]], - ], - device=device, - dtype=dtype, - ) - - -def get_hysteresis_kernel(device=None, dtype=None): - """Utility function that returns the 3x3 kernels for the Canny hysteresis.""" - return torch.tensor( - [ - [[[0.0, 0.0, 0.0], [0.0, 0.0, 1.0], [0.0, 0.0, 0.0]]], - [[[0.0, 0.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 1.0]]], - [[[0.0, 0.0, 0.0], [0.0, 0.0, 0.0], [0.0, 1.0, 0.0]]], - [[[0.0, 0.0, 0.0], [0.0, 0.0, 0.0], [1.0, 0.0, 0.0]]], - [[[0.0, 0.0, 0.0], [1.0, 0.0, 0.0], [0.0, 0.0, 0.0]]], - [[[1.0, 0.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]], - [[[0.0, 1.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]], - [[[0.0, 0.0, 1.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]], - ], - device=device, - dtype=dtype, - ) - -def gaussian_blur_2d(img, kernel_size, sigma): - ksize_half = (kernel_size - 1) * 0.5 - - x = torch.linspace(-ksize_half, ksize_half, steps=kernel_size) - - pdf = torch.exp(-0.5 * (x / sigma).pow(2)) - - x_kernel = pdf / pdf.sum() - x_kernel = x_kernel.to(device=img.device, dtype=img.dtype) - - kernel2d = torch.mm(x_kernel[:, None], x_kernel[None, :]) - kernel2d = kernel2d.expand(img.shape[-3], 1, kernel2d.shape[0], kernel2d.shape[1]) - - padding = [kernel_size // 2, kernel_size // 2, kernel_size // 2, kernel_size // 2] - - img = torch.nn.functional.pad(img, padding, mode="reflect") - img = torch.nn.functional.conv2d(img, kernel2d, groups=img.shape[-3]) - - return img - -def get_sobel_kernel2d(device=None, dtype=None): - kernel_x = torch.tensor([[-1.0, 0.0, 1.0], [-2.0, 0.0, 2.0], [-1.0, 0.0, 1.0]], device=device, dtype=dtype) - kernel_y = kernel_x.transpose(0, 1) - return torch.stack([kernel_x, kernel_y]) - -def spatial_gradient(input, normalized: bool = True): - r"""Compute the first order image derivative in both x and y using a Sobel operator. - .. image:: _static/img/spatial_gradient.png - Args: - input: input image tensor with shape :math:`(B, C, H, W)`. - mode: derivatives modality, can be: `sobel` or `diff`. - order: the order of the derivatives. - normalized: whether the output is normalized. - Return: - the derivatives of the input feature map. with shape :math:`(B, C, 2, H, W)`. - .. note:: - See a working example `here `__. - Examples: - >>> input = torch.rand(1, 3, 4, 4) - >>> output = spatial_gradient(input) # 1x3x2x4x4 - >>> output.shape - torch.Size([1, 3, 2, 4, 4]) - """ - # KORNIA_CHECK_IS_TENSOR(input) - # KORNIA_CHECK_SHAPE(input, ['B', 'C', 'H', 'W']) - - # allocate kernel - kernel = get_sobel_kernel2d(device=input.device, dtype=input.dtype) - if normalized: - kernel = normalize_kernel2d(kernel) - - # prepare kernel - b, c, h, w = input.shape - tmp_kernel = kernel[:, None, ...] - - # Pad with "replicate for spatial dims, but with zeros for channel - spatial_pad = [kernel.size(1) // 2, kernel.size(1) // 2, kernel.size(2) // 2, kernel.size(2) // 2] - out_channels: int = 2 - padded_inp = torch.nn.functional.pad(input.reshape(b * c, 1, h, w), spatial_pad, 'replicate') - out = F.conv2d(padded_inp, tmp_kernel, groups=1, padding=0, stride=1) - return out.reshape(b, c, out_channels, h, w) - -def rgb_to_grayscale(image, rgb_weights = None): - r"""Convert a RGB image to grayscale version of image. - - .. image:: _static/img/rgb_to_grayscale.png - - The image data is assumed to be in the range of (0, 1). - - Args: - image: RGB image to be converted to grayscale with shape :math:`(*,3,H,W)`. - rgb_weights: Weights that will be applied on each channel (RGB). - The sum of the weights should add up to one. - Returns: - grayscale version of the image with shape :math:`(*,1,H,W)`. - - .. note:: - See a working example `here `__. - - Example: - >>> input = torch.rand(2, 3, 4, 5) - >>> gray = rgb_to_grayscale(input) # 2x1x4x5 - """ - - if len(image.shape) < 3 or image.shape[-3] != 3: - raise ValueError(f"Input size must have a shape of (*, 3, H, W). Got {image.shape}") - - if rgb_weights is None: - # 8 bit images - if image.dtype == torch.uint8: - rgb_weights = torch.tensor([76, 150, 29], device=image.device, dtype=torch.uint8) - # floating point images - elif image.dtype in (torch.float16, torch.float32, torch.float64): - rgb_weights = torch.tensor([0.299, 0.587, 0.114], device=image.device, dtype=image.dtype) - else: - raise TypeError(f"Unknown data type: {image.dtype}") - else: - # is tensor that we make sure is in the same device/dtype - rgb_weights = rgb_weights.to(image) - - # unpack the color image channels with RGB order - r: Tensor = image[..., 0:1, :, :] - g: Tensor = image[..., 1:2, :, :] - b: Tensor = image[..., 2:3, :, :] - - w_r, w_g, w_b = rgb_weights.unbind() - return w_r * r + w_g * g + w_b * b - -def canny( - input, - low_threshold = 0.1, - high_threshold = 0.2, - kernel_size = 5, - sigma = 1, - hysteresis = True, - eps = 1e-6, -): - r"""Find edges of the input image and filters them using the Canny algorithm. - .. image:: _static/img/canny.png - Args: - input: input image tensor with shape :math:`(B,C,H,W)`. - low_threshold: lower threshold for the hysteresis procedure. - high_threshold: upper threshold for the hysteresis procedure. - kernel_size: the size of the kernel for the gaussian blur. - sigma: the standard deviation of the kernel for the gaussian blur. - hysteresis: if True, applies the hysteresis edge tracking. - Otherwise, the edges are divided between weak (0.5) and strong (1) edges. - eps: regularization number to avoid NaN during backprop. - Returns: - - the canny edge magnitudes map, shape of :math:`(B,1,H,W)`. - - the canny edge detection filtered by thresholds and hysteresis, shape of :math:`(B,1,H,W)`. - .. note:: - See a working example `here `__. - Example: - >>> input = torch.rand(5, 3, 4, 4) - >>> magnitude, edges = canny(input) # 5x3x4x4 - >>> magnitude.shape - torch.Size([5, 1, 4, 4]) - >>> edges.shape - torch.Size([5, 1, 4, 4]) - """ - # KORNIA_CHECK_IS_TENSOR(input) - # KORNIA_CHECK_SHAPE(input, ['B', 'C', 'H', 'W']) - # KORNIA_CHECK( - # low_threshold <= high_threshold, - # "Invalid input thresholds. low_threshold should be smaller than the high_threshold. Got: " - # f"{low_threshold}>{high_threshold}", - # ) - # KORNIA_CHECK(0 < low_threshold < 1, f'Invalid low threshold. Should be in range (0, 1). Got: {low_threshold}') - # KORNIA_CHECK(0 < high_threshold < 1, f'Invalid high threshold. Should be in range (0, 1). Got: {high_threshold}') - - device = input.device - dtype = input.dtype - - # To Grayscale - if input.shape[1] == 3: - input = rgb_to_grayscale(input) - - # Gaussian filter - blurred: Tensor = gaussian_blur_2d(input, kernel_size, sigma) - - # Compute the gradients - gradients: Tensor = spatial_gradient(blurred, normalized=False) - - # Unpack the edges - gx: Tensor = gradients[:, :, 0] - gy: Tensor = gradients[:, :, 1] - - # Compute gradient magnitude and angle - magnitude: Tensor = torch.sqrt(gx * gx + gy * gy + eps) - angle: Tensor = torch.atan2(gy, gx) - - # Radians to Degrees - angle = 180.0 * angle / math.pi - - # Round angle to the nearest 45 degree - angle = torch.round(angle / 45) * 45 - - # Non-maximal suppression - nms_kernels: Tensor = get_canny_nms_kernel(device, dtype) - nms_magnitude: Tensor = F.conv2d(magnitude, nms_kernels, padding=nms_kernels.shape[-1] // 2) - - # Get the indices for both directions - positive_idx: Tensor = (angle / 45) % 8 - positive_idx = positive_idx.long() - - negative_idx: Tensor = ((angle / 45) + 4) % 8 - negative_idx = negative_idx.long() - - # Apply the non-maximum suppression to the different directions - channel_select_filtered_positive: Tensor = torch.gather(nms_magnitude, 1, positive_idx) - channel_select_filtered_negative: Tensor = torch.gather(nms_magnitude, 1, negative_idx) - - channel_select_filtered: Tensor = torch.stack( - [channel_select_filtered_positive, channel_select_filtered_negative], 1 - ) - - is_max: Tensor = channel_select_filtered.min(dim=1)[0] > 0.0 - - magnitude = magnitude * is_max - - # Threshold - edges: Tensor = F.threshold(magnitude, low_threshold, 0.0) - - low: Tensor = magnitude > low_threshold - high: Tensor = magnitude > high_threshold - - edges = low * 0.5 + high * 0.5 - edges = edges.to(dtype) - - # Hysteresis - if hysteresis: - edges_old: Tensor = -torch.ones(edges.shape, device=edges.device, dtype=dtype) - hysteresis_kernels: Tensor = get_hysteresis_kernel(device, dtype) - - while ((edges_old - edges).abs() != 0).any(): - weak: Tensor = (edges == 0.5).float() - strong: Tensor = (edges == 1).float() - - hysteresis_magnitude: Tensor = F.conv2d( - edges, hysteresis_kernels, padding=hysteresis_kernels.shape[-1] // 2 - ) - hysteresis_magnitude = (hysteresis_magnitude == 1).any(1, keepdim=True).to(dtype) - hysteresis_magnitude = hysteresis_magnitude * weak + strong - - edges_old = edges.clone() - edges = hysteresis_magnitude + (hysteresis_magnitude == 0) * weak * 0.5 - - edges = hysteresis_magnitude - - return magnitude, edges - class Canny: @classmethod diff --git a/comfy_extras/nodes_clip_sdxl.py b/comfy_extras/nodes_clip_sdxl.py index dcf8859fa..3087b917b 100644 --- a/comfy_extras/nodes_clip_sdxl.py +++ b/comfy_extras/nodes_clip_sdxl.py @@ -8,7 +8,7 @@ class CLIPTextEncodeSDXLRefiner: "ascore": ("FLOAT", {"default": 6.0, "min": 0.0, "max": 1000.0, "step": 0.01}), "width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), "height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), - "text": ("STRING", {"multiline": True}), "clip": ("CLIP", ), + "text": ("STRING", {"multiline": True, "dynamicPrompts": True}), "clip": ("CLIP", ), }} RETURN_TYPES = ("CONDITIONING",) FUNCTION = "encode" @@ -30,8 +30,8 @@ class CLIPTextEncodeSDXL: "crop_h": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION}), "target_width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), "target_height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), - "text_g": ("STRING", {"multiline": True, "default": "CLIP_G"}), "clip": ("CLIP", ), - "text_l": ("STRING", {"multiline": True, "default": "CLIP_L"}), "clip": ("CLIP", ), + "text_g": ("STRING", {"multiline": True, "dynamicPrompts": True}), "clip": ("CLIP", ), + "text_l": ("STRING", {"multiline": True, "dynamicPrompts": True}), "clip": ("CLIP", ), }} RETURN_TYPES = ("CONDITIONING",) FUNCTION = "encode" diff --git a/comfy_extras/nodes_cond.py b/comfy_extras/nodes_cond.py new file mode 100644 index 000000000..4c3a1d5bf --- /dev/null +++ b/comfy_extras/nodes_cond.py @@ -0,0 +1,25 @@ + + +class CLIPTextEncodeControlnet: + @classmethod + def INPUT_TYPES(s): + return {"required": {"clip": ("CLIP", ), "conditioning": ("CONDITIONING", ), "text": ("STRING", {"multiline": True, "dynamicPrompts": True})}} + RETURN_TYPES = ("CONDITIONING",) + FUNCTION = "encode" + + CATEGORY = "_for_testing/conditioning" + + def encode(self, clip, conditioning, text): + tokens = clip.tokenize(text) + cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True) + c = [] + for t in conditioning: + n = [t[0], t[1].copy()] + n[1]['cross_attn_controlnet'] = cond + n[1]['pooled_output_controlnet'] = pooled + c.append(n) + return (c, ) + +NODE_CLASS_MAPPINGS = { + "CLIPTextEncodeControlnet": CLIPTextEncodeControlnet +} diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index f02cb5ef7..47f08bf60 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -4,6 +4,7 @@ from comfy.k_diffusion import sampling as k_diffusion_sampling import latent_preview import torch import comfy.utils +import node_helpers class BasicScheduler: @@ -24,10 +25,11 @@ class BasicScheduler: def get_sigmas(self, model, scheduler, steps, denoise): total_steps = steps if denoise < 1.0: + if denoise <= 0.0: + return (torch.FloatTensor([]),) total_steps = int(steps/denoise) - comfy.model_management.load_models_gpu([model]) - sigmas = comfy.samplers.calculate_sigmas_scheduler(model.model, scheduler, total_steps).cpu() + sigmas = comfy.samplers.calculate_sigmas(model.get_model_object("model_sampling"), scheduler, total_steps).cpu() sigmas = sigmas[-(steps + 1):] return (sigmas, ) @@ -37,8 +39,8 @@ class KarrasScheduler: def INPUT_TYPES(s): return {"required": {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), + "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), + "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), "rho": ("FLOAT", {"default": 7.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), } } @@ -56,8 +58,8 @@ class ExponentialScheduler: def INPUT_TYPES(s): return {"required": {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), + "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), + "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), } } RETURN_TYPES = ("SIGMAS",) @@ -74,8 +76,8 @@ class PolyexponentialScheduler: def INPUT_TYPES(s): return {"required": {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), + "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), + "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), "rho": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), } } @@ -105,9 +107,7 @@ class SDTurboScheduler: def get_sigmas(self, model, steps, denoise): start_step = 10 - int(10 * denoise) timesteps = torch.flip(torch.arange(1, 11) * 100 - 1, (0,))[start_step:start_step + steps] - inner_model = model.patch_model(patch_weights=False) - sigmas = inner_model.model_sampling.sigma(timesteps) - model.unpatch_model() + sigmas = model.get_model_object("model_sampling").sigma(timesteps) sigmas = torch.cat([sigmas, sigmas.new_zeros([1])]) return (sigmas, ) @@ -116,8 +116,8 @@ class VPScheduler: def INPUT_TYPES(s): return {"required": {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "beta_d": ("FLOAT", {"default": 19.9, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), #TODO: fix default values - "beta_min": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), + "beta_d": ("FLOAT", {"default": 19.9, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), #TODO: fix default values + "beta_min": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), "eps_s": ("FLOAT", {"default": 0.001, "min": 0.0, "max": 1.0, "step":0.0001, "round": False}), } } @@ -139,6 +139,7 @@ class SplitSigmas: } } RETURN_TYPES = ("SIGMAS","SIGMAS") + RETURN_NAMES = ("high_sigmas", "low_sigmas") CATEGORY = "sampling/custom_sampling/sigmas" FUNCTION = "get_sigmas" @@ -148,6 +149,27 @@ class SplitSigmas: sigmas2 = sigmas[step:] return (sigmas1, sigmas2) +class SplitSigmasDenoise: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"sigmas": ("SIGMAS", ), + "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), + } + } + RETURN_TYPES = ("SIGMAS","SIGMAS") + RETURN_NAMES = ("high_sigmas", "low_sigmas") + CATEGORY = "sampling/custom_sampling/sigmas" + + FUNCTION = "get_sigmas" + + def get_sigmas(self, sigmas, denoise): + steps = max(sigmas.shape[-1] - 1, 0) + total_steps = round(steps * denoise) + sigmas1 = sigmas[:-(total_steps)] + sigmas2 = sigmas[-(total_steps + 1):] + return (sigmas1, sigmas2) + class FlipSigmas: @classmethod def INPUT_TYPES(s): @@ -161,6 +183,9 @@ class FlipSigmas: FUNCTION = "get_sigmas" def get_sigmas(self, sigmas): + if len(sigmas) == 0: + return (sigmas,) + sigmas = sigmas.flip(0) if sigmas[0] == 0: sigmas[0] = 0.0001 @@ -182,6 +207,28 @@ class KSamplerSelect: sampler = comfy.samplers.sampler_object(sampler_name) return (sampler, ) +class SamplerDPMPP_3M_SDE: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "noise_device": (['gpu', 'cpu'], ), + } + } + RETURN_TYPES = ("SAMPLER",) + CATEGORY = "sampling/custom_sampling/samplers" + + FUNCTION = "get_sampler" + + def get_sampler(self, eta, s_noise, noise_device): + if noise_device == 'cpu': + sampler_name = "dpmpp_3m_sde" + else: + sampler_name = "dpmpp_3m_sde_gpu" + sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise}) + return (sampler, ) + class SamplerDPMPP_2M_SDE: @classmethod def INPUT_TYPES(s): @@ -229,6 +276,84 @@ class SamplerDPMPP_SDE: sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "r": r}) return (sampler, ) +class SamplerEulerAncestral: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + } + } + RETURN_TYPES = ("SAMPLER",) + CATEGORY = "sampling/custom_sampling/samplers" + + FUNCTION = "get_sampler" + + def get_sampler(self, eta, s_noise): + sampler = comfy.samplers.ksampler("euler_ancestral", {"eta": eta, "s_noise": s_noise}) + return (sampler, ) + +class SamplerLMS: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"order": ("INT", {"default": 4, "min": 1, "max": 100}), + } + } + RETURN_TYPES = ("SAMPLER",) + CATEGORY = "sampling/custom_sampling/samplers" + + FUNCTION = "get_sampler" + + def get_sampler(self, order): + sampler = comfy.samplers.ksampler("lms", {"order": order}) + return (sampler, ) + +class SamplerDPMAdaptative: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"order": ("INT", {"default": 3, "min": 2, "max": 3}), + "rtol": ("FLOAT", {"default": 0.05, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "atol": ("FLOAT", {"default": 0.0078, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "h_init": ("FLOAT", {"default": 0.05, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "pcoeff": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "icoeff": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "dcoeff": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "accept_safety": ("FLOAT", {"default": 0.81, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "eta": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), + } + } + RETURN_TYPES = ("SAMPLER",) + CATEGORY = "sampling/custom_sampling/samplers" + + FUNCTION = "get_sampler" + + def get_sampler(self, order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise): + sampler = comfy.samplers.ksampler("dpm_adaptive", {"order": order, "rtol": rtol, "atol": atol, "h_init": h_init, "pcoeff": pcoeff, + "icoeff": icoeff, "dcoeff": dcoeff, "accept_safety": accept_safety, "eta": eta, + "s_noise":s_noise }) + return (sampler, ) + +class Noise_EmptyNoise: + def __init__(self): + self.seed = 0 + + def generate_noise(self, input_latent): + latent_image = input_latent["samples"] + return torch.zeros(latent_image.shape, dtype=latent_image.dtype, layout=latent_image.layout, device="cpu") + + +class Noise_RandomNoise: + def __init__(self, seed): + self.seed = seed + + def generate_noise(self, input_latent): + latent_image = input_latent["samples"] + batch_inds = input_latent["batch_index"] if "batch_index" in input_latent else None + return comfy.sample.prepare_noise(latent_image, self.seed, batch_inds) + class SamplerCustom: @classmethod def INPUT_TYPES(s): @@ -256,10 +381,9 @@ class SamplerCustom: latent = latent_image latent_image = latent["samples"] if not add_noise: - noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu") + noise = Noise_EmptyNoise().generate_noise(latent) else: - batch_inds = latent["batch_index"] if "batch_index" in latent else None - noise = comfy.sample.prepare_noise(latent_image, noise_seed, batch_inds) + noise = Noise_RandomNoise(noise_seed).generate_noise(latent) noise_mask = None if "noise_mask" in latent: @@ -280,6 +404,207 @@ class SamplerCustom: out_denoised = out return (out, out_denoised) +class Guider_Basic(comfy.samplers.CFGGuider): + def set_conds(self, positive): + self.inner_set_conds({"positive": positive}) + +class BasicGuider: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"model": ("MODEL",), + "conditioning": ("CONDITIONING", ), + } + } + + RETURN_TYPES = ("GUIDER",) + + FUNCTION = "get_guider" + CATEGORY = "sampling/custom_sampling/guiders" + + def get_guider(self, model, conditioning): + guider = Guider_Basic(model) + guider.set_conds(conditioning) + return (guider,) + +class CFGGuider: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"model": ("MODEL",), + "positive": ("CONDITIONING", ), + "negative": ("CONDITIONING", ), + "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), + } + } + + RETURN_TYPES = ("GUIDER",) + + FUNCTION = "get_guider" + CATEGORY = "sampling/custom_sampling/guiders" + + def get_guider(self, model, positive, negative, cfg): + guider = comfy.samplers.CFGGuider(model) + guider.set_conds(positive, negative) + guider.set_cfg(cfg) + return (guider,) + +class Guider_DualCFG(comfy.samplers.CFGGuider): + def set_cfg(self, cfg1, cfg2): + self.cfg1 = cfg1 + self.cfg2 = cfg2 + + def set_conds(self, positive, middle, negative): + middle = node_helpers.conditioning_set_values(middle, {"prompt_type": "negative"}) + self.inner_set_conds({"positive": positive, "middle": middle, "negative": negative}) + + def predict_noise(self, x, timestep, model_options={}, seed=None): + negative_cond = self.conds.get("negative", None) + middle_cond = self.conds.get("middle", None) + + out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, self.conds.get("positive", None)], x, timestep, model_options) + return comfy.samplers.cfg_function(self.inner_model, out[1], out[0], self.cfg2, x, timestep, model_options=model_options, cond=middle_cond, uncond=negative_cond) + (out[2] - out[1]) * self.cfg1 + +class DualCFGGuider: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"model": ("MODEL",), + "cond1": ("CONDITIONING", ), + "cond2": ("CONDITIONING", ), + "negative": ("CONDITIONING", ), + "cfg_conds": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), + "cfg_cond2_negative": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), + } + } + + RETURN_TYPES = ("GUIDER",) + + FUNCTION = "get_guider" + CATEGORY = "sampling/custom_sampling/guiders" + + def get_guider(self, model, cond1, cond2, negative, cfg_conds, cfg_cond2_negative): + guider = Guider_DualCFG(model) + guider.set_conds(cond1, cond2, negative) + guider.set_cfg(cfg_conds, cfg_cond2_negative) + return (guider,) + +class DisableNoise: + @classmethod + def INPUT_TYPES(s): + return {"required":{ + } + } + + RETURN_TYPES = ("NOISE",) + FUNCTION = "get_noise" + CATEGORY = "sampling/custom_sampling/noise" + + def get_noise(self): + return (Noise_EmptyNoise(),) + + +class RandomNoise(DisableNoise): + @classmethod + def INPUT_TYPES(s): + return {"required":{ + "noise_seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}), + } + } + + def get_noise(self, noise_seed): + return (Noise_RandomNoise(noise_seed),) + + +class SamplerCustomAdvanced: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"noise": ("NOISE", ), + "guider": ("GUIDER", ), + "sampler": ("SAMPLER", ), + "sigmas": ("SIGMAS", ), + "latent_image": ("LATENT", ), + } + } + + RETURN_TYPES = ("LATENT","LATENT") + RETURN_NAMES = ("output", "denoised_output") + + FUNCTION = "sample" + + CATEGORY = "sampling/custom_sampling" + + def sample(self, noise, guider, sampler, sigmas, latent_image): + latent = latent_image + latent_image = latent["samples"] + + noise_mask = None + if "noise_mask" in latent: + noise_mask = latent["noise_mask"] + + x0_output = {} + callback = latent_preview.prepare_callback(guider.model_patcher, sigmas.shape[-1] - 1, x0_output) + + disable_pbar = not comfy.utils.PROGRESS_BAR_ENABLED + samples = guider.sample(noise.generate_noise(latent), latent_image, sampler, sigmas, denoise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=noise.seed) + samples = samples.to(comfy.model_management.intermediate_device()) + + out = latent.copy() + out["samples"] = samples + if "x0" in x0_output: + out_denoised = latent.copy() + out_denoised["samples"] = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu()) + else: + out_denoised = out + return (out, out_denoised) + +class AddNoise: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"model": ("MODEL",), + "noise": ("NOISE", ), + "sigmas": ("SIGMAS", ), + "latent_image": ("LATENT", ), + } + } + + RETURN_TYPES = ("LATENT",) + + FUNCTION = "add_noise" + + CATEGORY = "_for_testing/custom_sampling/noise" + + def add_noise(self, model, noise, sigmas, latent_image): + if len(sigmas) == 0: + return latent_image + + latent = latent_image + latent_image = latent["samples"] + + noisy = noise.generate_noise(latent) + + model_sampling = model.get_model_object("model_sampling") + process_latent_out = model.get_model_object("process_latent_out") + process_latent_in = model.get_model_object("process_latent_in") + + if len(sigmas) > 1: + scale = torch.abs(sigmas[0] - sigmas[-1]) + else: + scale = sigmas[0] + + if torch.count_nonzero(latent_image) > 0: #Don't shift the empty latent image. + latent_image = process_latent_in(latent_image) + noisy = model_sampling.noise_scaling(scale, noisy, latent_image) + noisy = process_latent_out(noisy) + noisy = torch.nan_to_num(noisy, nan=0.0, posinf=0.0, neginf=0.0) + + out = latent.copy() + out["samples"] = noisy + return (out,) + + NODE_CLASS_MAPPINGS = { "SamplerCustom": SamplerCustom, "BasicScheduler": BasicScheduler, @@ -289,8 +614,21 @@ NODE_CLASS_MAPPINGS = { "VPScheduler": VPScheduler, "SDTurboScheduler": SDTurboScheduler, "KSamplerSelect": KSamplerSelect, + "SamplerEulerAncestral": SamplerEulerAncestral, + "SamplerLMS": SamplerLMS, + "SamplerDPMPP_3M_SDE": SamplerDPMPP_3M_SDE, "SamplerDPMPP_2M_SDE": SamplerDPMPP_2M_SDE, "SamplerDPMPP_SDE": SamplerDPMPP_SDE, + "SamplerDPMAdaptative": SamplerDPMAdaptative, "SplitSigmas": SplitSigmas, + "SplitSigmasDenoise": SplitSigmasDenoise, "FlipSigmas": FlipSigmas, + + "CFGGuider": CFGGuider, + "DualCFGGuider": DualCFGGuider, + "BasicGuider": BasicGuider, + "RandomNoise": RandomNoise, + "DisableNoise": DisableNoise, + "AddNoise": AddNoise, + "SamplerCustomAdvanced": SamplerCustomAdvanced, } diff --git a/comfy_extras/nodes_differential_diffusion.py b/comfy_extras/nodes_differential_diffusion.py new file mode 100644 index 000000000..98dbbf102 --- /dev/null +++ b/comfy_extras/nodes_differential_diffusion.py @@ -0,0 +1,42 @@ +# code adapted from https://github.com/exx8/differential-diffusion + +import torch + +class DifferentialDiffusion(): + @classmethod + def INPUT_TYPES(s): + return {"required": {"model": ("MODEL", ), + }} + RETURN_TYPES = ("MODEL",) + FUNCTION = "apply" + CATEGORY = "_for_testing" + INIT = False + + def apply(self, model): + model = model.clone() + model.set_model_denoise_mask_function(self.forward) + return (model,) + + def forward(self, sigma: torch.Tensor, denoise_mask: torch.Tensor, extra_options: dict): + model = extra_options["model"] + step_sigmas = extra_options["sigmas"] + sigma_to = model.inner_model.model_sampling.sigma_min + if step_sigmas[-1] > sigma_to: + sigma_to = step_sigmas[-1] + sigma_from = step_sigmas[0] + + ts_from = model.inner_model.model_sampling.timestep(sigma_from) + ts_to = model.inner_model.model_sampling.timestep(sigma_to) + current_ts = model.inner_model.model_sampling.timestep(sigma[0]) + + threshold = (current_ts - ts_to) / (ts_from - ts_to) + + return (denoise_mask >= threshold).to(denoise_mask.dtype) + + +NODE_CLASS_MAPPINGS = { + "DifferentialDiffusion": DifferentialDiffusion, +} +NODE_DISPLAY_NAME_MAPPINGS = { + "DifferentialDiffusion": "Differential Diffusion", +} diff --git a/comfy_extras/nodes_freelunch.py b/comfy_extras/nodes_freelunch.py index 7512b841d..c5ebcf26f 100644 --- a/comfy_extras/nodes_freelunch.py +++ b/comfy_extras/nodes_freelunch.py @@ -1,7 +1,7 @@ #code originally taken from: https://github.com/ChenyangSi/FreeU (under MIT License) import torch - +import logging def Fourier_filter(x, threshold, scale): # FFT @@ -34,7 +34,7 @@ class FreeU: RETURN_TYPES = ("MODEL",) FUNCTION = "patch" - CATEGORY = "_for_testing" + CATEGORY = "model_patches" def patch(self, model, b1, b2, s1, s2): model_channels = model.model.model_config.unet_config["model_channels"] @@ -42,14 +42,14 @@ class FreeU: on_cpu_devices = {} def output_block_patch(h, hsp, transformer_options): - scale = scale_dict.get(h.shape[1], None) + scale = scale_dict.get(int(h.shape[1]), None) if scale is not None: h[:,:h.shape[1] // 2] = h[:,:h.shape[1] // 2] * scale[0] if hsp.device not in on_cpu_devices: try: hsp = Fourier_filter(hsp, threshold=1, scale=scale[1]) except: - print("Device", hsp.device, "does not support the torch.fft functions used in the FreeU node, switching to CPU.") + logging.warning("Device {} does not support the torch.fft functions used in the FreeU node, switching to CPU.".format(hsp.device)) on_cpu_devices[hsp.device] = True hsp = Fourier_filter(hsp.cpu(), threshold=1, scale=scale[1]).to(hsp.device) else: @@ -73,7 +73,7 @@ class FreeU_V2: RETURN_TYPES = ("MODEL",) FUNCTION = "patch" - CATEGORY = "_for_testing" + CATEGORY = "model_patches" def patch(self, model, b1, b2, s1, s2): model_channels = model.model.model_config.unet_config["model_channels"] @@ -81,7 +81,7 @@ class FreeU_V2: on_cpu_devices = {} def output_block_patch(h, hsp, transformer_options): - scale = scale_dict.get(h.shape[1], None) + scale = scale_dict.get(int(h.shape[1]), None) if scale is not None: hidden_mean = h.mean(1).unsqueeze(1) B = hidden_mean.shape[0] @@ -95,7 +95,7 @@ class FreeU_V2: try: hsp = Fourier_filter(hsp, threshold=1, scale=scale[1]) except: - print("Device", hsp.device, "does not support the torch.fft functions used in the FreeU node, switching to CPU.") + logging.warning("Device {} does not support the torch.fft functions used in the FreeU node, switching to CPU.".format(hsp.device)) on_cpu_devices[hsp.device] = True hsp = Fourier_filter(hsp.cpu(), threshold=1, scale=scale[1]).to(hsp.device) else: diff --git a/comfy_extras/nodes_hypernetwork.py b/comfy_extras/nodes_hypernetwork.py index f692945a8..cafafa6ab 100644 --- a/comfy_extras/nodes_hypernetwork.py +++ b/comfy_extras/nodes_hypernetwork.py @@ -1,6 +1,7 @@ import comfy.utils import folder_paths import torch +import logging def load_hypernetwork_patch(path, strength): sd = comfy.utils.load_torch_file(path, safe_load=True) @@ -23,7 +24,7 @@ def load_hypernetwork_patch(path, strength): } if activation_func not in valid_activation: - print("Unsupported Hypernetwork format, if you report it I might implement it.", path, " ", activation_func, is_layer_norm, use_dropout, activate_output, last_layer_dropout) + logging.error("Unsupported Hypernetwork format, if you report it I might implement it. {} {} {} {} {} {}".format(path, activation_func, is_layer_norm, use_dropout, activate_output, last_layer_dropout)) return None out = {} diff --git a/comfy_extras/nodes_hypertile.py b/comfy_extras/nodes_hypertile.py index 544ae1b22..ae55d23dd 100644 --- a/comfy_extras/nodes_hypertile.py +++ b/comfy_extras/nodes_hypertile.py @@ -32,7 +32,7 @@ class HyperTile: RETURN_TYPES = ("MODEL",) FUNCTION = "patch" - CATEGORY = "_for_testing" + CATEGORY = "model_patches" def patch(self, model, tile_size, swap_size, max_depth, scale_depth): model_channels = model.model.model_config.unet_config["model_channels"] diff --git a/comfy_extras/nodes_images.py b/comfy_extras/nodes_images.py index aa80f5269..af37666b2 100644 --- a/comfy_extras/nodes_images.py +++ b/comfy_extras/nodes_images.py @@ -37,7 +37,7 @@ class RepeatImageBatch: @classmethod def INPUT_TYPES(s): return {"required": { "image": ("IMAGE",), - "amount": ("INT", {"default": 1, "min": 1, "max": 64}), + "amount": ("INT", {"default": 1, "min": 1, "max": 4096}), }} RETURN_TYPES = ("IMAGE",) FUNCTION = "repeat" @@ -48,6 +48,25 @@ class RepeatImageBatch: s = image.repeat((amount, 1,1,1)) return (s,) +class ImageFromBatch: + @classmethod + def INPUT_TYPES(s): + return {"required": { "image": ("IMAGE",), + "batch_index": ("INT", {"default": 0, "min": 0, "max": 4095}), + "length": ("INT", {"default": 1, "min": 1, "max": 4096}), + }} + RETURN_TYPES = ("IMAGE",) + FUNCTION = "frombatch" + + CATEGORY = "image/batch" + + def frombatch(self, image, batch_index, length): + s_in = image + batch_index = min(s_in.shape[0] - 1, batch_index) + length = min(s_in.shape[0] - batch_index, length) + s = s_in[batch_index:batch_index + length].clone() + return (s,) + class SaveAnimatedWEBP: def __init__(self): self.output_dir = folder_paths.get_output_directory() @@ -170,6 +189,7 @@ class SaveAnimatedPNG: NODE_CLASS_MAPPINGS = { "ImageCrop": ImageCrop, "RepeatImageBatch": RepeatImageBatch, + "ImageFromBatch": ImageFromBatch, "SaveAnimatedWEBP": SaveAnimatedWEBP, "SaveAnimatedPNG": SaveAnimatedPNG, } diff --git a/comfy_extras/nodes_ip2p.py b/comfy_extras/nodes_ip2p.py new file mode 100644 index 000000000..c2e70a84c --- /dev/null +++ b/comfy_extras/nodes_ip2p.py @@ -0,0 +1,45 @@ +import torch + +class InstructPixToPixConditioning: + @classmethod + def INPUT_TYPES(s): + return {"required": {"positive": ("CONDITIONING", ), + "negative": ("CONDITIONING", ), + "vae": ("VAE", ), + "pixels": ("IMAGE", ), + }} + + RETURN_TYPES = ("CONDITIONING","CONDITIONING","LATENT") + RETURN_NAMES = ("positive", "negative", "latent") + FUNCTION = "encode" + + CATEGORY = "conditioning/instructpix2pix" + + def encode(self, positive, negative, pixels, vae): + x = (pixels.shape[1] // 8) * 8 + y = (pixels.shape[2] // 8) * 8 + + if pixels.shape[1] != x or pixels.shape[2] != y: + x_offset = (pixels.shape[1] % 8) // 2 + y_offset = (pixels.shape[2] % 8) // 2 + pixels = pixels[:,x_offset:x + x_offset, y_offset:y + y_offset,:] + + concat_latent = vae.encode(pixels) + + out_latent = {} + out_latent["samples"] = torch.zeros_like(concat_latent) + + out = [] + for conditioning in [positive, negative]: + c = [] + for t in conditioning: + d = t[1].copy() + d["concat_latent_image"] = concat_latent + n = [t[0], d] + c.append(n) + out.append(c) + return (out[0], out[1], out_latent) + +NODE_CLASS_MAPPINGS = { + "InstructPixToPixConditioning": InstructPixToPixConditioning, +} diff --git a/comfy_extras/nodes_latent.py b/comfy_extras/nodes_latent.py index 2eefc4c55..eabae0885 100644 --- a/comfy_extras/nodes_latent.py +++ b/comfy_extras/nodes_latent.py @@ -122,10 +122,34 @@ class LatentBatch: samples_out["batch_index"] = samples1.get("batch_index", [x for x in range(0, s1.shape[0])]) + samples2.get("batch_index", [x for x in range(0, s2.shape[0])]) return (samples_out,) +class LatentBatchSeedBehavior: + @classmethod + def INPUT_TYPES(s): + return {"required": { "samples": ("LATENT",), + "seed_behavior": (["random", "fixed"],{"default": "fixed"}),}} + + RETURN_TYPES = ("LATENT",) + FUNCTION = "op" + + CATEGORY = "latent/advanced" + + def op(self, samples, seed_behavior): + samples_out = samples.copy() + latent = samples["samples"] + if seed_behavior == "random": + if 'batch_index' in samples_out: + samples_out.pop('batch_index') + elif seed_behavior == "fixed": + batch_number = samples_out.get("batch_index", [0])[0] + samples_out["batch_index"] = [batch_number] * latent.shape[0] + + return (samples_out,) + NODE_CLASS_MAPPINGS = { "LatentAdd": LatentAdd, "LatentSubtract": LatentSubtract, "LatentMultiply": LatentMultiply, "LatentInterpolate": LatentInterpolate, "LatentBatch": LatentBatch, + "LatentBatchSeedBehavior": LatentBatchSeedBehavior, } diff --git a/comfy_extras/nodes_mask.py b/comfy_extras/nodes_mask.py index a7d164bf7..29589b4ab 100644 --- a/comfy_extras/nodes_mask.py +++ b/comfy_extras/nodes_mask.py @@ -341,6 +341,24 @@ class GrowMask: out.append(output) return (torch.stack(out, dim=0),) +class ThresholdMask: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "mask": ("MASK",), + "value": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}), + } + } + + CATEGORY = "mask" + + RETURN_TYPES = ("MASK",) + FUNCTION = "image_to_mask" + + def image_to_mask(self, mask, value): + mask = (mask > value).float() + return (mask,) NODE_CLASS_MAPPINGS = { @@ -355,6 +373,7 @@ NODE_CLASS_MAPPINGS = { "MaskComposite": MaskComposite, "FeatherMask": FeatherMask, "GrowMask": GrowMask, + "ThresholdMask": ThresholdMask, } NODE_DISPLAY_NAME_MAPPINGS = { diff --git a/comfy_extras/nodes_model_advanced.py b/comfy_extras/nodes_model_advanced.py index 541ce8fa5..21af4b733 100644 --- a/comfy_extras/nodes_model_advanced.py +++ b/comfy_extras/nodes_model_advanced.py @@ -1,6 +1,7 @@ import folder_paths import comfy.sd import comfy.model_sampling +import comfy.latent_formats import torch class LCM(comfy.model_sampling.EPS): @@ -17,6 +18,10 @@ class LCM(comfy.model_sampling.EPS): return c_out * x0 + c_skip * model_input +class X0(comfy.model_sampling.EPS): + def calculate_denoised(self, sigma, model_output, model_input): + return model_output + class ModelSamplingDiscreteDistilled(comfy.model_sampling.ModelSamplingDiscrete): original_timesteps = 50 @@ -68,7 +73,7 @@ class ModelSamplingDiscrete: @classmethod def INPUT_TYPES(s): return {"required": { "model": ("MODEL",), - "sampling": (["eps", "v_prediction", "lcm"],), + "sampling": (["eps", "v_prediction", "lcm", "x0"],), "zsnr": ("BOOLEAN", {"default": False}), }} @@ -88,6 +93,8 @@ class ModelSamplingDiscrete: elif sampling == "lcm": sampling_type = LCM sampling_base = ModelSamplingDiscreteDistilled + elif sampling == "x0": + sampling_type = X0 class ModelSamplingAdvanced(sampling_base, sampling_type): pass @@ -99,11 +106,37 @@ class ModelSamplingDiscrete: m.add_object_patch("model_sampling", model_sampling) return (m, ) +class ModelSamplingStableCascade: + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "shift": ("FLOAT", {"default": 2.0, "min": 0.0, "max": 100.0, "step":0.01}), + }} + + RETURN_TYPES = ("MODEL",) + FUNCTION = "patch" + + CATEGORY = "advanced/model" + + def patch(self, model, shift): + m = model.clone() + + sampling_base = comfy.model_sampling.StableCascadeSampling + sampling_type = comfy.model_sampling.EPS + + class ModelSamplingAdvanced(sampling_base, sampling_type): + pass + + model_sampling = ModelSamplingAdvanced(model.model.model_config) + model_sampling.set_parameters(shift) + m.add_object_patch("model_sampling", model_sampling) + return (m, ) + class ModelSamplingContinuousEDM: @classmethod def INPUT_TYPES(s): return {"required": { "model": ("MODEL",), - "sampling": (["v_prediction", "eps"],), + "sampling": (["v_prediction", "edm_playground_v2.5", "eps"],), "sigma_max": ("FLOAT", {"default": 120.0, "min": 0.0, "max": 1000.0, "step":0.001, "round": False}), "sigma_min": ("FLOAT", {"default": 0.002, "min": 0.0, "max": 1000.0, "step":0.001, "round": False}), }} @@ -116,17 +149,25 @@ class ModelSamplingContinuousEDM: def patch(self, model, sampling, sigma_max, sigma_min): m = model.clone() + latent_format = None + sigma_data = 1.0 if sampling == "eps": sampling_type = comfy.model_sampling.EPS elif sampling == "v_prediction": sampling_type = comfy.model_sampling.V_PREDICTION + elif sampling == "edm_playground_v2.5": + sampling_type = comfy.model_sampling.EDM + sigma_data = 0.5 + latent_format = comfy.latent_formats.SDXL_Playground_2_5() class ModelSamplingAdvanced(comfy.model_sampling.ModelSamplingContinuousEDM, sampling_type): pass model_sampling = ModelSamplingAdvanced(model.model.model_config) - model_sampling.set_sigma_range(sigma_min, sigma_max) + model_sampling.set_parameters(sigma_min, sigma_max, sigma_data) m.add_object_patch("model_sampling", model_sampling) + if latent_format is not None: + m.add_object_patch("latent_format", latent_format) return (m, ) class RescaleCFG: @@ -171,5 +212,6 @@ class RescaleCFG: NODE_CLASS_MAPPINGS = { "ModelSamplingDiscrete": ModelSamplingDiscrete, "ModelSamplingContinuousEDM": ModelSamplingContinuousEDM, + "ModelSamplingStableCascade": ModelSamplingStableCascade, "RescaleCFG": RescaleCFG, } diff --git a/comfy_extras/nodes_model_downscale.py b/comfy_extras/nodes_model_downscale.py index 48bcc6892..58b5073ec 100644 --- a/comfy_extras/nodes_model_downscale.py +++ b/comfy_extras/nodes_model_downscale.py @@ -20,8 +20,9 @@ class PatchModelAddDownscale: CATEGORY = "_for_testing" def patch(self, model, block_number, downscale_factor, start_percent, end_percent, downscale_after_skip, downscale_method, upscale_method): - sigma_start = model.model.model_sampling.percent_to_sigma(start_percent) - sigma_end = model.model.model_sampling.percent_to_sigma(end_percent) + model_sampling = model.get_model_object("model_sampling") + sigma_start = model_sampling.percent_to_sigma(start_percent) + sigma_end = model_sampling.percent_to_sigma(end_percent) def input_block_patch(h, transformer_options): if transformer_options["block"][1] == block_number: diff --git a/comfy_extras/nodes_model_merging.py b/comfy_extras/nodes_model_merging.py index dad1dd637..bb15112f4 100644 --- a/comfy_extras/nodes_model_merging.py +++ b/comfy_extras/nodes_model_merging.py @@ -2,7 +2,9 @@ import comfy.sd import comfy.utils import comfy.model_base import comfy.model_management +import comfy.model_sampling +import torch import folder_paths import json import os @@ -87,6 +89,50 @@ class CLIPMergeSimple: m.add_patches({k: kp[k]}, 1.0 - ratio, ratio) return (m, ) + +class CLIPSubtract: + @classmethod + def INPUT_TYPES(s): + return {"required": { "clip1": ("CLIP",), + "clip2": ("CLIP",), + "multiplier": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}), + }} + RETURN_TYPES = ("CLIP",) + FUNCTION = "merge" + + CATEGORY = "advanced/model_merging" + + def merge(self, clip1, clip2, multiplier): + m = clip1.clone() + kp = clip2.get_key_patches() + for k in kp: + if k.endswith(".position_ids") or k.endswith(".logit_scale"): + continue + m.add_patches({k: kp[k]}, - multiplier, multiplier) + return (m, ) + + +class CLIPAdd: + @classmethod + def INPUT_TYPES(s): + return {"required": { "clip1": ("CLIP",), + "clip2": ("CLIP",), + }} + RETURN_TYPES = ("CLIP",) + FUNCTION = "merge" + + CATEGORY = "advanced/model_merging" + + def merge(self, clip1, clip2): + m = clip1.clone() + kp = clip2.get_key_patches() + for k in kp: + if k.endswith(".position_ids") or k.endswith(".logit_scale"): + continue + m.add_patches({k: kp[k]}, 1.0, 1.0) + return (m, ) + + class ModelMergeBlocks: @classmethod def INPUT_TYPES(s): @@ -119,6 +165,60 @@ class ModelMergeBlocks: m.add_patches({k: kp[k]}, 1.0 - ratio, ratio) return (m, ) +def save_checkpoint(model, clip=None, vae=None, clip_vision=None, filename_prefix=None, output_dir=None, prompt=None, extra_pnginfo=None): + full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, output_dir) + prompt_info = "" + if prompt is not None: + prompt_info = json.dumps(prompt) + + metadata = {} + + enable_modelspec = True + if isinstance(model.model, comfy.model_base.SDXL): + if isinstance(model.model, comfy.model_base.SDXL_instructpix2pix): + metadata["modelspec.architecture"] = "stable-diffusion-xl-v1-edit" + else: + metadata["modelspec.architecture"] = "stable-diffusion-xl-v1-base" + elif isinstance(model.model, comfy.model_base.SDXLRefiner): + metadata["modelspec.architecture"] = "stable-diffusion-xl-v1-refiner" + elif isinstance(model.model, comfy.model_base.SVD_img2vid): + metadata["modelspec.architecture"] = "stable-video-diffusion-img2vid-v1" + else: + enable_modelspec = False + + if enable_modelspec: + metadata["modelspec.sai_model_spec"] = "1.0.0" + metadata["modelspec.implementation"] = "sgm" + metadata["modelspec.title"] = "{} {}".format(filename, counter) + + #TODO: + # "stable-diffusion-v1", "stable-diffusion-v1-inpainting", "stable-diffusion-v2-512", + # "stable-diffusion-v2-768-v", "stable-diffusion-v2-unclip-l", "stable-diffusion-v2-unclip-h", + # "v2-inpainting" + + extra_keys = {} + model_sampling = model.get_model_object("model_sampling") + if isinstance(model_sampling, comfy.model_sampling.ModelSamplingContinuousEDM): + if isinstance(model_sampling, comfy.model_sampling.V_PREDICTION): + extra_keys["edm_vpred.sigma_max"] = torch.tensor(model_sampling.sigma_max).float() + extra_keys["edm_vpred.sigma_min"] = torch.tensor(model_sampling.sigma_min).float() + + if model.model.model_type == comfy.model_base.ModelType.EPS: + metadata["modelspec.predict_key"] = "epsilon" + elif model.model.model_type == comfy.model_base.ModelType.V_PREDICTION: + metadata["modelspec.predict_key"] = "v" + + if not args.disable_metadata: + metadata["prompt"] = prompt_info + if extra_pnginfo is not None: + for x in extra_pnginfo: + metadata[x] = json.dumps(extra_pnginfo[x]) + + output_checkpoint = f"{filename}_{counter:05}_.safetensors" + output_checkpoint = os.path.join(full_output_folder, output_checkpoint) + + comfy.sd.save_checkpoint(output_checkpoint, model, clip, vae, clip_vision, metadata=metadata, extra_keys=extra_keys) + class CheckpointSave: def __init__(self): self.output_dir = folder_paths.get_output_directory() @@ -137,46 +237,7 @@ class CheckpointSave: CATEGORY = "advanced/model_merging" def save(self, model, clip, vae, filename_prefix, prompt=None, extra_pnginfo=None): - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir) - prompt_info = "" - if prompt is not None: - prompt_info = json.dumps(prompt) - - metadata = {} - - enable_modelspec = True - if isinstance(model.model, comfy.model_base.SDXL): - metadata["modelspec.architecture"] = "stable-diffusion-xl-v1-base" - elif isinstance(model.model, comfy.model_base.SDXLRefiner): - metadata["modelspec.architecture"] = "stable-diffusion-xl-v1-refiner" - else: - enable_modelspec = False - - if enable_modelspec: - metadata["modelspec.sai_model_spec"] = "1.0.0" - metadata["modelspec.implementation"] = "sgm" - metadata["modelspec.title"] = "{} {}".format(filename, counter) - - #TODO: - # "stable-diffusion-v1", "stable-diffusion-v1-inpainting", "stable-diffusion-v2-512", - # "stable-diffusion-v2-768-v", "stable-diffusion-v2-unclip-l", "stable-diffusion-v2-unclip-h", - # "v2-inpainting" - - if model.model.model_type == comfy.model_base.ModelType.EPS: - metadata["modelspec.predict_key"] = "epsilon" - elif model.model.model_type == comfy.model_base.ModelType.V_PREDICTION: - metadata["modelspec.predict_key"] = "v" - - if not args.disable_metadata: - metadata["prompt"] = prompt_info - if extra_pnginfo is not None: - for x in extra_pnginfo: - metadata[x] = json.dumps(extra_pnginfo[x]) - - output_checkpoint = f"{filename}_{counter:05}_.safetensors" - output_checkpoint = os.path.join(full_output_folder, output_checkpoint) - - comfy.sd.save_checkpoint(output_checkpoint, model, clip, vae, metadata=metadata) + save_checkpoint(model, clip=clip, vae=vae, filename_prefix=filename_prefix, output_dir=self.output_dir, prompt=prompt, extra_pnginfo=extra_pnginfo) return {} class CLIPSave: @@ -206,7 +267,7 @@ class CLIPSave: for x in extra_pnginfo: metadata[x] = json.dumps(extra_pnginfo[x]) - comfy.model_management.load_models_gpu([clip.load_model()]) + comfy.model_management.load_models_gpu([clip.load_model()], force_patch_weights=True) clip_sd = clip.get_sd() for prefix in ["clip_l.", "clip_g.", ""]: @@ -276,6 +337,8 @@ NODE_CLASS_MAPPINGS = { "ModelMergeAdd": ModelAdd, "CheckpointSave": CheckpointSave, "CLIPMergeSimple": CLIPMergeSimple, + "CLIPMergeSubtract": CLIPSubtract, + "CLIPMergeAdd": CLIPAdd, "CLIPSave": CLIPSave, "VAESave": VAESave, } diff --git a/comfy_extras/nodes_model_merging_model_specific.py b/comfy_extras/nodes_model_merging_model_specific.py new file mode 100644 index 000000000..f2d008d8b --- /dev/null +++ b/comfy_extras/nodes_model_merging_model_specific.py @@ -0,0 +1,60 @@ +import comfy_extras.nodes_model_merging + +class ModelMergeSD1(comfy_extras.nodes_model_merging.ModelMergeBlocks): + CATEGORY = "advanced/model_merging/model_specific" + @classmethod + def INPUT_TYPES(s): + arg_dict = { "model1": ("MODEL",), + "model2": ("MODEL",)} + + argument = ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}) + + arg_dict["time_embed."] = argument + arg_dict["label_emb."] = argument + + for i in range(12): + arg_dict["input_blocks.{}.".format(i)] = argument + + for i in range(3): + arg_dict["middle_block.{}.".format(i)] = argument + + for i in range(12): + arg_dict["output_blocks.{}.".format(i)] = argument + + arg_dict["out."] = argument + + return {"required": arg_dict} + + +class ModelMergeSDXL(comfy_extras.nodes_model_merging.ModelMergeBlocks): + CATEGORY = "advanced/model_merging/model_specific" + + @classmethod + def INPUT_TYPES(s): + arg_dict = { "model1": ("MODEL",), + "model2": ("MODEL",)} + + argument = ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}) + + arg_dict["time_embed."] = argument + arg_dict["label_emb."] = argument + + for i in range(9): + arg_dict["input_blocks.{}".format(i)] = argument + + for i in range(3): + arg_dict["middle_block.{}".format(i)] = argument + + for i in range(9): + arg_dict["output_blocks.{}".format(i)] = argument + + arg_dict["out."] = argument + + return {"required": arg_dict} + + +NODE_CLASS_MAPPINGS = { + "ModelMergeSD1": ModelMergeSD1, + "ModelMergeSD2": ModelMergeSD1, #SD1 and SD2 have the same blocks + "ModelMergeSDXL": ModelMergeSDXL, +} diff --git a/comfy_extras/nodes_morphology.py b/comfy_extras/nodes_morphology.py new file mode 100644 index 000000000..071521d87 --- /dev/null +++ b/comfy_extras/nodes_morphology.py @@ -0,0 +1,49 @@ +import torch +import comfy.model_management + +from kornia.morphology import dilation, erosion, opening, closing, gradient, top_hat, bottom_hat + + +class Morphology: + @classmethod + def INPUT_TYPES(s): + return {"required": {"image": ("IMAGE",), + "operation": (["erode", "dilate", "open", "close", "gradient", "bottom_hat", "top_hat"],), + "kernel_size": ("INT", {"default": 3, "min": 3, "max": 999, "step": 1}), + }} + + RETURN_TYPES = ("IMAGE",) + FUNCTION = "process" + + CATEGORY = "image/postprocessing" + + def process(self, image, operation, kernel_size): + device = comfy.model_management.get_torch_device() + kernel = torch.ones(kernel_size, kernel_size, device=device) + image_k = image.to(device).movedim(-1, 1) + if operation == "erode": + output = erosion(image_k, kernel) + elif operation == "dilate": + output = dilation(image_k, kernel) + elif operation == "open": + output = opening(image_k, kernel) + elif operation == "close": + output = closing(image_k, kernel) + elif operation == "gradient": + output = gradient(image_k, kernel) + elif operation == "top_hat": + output = top_hat(image_k, kernel) + elif operation == "bottom_hat": + output = bottom_hat(image_k, kernel) + else: + raise ValueError(f"Invalid operation {operation} for morphology. Must be one of 'erode', 'dilate', 'open', 'close', 'gradient', 'tophat', 'bottomhat'") + img_out = output.to(comfy.model_management.intermediate_device()).movedim(1, -1) + return (img_out,) + +NODE_CLASS_MAPPINGS = { + "Morphology": Morphology, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "Morphology": "ImageMorphology", +} \ No newline at end of file diff --git a/comfy_extras/nodes_pag.py b/comfy_extras/nodes_pag.py new file mode 100644 index 000000000..63f43fd62 --- /dev/null +++ b/comfy_extras/nodes_pag.py @@ -0,0 +1,56 @@ +#Modified/simplified version of the node from: https://github.com/pamparamm/sd-perturbed-attention +#If you want the one with more options see the above repo. + +#My modified one here is more basic but has less chances of breaking with ComfyUI updates. + +import comfy.model_patcher +import comfy.samplers + +class PerturbedAttentionGuidance: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "model": ("MODEL",), + "scale": ("FLOAT", {"default": 3.0, "min": 0.0, "max": 100.0, "step": 0.1, "round": 0.01}), + } + } + + RETURN_TYPES = ("MODEL",) + FUNCTION = "patch" + + CATEGORY = "_for_testing" + + def patch(self, model, scale): + unet_block = "middle" + unet_block_id = 0 + m = model.clone() + + def perturbed_attention(q, k, v, extra_options, mask=None): + return v + + def post_cfg_function(args): + model = args["model"] + cond_pred = args["cond_denoised"] + cond = args["cond"] + cfg_result = args["denoised"] + sigma = args["sigma"] + model_options = args["model_options"].copy() + x = args["input"] + + if scale == 0: + return cfg_result + + # Replace Self-attention with PAG + model_options = comfy.model_patcher.set_model_options_patch_replace(model_options, perturbed_attention, "attn1", unet_block, unet_block_id) + (pag,) = comfy.samplers.calc_cond_batch(model, [cond], x, sigma, model_options) + + return cfg_result + (cond_pred - pag) * scale + + m.set_model_sampler_post_cfg_function(post_cfg_function) + + return (m,) + +NODE_CLASS_MAPPINGS = { + "PerturbedAttentionGuidance": PerturbedAttentionGuidance, +} diff --git a/comfy_extras/nodes_perpneg.py b/comfy_extras/nodes_perpneg.py index 45e4d418f..546276aa1 100644 --- a/comfy_extras/nodes_perpneg.py +++ b/comfy_extras/nodes_perpneg.py @@ -1,16 +1,26 @@ import torch import comfy.model_management -import comfy.sample +import comfy.sampler_helpers import comfy.samplers import comfy.utils +import node_helpers +def perp_neg(x, noise_pred_pos, noise_pred_neg, noise_pred_nocond, neg_scale, cond_scale): + pos = noise_pred_pos - noise_pred_nocond + neg = noise_pred_neg - noise_pred_nocond + perp = neg - ((torch.mul(neg, pos).sum())/(torch.norm(pos)**2)) * pos + perp_neg = perp * neg_scale + cfg_result = noise_pred_nocond + cond_scale*(pos - perp_neg) + return cfg_result + +#TODO: This node should be removed, it has been replaced with PerpNegGuider class PerpNeg: @classmethod def INPUT_TYPES(s): return {"required": {"model": ("MODEL", ), "empty_conditioning": ("CONDITIONING", ), - "neg_scale": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0}), + "neg_scale": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01}), }} RETURN_TYPES = ("MODEL",) FUNCTION = "patch" @@ -19,7 +29,7 @@ class PerpNeg: def patch(self, model, empty_conditioning, neg_scale): m = model.clone() - nocond = comfy.sample.convert_cond(empty_conditioning) + nocond = comfy.sampler_helpers.convert_cond(empty_conditioning) def cfg_function(args): model = args["model"] @@ -31,14 +41,9 @@ class PerpNeg: model_options = args["model_options"] nocond_processed = comfy.samplers.encode_model_conds(model.extra_conds, nocond, x, x.device, "negative") - (noise_pred_nocond, _) = comfy.samplers.calc_cond_uncond_batch(model, nocond_processed, None, x, sigma, model_options) + (noise_pred_nocond,) = comfy.samplers.calc_cond_batch(model, [nocond_processed], x, sigma, model_options) - pos = noise_pred_pos - noise_pred_nocond - neg = noise_pred_neg - noise_pred_nocond - perp = ((torch.mul(pos, neg).sum())/(torch.norm(neg)**2)) * neg - perp_neg = perp * neg_scale - cfg_result = noise_pred_nocond + cond_scale*(pos - perp_neg) - cfg_result = x - cfg_result + cfg_result = x - perp_neg(x, noise_pred_pos, noise_pred_neg, noise_pred_nocond, neg_scale, cond_scale) return cfg_result m.set_model_sampler_cfg_function(cfg_function) @@ -46,10 +51,78 @@ class PerpNeg: return (m, ) +class Guider_PerpNeg(comfy.samplers.CFGGuider): + def set_conds(self, positive, negative, empty_negative_prompt): + empty_negative_prompt = node_helpers.conditioning_set_values(empty_negative_prompt, {"prompt_type": "negative"}) + self.inner_set_conds({"positive": positive, "empty_negative_prompt": empty_negative_prompt, "negative": negative}) + + def set_cfg(self, cfg, neg_scale): + self.cfg = cfg + self.neg_scale = neg_scale + + def predict_noise(self, x, timestep, model_options={}, seed=None): + # in CFGGuider.predict_noise, we call sampling_function(), which uses cfg_function() to compute pos & neg + # but we'd rather do a single batch of sampling pos, neg, and empty, so we call calc_cond_batch([pos,neg,empty]) directly + + positive_cond = self.conds.get("positive", None) + negative_cond = self.conds.get("negative", None) + empty_cond = self.conds.get("empty_negative_prompt", None) + + (noise_pred_pos, noise_pred_neg, noise_pred_empty) = \ + comfy.samplers.calc_cond_batch(self.inner_model, [positive_cond, negative_cond, empty_cond], x, timestep, model_options) + cfg_result = perp_neg(x, noise_pred_pos, noise_pred_neg, noise_pred_empty, self.neg_scale, self.cfg) + + # normally this would be done in cfg_function, but we skipped + # that for efficiency: we can compute the noise predictions in + # a single call to calc_cond_batch() (rather than two) + # so we replicate the hook here + for fn in model_options.get("sampler_post_cfg_function", []): + args = { + "denoised": cfg_result, + "cond": positive_cond, + "uncond": negative_cond, + "model": self.inner_model, + "uncond_denoised": noise_pred_neg, + "cond_denoised": noise_pred_pos, + "sigma": timestep, + "model_options": model_options, + "input": x, + # not in the original call in samplers.py:cfg_function, but made available for future hooks + "empty_cond": empty_cond, + "empty_cond_denoised": noise_pred_empty,} + cfg_result = fn(args) + + return cfg_result + +class PerpNegGuider: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"model": ("MODEL",), + "positive": ("CONDITIONING", ), + "negative": ("CONDITIONING", ), + "empty_conditioning": ("CONDITIONING", ), + "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), + "neg_scale": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01}), + } + } + + RETURN_TYPES = ("GUIDER",) + + FUNCTION = "get_guider" + CATEGORY = "_for_testing" + + def get_guider(self, model, positive, negative, empty_conditioning, cfg, neg_scale): + guider = Guider_PerpNeg(model) + guider.set_conds(positive, negative, empty_conditioning) + guider.set_cfg(cfg, neg_scale) + return (guider,) + NODE_CLASS_MAPPINGS = { "PerpNeg": PerpNeg, + "PerpNegGuider": PerpNegGuider, } NODE_DISPLAY_NAME_MAPPINGS = { - "PerpNeg": "Perp-Neg", + "PerpNeg": "Perp-Neg (DEPRECATED by PerpNegGuider)", } diff --git a/comfy_extras/nodes_photomaker.py b/comfy_extras/nodes_photomaker.py new file mode 100644 index 000000000..29d127d74 --- /dev/null +++ b/comfy_extras/nodes_photomaker.py @@ -0,0 +1,187 @@ +import torch +import torch.nn as nn +import folder_paths +import comfy.clip_model +import comfy.clip_vision +import comfy.ops + +# code for model from: https://github.com/TencentARC/PhotoMaker/blob/main/photomaker/model.py under Apache License Version 2.0 +VISION_CONFIG_DICT = { + "hidden_size": 1024, + "image_size": 224, + "intermediate_size": 4096, + "num_attention_heads": 16, + "num_channels": 3, + "num_hidden_layers": 24, + "patch_size": 14, + "projection_dim": 768, + "hidden_act": "quick_gelu", +} + +class MLP(nn.Module): + def __init__(self, in_dim, out_dim, hidden_dim, use_residual=True, operations=comfy.ops): + super().__init__() + if use_residual: + assert in_dim == out_dim + self.layernorm = operations.LayerNorm(in_dim) + self.fc1 = operations.Linear(in_dim, hidden_dim) + self.fc2 = operations.Linear(hidden_dim, out_dim) + self.use_residual = use_residual + self.act_fn = nn.GELU() + + def forward(self, x): + residual = x + x = self.layernorm(x) + x = self.fc1(x) + x = self.act_fn(x) + x = self.fc2(x) + if self.use_residual: + x = x + residual + return x + + +class FuseModule(nn.Module): + def __init__(self, embed_dim, operations): + super().__init__() + self.mlp1 = MLP(embed_dim * 2, embed_dim, embed_dim, use_residual=False, operations=operations) + self.mlp2 = MLP(embed_dim, embed_dim, embed_dim, use_residual=True, operations=operations) + self.layer_norm = operations.LayerNorm(embed_dim) + + def fuse_fn(self, prompt_embeds, id_embeds): + stacked_id_embeds = torch.cat([prompt_embeds, id_embeds], dim=-1) + stacked_id_embeds = self.mlp1(stacked_id_embeds) + prompt_embeds + stacked_id_embeds = self.mlp2(stacked_id_embeds) + stacked_id_embeds = self.layer_norm(stacked_id_embeds) + return stacked_id_embeds + + def forward( + self, + prompt_embeds, + id_embeds, + class_tokens_mask, + ) -> torch.Tensor: + # id_embeds shape: [b, max_num_inputs, 1, 2048] + id_embeds = id_embeds.to(prompt_embeds.dtype) + num_inputs = class_tokens_mask.sum().unsqueeze(0) # TODO: check for training case + batch_size, max_num_inputs = id_embeds.shape[:2] + # seq_length: 77 + seq_length = prompt_embeds.shape[1] + # flat_id_embeds shape: [b*max_num_inputs, 1, 2048] + flat_id_embeds = id_embeds.view( + -1, id_embeds.shape[-2], id_embeds.shape[-1] + ) + # valid_id_mask [b*max_num_inputs] + valid_id_mask = ( + torch.arange(max_num_inputs, device=flat_id_embeds.device)[None, :] + < num_inputs[:, None] + ) + valid_id_embeds = flat_id_embeds[valid_id_mask.flatten()] + + prompt_embeds = prompt_embeds.view(-1, prompt_embeds.shape[-1]) + class_tokens_mask = class_tokens_mask.view(-1) + valid_id_embeds = valid_id_embeds.view(-1, valid_id_embeds.shape[-1]) + # slice out the image token embeddings + image_token_embeds = prompt_embeds[class_tokens_mask] + stacked_id_embeds = self.fuse_fn(image_token_embeds, valid_id_embeds) + assert class_tokens_mask.sum() == stacked_id_embeds.shape[0], f"{class_tokens_mask.sum()} != {stacked_id_embeds.shape[0]}" + prompt_embeds.masked_scatter_(class_tokens_mask[:, None], stacked_id_embeds.to(prompt_embeds.dtype)) + updated_prompt_embeds = prompt_embeds.view(batch_size, seq_length, -1) + return updated_prompt_embeds + +class PhotoMakerIDEncoder(comfy.clip_model.CLIPVisionModelProjection): + def __init__(self): + self.load_device = comfy.model_management.text_encoder_device() + offload_device = comfy.model_management.text_encoder_offload_device() + dtype = comfy.model_management.text_encoder_dtype(self.load_device) + + super().__init__(VISION_CONFIG_DICT, dtype, offload_device, comfy.ops.manual_cast) + self.visual_projection_2 = comfy.ops.manual_cast.Linear(1024, 1280, bias=False) + self.fuse_module = FuseModule(2048, comfy.ops.manual_cast) + + def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask): + b, num_inputs, c, h, w = id_pixel_values.shape + id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w) + + shared_id_embeds = self.vision_model(id_pixel_values)[2] + id_embeds = self.visual_projection(shared_id_embeds) + id_embeds_2 = self.visual_projection_2(shared_id_embeds) + + id_embeds = id_embeds.view(b, num_inputs, 1, -1) + id_embeds_2 = id_embeds_2.view(b, num_inputs, 1, -1) + + id_embeds = torch.cat((id_embeds, id_embeds_2), dim=-1) + updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask) + + return updated_prompt_embeds + + +class PhotoMakerLoader: + @classmethod + def INPUT_TYPES(s): + return {"required": { "photomaker_model_name": (folder_paths.get_filename_list("photomaker"), )}} + + RETURN_TYPES = ("PHOTOMAKER",) + FUNCTION = "load_photomaker_model" + + CATEGORY = "_for_testing/photomaker" + + def load_photomaker_model(self, photomaker_model_name): + photomaker_model_path = folder_paths.get_full_path("photomaker", photomaker_model_name) + photomaker_model = PhotoMakerIDEncoder() + data = comfy.utils.load_torch_file(photomaker_model_path, safe_load=True) + if "id_encoder" in data: + data = data["id_encoder"] + photomaker_model.load_state_dict(data) + return (photomaker_model,) + + +class PhotoMakerEncode: + @classmethod + def INPUT_TYPES(s): + return {"required": { "photomaker": ("PHOTOMAKER",), + "image": ("IMAGE",), + "clip": ("CLIP", ), + "text": ("STRING", {"multiline": True, "dynamicPrompts": True, "default": "photograph of photomaker"}), + }} + + RETURN_TYPES = ("CONDITIONING",) + FUNCTION = "apply_photomaker" + + CATEGORY = "_for_testing/photomaker" + + def apply_photomaker(self, photomaker, image, clip, text): + special_token = "photomaker" + pixel_values = comfy.clip_vision.clip_preprocess(image.to(photomaker.load_device)).float() + try: + index = text.split(" ").index(special_token) + 1 + except ValueError: + index = -1 + tokens = clip.tokenize(text, return_word_ids=True) + out_tokens = {} + for k in tokens: + out_tokens[k] = [] + for t in tokens[k]: + f = list(filter(lambda x: x[2] != index, t)) + while len(f) < len(t): + f.append(t[-1]) + out_tokens[k].append(f) + + cond, pooled = clip.encode_from_tokens(out_tokens, return_pooled=True) + + if index > 0: + token_index = index - 1 + num_id_images = 1 + class_tokens_mask = [True if token_index <= i < token_index+num_id_images else False for i in range(77)] + out = photomaker(id_pixel_values=pixel_values.unsqueeze(0), prompt_embeds=cond.to(photomaker.load_device), + class_tokens_mask=torch.tensor(class_tokens_mask, dtype=torch.bool, device=photomaker.load_device).unsqueeze(0)) + else: + out = cond + + return ([[out, {"pooled_output": pooled}]], ) + + +NODE_CLASS_MAPPINGS = { + "PhotoMakerLoader": PhotoMakerLoader, + "PhotoMakerEncode": PhotoMakerEncode, +} + diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 71660f8a5..68f6ef51e 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -5,6 +5,7 @@ from PIL import Image import math import comfy.utils +import comfy.model_management class Blend: @@ -33,6 +34,7 @@ class Blend: CATEGORY = "image/postprocessing" def blend_images(self, image1: torch.Tensor, image2: torch.Tensor, blend_factor: float, blend_mode: str): + image2 = image2.to(image1.device) if image1.shape != image2.shape: image2 = image2.permute(0, 3, 1, 2) image2 = comfy.utils.common_upscale(image2, image1.shape[2], image1.shape[1], upscale_method='bicubic', crop='center') @@ -101,6 +103,7 @@ class Blur: if blur_radius == 0: return (image,) + image = image.to(comfy.model_management.get_torch_device()) batch_size, height, width, channels = image.shape kernel_size = blur_radius * 2 + 1 @@ -111,7 +114,7 @@ class Blur: blurred = F.conv2d(padded_image, kernel, padding=kernel_size // 2, groups=channels)[:,:,blur_radius:-blur_radius, blur_radius:-blur_radius] blurred = blurred.permute(0, 2, 3, 1) - return (blurred,) + return (blurred.to(comfy.model_management.intermediate_device()),) class Quantize: def __init__(self): @@ -203,13 +206,13 @@ class Sharpen: "default": 1.0, "min": 0.1, "max": 10.0, - "step": 0.1 + "step": 0.01 }), "alpha": ("FLOAT", { "default": 1.0, "min": 0.0, "max": 5.0, - "step": 0.1 + "step": 0.01 }), }, } @@ -224,6 +227,7 @@ class Sharpen: return (image,) batch_size, height, width, channels = image.shape + image = image.to(comfy.model_management.get_torch_device()) kernel_size = sharpen_radius * 2 + 1 kernel = gaussian_kernel(kernel_size, sigma, device=image.device) * -(alpha*10) @@ -238,7 +242,7 @@ class Sharpen: result = torch.clamp(sharpened, 0, 1) - return (result,) + return (result.to(comfy.model_management.intermediate_device()),) class ImageScaleToTotalPixels: upscale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "lanczos"] diff --git a/comfy_extras/nodes_sag.py b/comfy_extras/nodes_sag.py index bbd380807..010e99744 100644 --- a/comfy_extras/nodes_sag.py +++ b/comfy_extras/nodes_sag.py @@ -4,13 +4,12 @@ import torch.nn.functional as F import math from einops import rearrange, repeat -import os -from comfy.ldm.modules.attention import optimized_attention, _ATTN_PRECISION +from comfy.ldm.modules.attention import optimized_attention import comfy.samplers # from comfy/ldm/modules/attention.py # but modified to return attention scores as well as output -def attention_basic_with_sim(q, k, v, heads, mask=None): +def attention_basic_with_sim(q, k, v, heads, mask=None, attn_precision=None): b, _, dim_head = q.shape dim_head //= heads scale = dim_head ** -0.5 @@ -26,7 +25,7 @@ def attention_basic_with_sim(q, k, v, heads, mask=None): ) # force cast to fp32 to avoid overflowing - if _ATTN_PRECISION =="fp32": + if attn_precision == torch.float32: sim = einsum('b i d, b j d -> b i j', q.float(), k.float()) * scale else: sim = einsum('b i d, b j d -> b i j', q, k) * scale @@ -121,13 +120,13 @@ class SelfAttentionGuidance: if 1 in cond_or_uncond: uncond_index = cond_or_uncond.index(1) # do the entire attention operation, but save the attention scores to attn_scores - (out, sim) = attention_basic_with_sim(q, k, v, heads=heads) + (out, sim) = attention_basic_with_sim(q, k, v, heads=heads, attn_precision=extra_options["attn_precision"]) # when using a higher batch size, I BELIEVE the result batch dimension is [uc1, ... ucn, c1, ... cn] n_slices = heads * b attn_scores = sim[n_slices * uncond_index:n_slices * (uncond_index+1)] return out else: - return optimized_attention(q, k, v, heads=heads) + return optimized_attention(q, k, v, heads=heads, attn_precision=extra_options["attn_precision"]) def post_cfg_function(args): nonlocal attn_scores @@ -150,7 +149,7 @@ class SelfAttentionGuidance: degraded = create_blur_map(uncond_pred, uncond_attn, sag_sigma, sag_threshold) degraded_noised = degraded + x - uncond_pred # call into the UNet - (sag, _) = comfy.samplers.calc_cond_uncond_batch(model, uncond, None, degraded_noised, sigma, model_options) + (sag,) = comfy.samplers.calc_cond_batch(model, [uncond], degraded_noised, sigma, model_options) return cfg_result + (degraded - sag) * sag_scale m.set_model_sampler_post_cfg_function(post_cfg_function, disable_cfg1_optimization=True) diff --git a/comfy_extras/nodes_sdupscale.py b/comfy_extras/nodes_sdupscale.py index 28c1cb0f1..bba67e8dd 100644 --- a/comfy_extras/nodes_sdupscale.py +++ b/comfy_extras/nodes_sdupscale.py @@ -1,5 +1,4 @@ import torch -import nodes import comfy.utils class SD_4XUpscale_Conditioning: diff --git a/comfy_extras/nodes_stable3d.py b/comfy_extras/nodes_stable3d.py index e02a9875a..be2e34c28 100644 --- a/comfy_extras/nodes_stable3d.py +++ b/comfy_extras/nodes_stable3d.py @@ -29,8 +29,8 @@ class StableZero123_Conditioning: "width": ("INT", {"default": 256, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), "height": ("INT", {"default": 256, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), - "elevation": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), - "azimuth": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), + "elevation": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), + "azimuth": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), }} RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT") RETURN_NAMES = ("positive", "negative", "latent") @@ -46,7 +46,7 @@ class StableZero123_Conditioning: encode_pixels = pixels[:,:,:,:3] t = vae.encode(encode_pixels) cam_embeds = camera_embeddings(elevation, azimuth) - cond = torch.cat([pooled, cam_embeds.repeat((pooled.shape[0], 1, 1))], dim=-1) + cond = torch.cat([pooled, cam_embeds.to(pooled.device).repeat((pooled.shape[0], 1, 1))], dim=-1) positive = [[cond, {"concat_latent_image": t}]] negative = [[torch.zeros_like(pooled), {"concat_latent_image": torch.zeros_like(t)}]] @@ -62,10 +62,10 @@ class StableZero123_Conditioning_Batched: "width": ("INT", {"default": 256, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), "height": ("INT", {"default": 256, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), - "elevation": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), - "azimuth": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), - "elevation_batch_increment": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), - "azimuth_batch_increment": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0}), + "elevation": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), + "azimuth": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), + "elevation_batch_increment": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), + "azimuth_batch_increment": ("FLOAT", {"default": 0.0, "min": -180.0, "max": 180.0, "step": 0.1, "round": False}), }} RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT") RETURN_NAMES = ("positive", "negative", "latent") @@ -95,8 +95,49 @@ class StableZero123_Conditioning_Batched: latent = torch.zeros([batch_size, 4, height // 8, width // 8]) return (positive, negative, {"samples":latent, "batch_index": [0] * batch_size}) +class SV3D_Conditioning: + @classmethod + def INPUT_TYPES(s): + return {"required": { "clip_vision": ("CLIP_VISION",), + "init_image": ("IMAGE",), + "vae": ("VAE",), + "width": ("INT", {"default": 576, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), + "height": ("INT", {"default": 576, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}), + "video_frames": ("INT", {"default": 21, "min": 1, "max": 4096}), + "elevation": ("FLOAT", {"default": 0.0, "min": -90.0, "max": 90.0, "step": 0.1, "round": False}), + }} + RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT") + RETURN_NAMES = ("positive", "negative", "latent") + + FUNCTION = "encode" + + CATEGORY = "conditioning/3d_models" + + def encode(self, clip_vision, init_image, vae, width, height, video_frames, elevation): + output = clip_vision.encode_image(init_image) + pooled = output.image_embeds.unsqueeze(0) + pixels = comfy.utils.common_upscale(init_image.movedim(-1,1), width, height, "bilinear", "center").movedim(1,-1) + encode_pixels = pixels[:,:,:,:3] + t = vae.encode(encode_pixels) + + azimuth = 0 + azimuth_increment = 360 / (max(video_frames, 2) - 1) + + elevations = [] + azimuths = [] + for i in range(video_frames): + elevations.append(elevation) + azimuths.append(azimuth) + azimuth += azimuth_increment + + positive = [[pooled, {"concat_latent_image": t, "elevation": elevations, "azimuth": azimuths}]] + negative = [[torch.zeros_like(pooled), {"concat_latent_image": torch.zeros_like(t), "elevation": elevations, "azimuth": azimuths}]] + latent = torch.zeros([video_frames, 4, height // 8, width // 8]) + return (positive, negative, {"samples":latent}) + NODE_CLASS_MAPPINGS = { "StableZero123_Conditioning": StableZero123_Conditioning, "StableZero123_Conditioning_Batched": StableZero123_Conditioning_Batched, + "SV3D_Conditioning": SV3D_Conditioning, } diff --git a/comfy_extras/nodes_stable_cascade.py b/comfy_extras/nodes_stable_cascade.py new file mode 100644 index 000000000..fcbbeb27f --- /dev/null +++ b/comfy_extras/nodes_stable_cascade.py @@ -0,0 +1,140 @@ +""" + This file is part of ComfyUI. + Copyright (C) 2024 Stability AI + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . +""" + +import torch +import nodes +import comfy.utils + + +class StableCascade_EmptyLatentImage: + def __init__(self, device="cpu"): + self.device = device + + @classmethod + def INPUT_TYPES(s): + return {"required": { + "width": ("INT", {"default": 1024, "min": 256, "max": nodes.MAX_RESOLUTION, "step": 8}), + "height": ("INT", {"default": 1024, "min": 256, "max": nodes.MAX_RESOLUTION, "step": 8}), + "compression": ("INT", {"default": 42, "min": 4, "max": 128, "step": 1}), + "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}) + }} + RETURN_TYPES = ("LATENT", "LATENT") + RETURN_NAMES = ("stage_c", "stage_b") + FUNCTION = "generate" + + CATEGORY = "latent/stable_cascade" + + def generate(self, width, height, compression, batch_size=1): + c_latent = torch.zeros([batch_size, 16, height // compression, width // compression]) + b_latent = torch.zeros([batch_size, 4, height // 4, width // 4]) + return ({ + "samples": c_latent, + }, { + "samples": b_latent, + }) + +class StableCascade_StageC_VAEEncode: + def __init__(self, device="cpu"): + self.device = device + + @classmethod + def INPUT_TYPES(s): + return {"required": { + "image": ("IMAGE",), + "vae": ("VAE", ), + "compression": ("INT", {"default": 42, "min": 4, "max": 128, "step": 1}), + }} + RETURN_TYPES = ("LATENT", "LATENT") + RETURN_NAMES = ("stage_c", "stage_b") + FUNCTION = "generate" + + CATEGORY = "latent/stable_cascade" + + def generate(self, image, vae, compression): + width = image.shape[-2] + height = image.shape[-3] + out_width = (width // compression) * vae.downscale_ratio + out_height = (height // compression) * vae.downscale_ratio + + s = comfy.utils.common_upscale(image.movedim(-1,1), out_width, out_height, "bicubic", "center").movedim(1,-1) + + c_latent = vae.encode(s[:,:,:,:3]) + b_latent = torch.zeros([c_latent.shape[0], 4, (height // 8) * 2, (width // 8) * 2]) + return ({ + "samples": c_latent, + }, { + "samples": b_latent, + }) + +class StableCascade_StageB_Conditioning: + @classmethod + def INPUT_TYPES(s): + return {"required": { "conditioning": ("CONDITIONING",), + "stage_c": ("LATENT",), + }} + RETURN_TYPES = ("CONDITIONING",) + + FUNCTION = "set_prior" + + CATEGORY = "conditioning/stable_cascade" + + def set_prior(self, conditioning, stage_c): + c = [] + for t in conditioning: + d = t[1].copy() + d['stable_cascade_prior'] = stage_c['samples'] + n = [t[0], d] + c.append(n) + return (c, ) + +class StableCascade_SuperResolutionControlnet: + def __init__(self, device="cpu"): + self.device = device + + @classmethod + def INPUT_TYPES(s): + return {"required": { + "image": ("IMAGE",), + "vae": ("VAE", ), + }} + RETURN_TYPES = ("IMAGE", "LATENT", "LATENT") + RETURN_NAMES = ("controlnet_input", "stage_c", "stage_b") + FUNCTION = "generate" + + CATEGORY = "_for_testing/stable_cascade" + + def generate(self, image, vae): + width = image.shape[-2] + height = image.shape[-3] + batch_size = image.shape[0] + controlnet_input = vae.encode(image[:,:,:,:3]).movedim(1, -1) + + c_latent = torch.zeros([batch_size, 16, height // 16, width // 16]) + b_latent = torch.zeros([batch_size, 4, height // 2, width // 2]) + return (controlnet_input, { + "samples": c_latent, + }, { + "samples": b_latent, + }) + +NODE_CLASS_MAPPINGS = { + "StableCascade_EmptyLatentImage": StableCascade_EmptyLatentImage, + "StableCascade_StageB_Conditioning": StableCascade_StageB_Conditioning, + "StableCascade_StageC_VAEEncode": StableCascade_StageC_VAEEncode, + "StableCascade_SuperResolutionControlnet": StableCascade_SuperResolutionControlnet, +} diff --git a/comfy_extras/nodes_upscale_model.py b/comfy_extras/nodes_upscale_model.py index 2b5e49a55..bca79ef2e 100644 --- a/comfy_extras/nodes_upscale_model.py +++ b/comfy_extras/nodes_upscale_model.py @@ -1,10 +1,19 @@ import os -from comfy_extras.chainner_models import model_loading +import logging +from spandrel import ModelLoader, ImageModelDescriptor from comfy import model_management import torch import comfy.utils import folder_paths +try: + from spandrel_extra_arches import EXTRA_REGISTRY + from spandrel import MAIN_REGISTRY + MAIN_REGISTRY.add(*EXTRA_REGISTRY) + logging.info("Successfully imported spandrel_extra_arches: support for non commercial upscale models.") +except: + pass + class UpscaleModelLoader: @classmethod def INPUT_TYPES(s): @@ -20,7 +29,11 @@ class UpscaleModelLoader: sd = comfy.utils.load_torch_file(model_path, safe_load=True) if "module.layers.0.residual_group.blocks.0.norm1.weight" in sd: sd = comfy.utils.state_dict_prefix_replace(sd, {"module.":""}) - out = model_loading.load_state_dict(sd).eval() + out = ModelLoader().load_from_state_dict(sd).eval() + + if not isinstance(out, ImageModelDescriptor): + raise Exception("Upscale model must be a single-image model.") + return (out, ) @@ -37,9 +50,14 @@ class ImageUpscaleWithModel: def upscale(self, upscale_model, image): device = model_management.get_torch_device() + + memory_required = model_management.module_size(upscale_model.model) + memory_required += (512 * 512 * 3) * image.element_size() * max(upscale_model.scale, 1.0) * 384.0 #The 384.0 is an estimate of how much some of these models take, TODO: make it more accurate + memory_required += image.nelement() * image.element_size() + model_management.free_memory(memory_required, device) + upscale_model.to(device) in_img = image.movedim(-1,-3).to(device) - free_memory = model_management.get_free_memory(device) tile = 512 overlap = 32 @@ -56,7 +74,7 @@ class ImageUpscaleWithModel: if tile < 128: raise e - upscale_model.cpu() + upscale_model.to("cpu") s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0) return (s,) diff --git a/comfy_extras/nodes_video_model.py b/comfy_extras/nodes_video_model.py index 26a717a38..1a0189ed4 100644 --- a/comfy_extras/nodes_video_model.py +++ b/comfy_extras/nodes_video_model.py @@ -3,6 +3,7 @@ import torch import comfy.utils import comfy.sd import folder_paths +import comfy_extras.nodes_model_merging class ImageOnlyCheckpointLoader: @@ -78,10 +79,54 @@ class VideoLinearCFGGuidance: m.set_model_sampler_cfg_function(linear_cfg) return (m, ) +class VideoTriangleCFGGuidance: + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "min_cfg": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.5, "round": 0.01}), + }} + RETURN_TYPES = ("MODEL",) + FUNCTION = "patch" + + CATEGORY = "sampling/video_models" + + def patch(self, model, min_cfg): + def linear_cfg(args): + cond = args["cond"] + uncond = args["uncond"] + cond_scale = args["cond_scale"] + period = 1.0 + values = torch.linspace(0, 1, cond.shape[0], device=cond.device) + values = 2 * (values / period - torch.floor(values / period + 0.5)).abs() + scale = (values * (cond_scale - min_cfg) + min_cfg).reshape((cond.shape[0], 1, 1, 1)) + + return uncond + scale * (cond - uncond) + + m = model.clone() + m.set_model_sampler_cfg_function(linear_cfg) + return (m, ) + +class ImageOnlyCheckpointSave(comfy_extras.nodes_model_merging.CheckpointSave): + CATEGORY = "_for_testing" + + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "clip_vision": ("CLIP_VISION",), + "vae": ("VAE",), + "filename_prefix": ("STRING", {"default": "checkpoints/ComfyUI"}),}, + "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},} + + def save(self, model, clip_vision, vae, filename_prefix, prompt=None, extra_pnginfo=None): + comfy_extras.nodes_model_merging.save_checkpoint(model, clip_vision=clip_vision, vae=vae, filename_prefix=filename_prefix, output_dir=self.output_dir, prompt=prompt, extra_pnginfo=extra_pnginfo) + return {} + NODE_CLASS_MAPPINGS = { "ImageOnlyCheckpointLoader": ImageOnlyCheckpointLoader, "SVD_img2vid_Conditioning": SVD_img2vid_Conditioning, "VideoLinearCFGGuidance": VideoLinearCFGGuidance, + "VideoTriangleCFGGuidance": VideoTriangleCFGGuidance, + "ImageOnlyCheckpointSave": ImageOnlyCheckpointSave, } NODE_DISPLAY_NAME_MAPPINGS = { diff --git a/comfy_extras/nodes_webcam.py b/comfy_extras/nodes_webcam.py new file mode 100644 index 000000000..32a0ba2f6 --- /dev/null +++ b/comfy_extras/nodes_webcam.py @@ -0,0 +1,33 @@ +import nodes +import folder_paths + +MAX_RESOLUTION = nodes.MAX_RESOLUTION + + +class WebcamCapture(nodes.LoadImage): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": ("WEBCAM", {}), + "width": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), + "height": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), + "capture_on_queue": ("BOOLEAN", {"default": True}), + } + } + RETURN_TYPES = ("IMAGE",) + FUNCTION = "load_capture" + + CATEGORY = "image" + + def load_capture(s, image, **kwargs): + return super().load_image(folder_paths.get_annotated_filepath(image)) + + +NODE_CLASS_MAPPINGS = { + "WebcamCapture": WebcamCapture, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "WebcamCapture": "Webcam Capture", +} \ No newline at end of file diff --git a/cuda_malloc.py b/cuda_malloc.py index 144cdacd3..eb2857c5f 100644 --- a/cuda_malloc.py +++ b/cuda_malloc.py @@ -1,6 +1,7 @@ import os import importlib.util from comfy.cli_args import args +import subprocess #Can't use pytorch to get the GPU names because the cuda malloc has to be set before the first import. def get_gpu_names(): @@ -34,14 +35,19 @@ def get_gpu_names(): return gpu_names return enum_display_devices() else: - return set() + gpu_names = set() + out = subprocess.check_output(['nvidia-smi', '-L']) + for l in out.split(b'\n'): + if len(l) > 0: + gpu_names.add(l.decode('utf-8').split(' (UUID')[0]) + return gpu_names blacklist = {"GeForce GTX TITAN X", "GeForce GTX 980", "GeForce GTX 970", "GeForce GTX 960", "GeForce GTX 950", "GeForce 945M", "GeForce 940M", "GeForce 930M", "GeForce 920M", "GeForce 910M", "GeForce GTX 750", "GeForce GTX 745", "Quadro K620", "Quadro K1200", "Quadro K2200", "Quadro M500", "Quadro M520", "Quadro M600", "Quadro M620", "Quadro M1000", "Quadro M1200", "Quadro M2000", "Quadro M2200", "Quadro M3000", "Quadro M4000", "Quadro M5000", "Quadro M5500", "Quadro M6000", "GeForce MX110", "GeForce MX130", "GeForce 830M", "GeForce 840M", "GeForce GTX 850M", "GeForce GTX 860M", - "GeForce GTX 1650", "GeForce GTX 1630" + "GeForce GTX 1650", "GeForce GTX 1630", "Tesla M4", "Tesla M6", "Tesla M10", "Tesla M40", "Tesla M60" } def cuda_malloc_supported(): diff --git a/custom_nodes/example_node.py.example b/custom_nodes/example_node.py.example index 733014f3c..f06632593 100644 --- a/custom_nodes/example_node.py.example +++ b/custom_nodes/example_node.py.example @@ -6,6 +6,8 @@ class Example: ------------- INPUT_TYPES (dict): Tell the main program input parameters of nodes. + IS_CHANGED: + optional method to control when the node is re executed. Attributes ---------- @@ -89,6 +91,20 @@ class Example: image = 1.0 - image return (image,) + """ + The node will always be re executed if any of the inputs change but + this method can be used to force the node to execute again even when the inputs don't change. + You can make this node return a number or a string. This value will be compared to the one returned the last time the node was + executed, if it is different the node will be executed again. + This method is used in the core repo for the LoadImage node where they return the image hash as a string, if the image hash + changes between executions the LoadImage node is executed again. + """ + #@classmethod + #def IS_CHANGED(s, image, string_field, int_field, float_field, print_to_screen): + # return "" + +# Set the web directory, any .js file in that directory will be loaded by the frontend as a frontend extension +# WEB_DIRECTORY = "./somejs" # A dictionary that contains all nodes you want to export with their names # NOTE: names should be globally unique diff --git a/custom_nodes/websocket_image_save.py b/custom_nodes/websocket_image_save.py new file mode 100644 index 000000000..5aa573642 --- /dev/null +++ b/custom_nodes/websocket_image_save.py @@ -0,0 +1,45 @@ +from PIL import Image, ImageOps +from io import BytesIO +import numpy as np +import struct +import comfy.utils +import time + +#You can use this node to save full size images through the websocket, the +#images will be sent in exactly the same format as the image previews: as +#binary images on the websocket with a 8 byte header indicating the type +#of binary message (first 4 bytes) and the image format (next 4 bytes). + +#Note that no metadata will be put in the images saved with this node. + +class SaveImageWebsocket: + @classmethod + def INPUT_TYPES(s): + return {"required": + {"images": ("IMAGE", ),} + } + + RETURN_TYPES = () + FUNCTION = "save_images" + + OUTPUT_NODE = True + + CATEGORY = "api/image" + + def save_images(self, images): + pbar = comfy.utils.ProgressBar(images.shape[0]) + step = 0 + for image in images: + i = 255. * image.cpu().numpy() + img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8)) + pbar.update_absolute(step, images.shape[0], ("PNG", img, None)) + step += 1 + + return {} + + def IS_CHANGED(s, images): + return time.time() + +NODE_CLASS_MAPPINGS = { + "SaveImageWebsocket": SaveImageWebsocket, +} diff --git a/execution.py b/execution.py index e91e9a410..76225a962 100644 --- a/execution.py +++ b/execution.py @@ -1,12 +1,9 @@ -import os import sys import copy -import json import logging import threading import heapq import traceback -import gc import inspect from typing import List, Literal, NamedTuple, Optional @@ -38,8 +35,7 @@ def get_input_data(inputs, class_def, unique_id, outputs={}, prompt={}, extra_da if h[x] == "PROMPT": input_data_all[x] = [prompt] if h[x] == "EXTRA_PNGINFO": - if "extra_pnginfo" in extra_data: - input_data_all[x] = [extra_data['extra_pnginfo']] + input_data_all[x] = [extra_data.get('extra_pnginfo', None)] if h[x] == "UNIQUE_ID": input_data_all[x] = [unique_id] return input_data_all @@ -180,7 +176,7 @@ def recursive_execute(server, prompt, outputs, current_item, extra_data, execute for node_id, node_outputs in outputs.items(): output_data_formatted[node_id] = [[format_value(x) for x in l] for l in node_outputs] - logging.error("!!! Exception during processing !!!") + logging.error(f"!!! Exception during processing!!! {ex}") logging.error(traceback.format_exc()) error_details = { @@ -197,8 +193,12 @@ def recursive_execute(server, prompt, outputs, current_item, extra_data, execute return (True, None, None) -def recursive_will_execute(prompt, outputs, current_item): +def recursive_will_execute(prompt, outputs, current_item, memo={}): unique_id = current_item + + if unique_id in memo: + return memo[unique_id] + inputs = prompt[unique_id]['inputs'] will_execute = [] if unique_id in outputs: @@ -210,9 +210,10 @@ def recursive_will_execute(prompt, outputs, current_item): input_unique_id = input_data[0] output_index = input_data[1] if input_unique_id not in outputs: - will_execute += recursive_will_execute(prompt, outputs, input_unique_id) + will_execute += recursive_will_execute(prompt, outputs, input_unique_id, memo) - return will_execute + [unique_id] + memo[unique_id] = will_execute + [unique_id] + return memo[unique_id] def recursive_output_delete_if_changed(prompt, old_prompt, outputs, current_item): unique_id = current_item @@ -367,7 +368,7 @@ class PromptExecutor: d = self.outputs_ui.pop(x) del d - comfy.model_management.cleanup_models() + comfy.model_management.cleanup_models(keep_clone_weights_loaded=True) self.add_message("execution_cached", { "nodes": list(current_outputs) , "prompt_id": prompt_id}, broadcast=False) @@ -380,7 +381,8 @@ class PromptExecutor: while len(to_execute) > 0: #always execute the output that depends on the least amount of unexecuted nodes first - to_execute = sorted(list(map(lambda a: (len(recursive_will_execute(prompt, self.outputs, a[-1])), a[-1]), to_execute))) + memo = {} + to_execute = sorted(list(map(lambda a: (len(recursive_will_execute(prompt, self.outputs, a[-1], memo)), a[-1]), to_execute))) output_node_id = to_execute.pop(0)[-1] # This call shouldn't raise anything if there's an error deep in @@ -620,8 +622,27 @@ def full_type_name(klass): def validate_prompt(prompt): outputs = set() for x in prompt: - class_ = nodes.NODE_CLASS_MAPPINGS[prompt[x]['class_type']] - if hasattr(class_, 'OUTPUT_NODE') and class_.OUTPUT_NODE == True: + if 'class_type' not in prompt[x]: + error = { + "type": "invalid_prompt", + "message": f"Cannot execute because a node is missing the class_type property.", + "details": f"Node ID '#{x}'", + "extra_info": {} + } + return (False, error, [], []) + + class_type = prompt[x]['class_type'] + class_ = nodes.NODE_CLASS_MAPPINGS.get(class_type, None) + if class_ is None: + error = { + "type": "invalid_prompt", + "message": f"Cannot execute because node {class_type} does not exist.", + "details": f"Node ID '#{x}'", + "extra_info": {} + } + return (False, error, [], []) + + if hasattr(class_, 'OUTPUT_NODE') and class_.OUTPUT_NODE is True: outputs.add(x) if len(outputs) == 0: @@ -751,7 +772,7 @@ class PromptQueue: if len(self.history) > MAXIMUM_HISTORY_SIZE: self.history.pop(next(iter(self.history))) - status_dict: dict|None = None + status_dict: Optional[dict] = None if status is not None: status_dict = copy.deepcopy(status._asdict()) diff --git a/folder_paths.py b/folder_paths.py index 641e5f0b2..234b73409 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -1,7 +1,8 @@ import os import time +import logging -supported_pt_extensions = set(['.ckpt', '.pt', '.bin', '.pth', '.safetensors']) +supported_pt_extensions = set(['.ckpt', '.pt', '.bin', '.pth', '.safetensors', '.pkl']) folder_names_and_paths = {} @@ -29,6 +30,8 @@ folder_names_and_paths["custom_nodes"] = ([os.path.join(base_path, "custom_nodes folder_names_and_paths["hypernetworks"] = ([os.path.join(models_dir, "hypernetworks")], supported_pt_extensions) +folder_names_and_paths["photomaker"] = ([os.path.join(models_dir, "photomaker")], supported_pt_extensions) + folder_names_and_paths["classifiers"] = ([os.path.join(models_dir, "classifiers")], {""}) output_directory = os.path.join(os.path.dirname(os.path.realpath(__file__)), "output") @@ -42,7 +45,7 @@ if not os.path.exists(input_directory): try: os.makedirs(input_directory) except: - print("Failed to create input directory") + logging.error("Failed to create input directory") def set_output_directory(output_dir): global output_directory @@ -138,15 +141,29 @@ def recursive_search(directory, excluded_dir_names=None): excluded_dir_names = [] result = [] - dirs = {directory: os.path.getmtime(directory)} + dirs = {} + + # Attempt to add the initial directory to dirs with error handling + try: + dirs[directory] = os.path.getmtime(directory) + except FileNotFoundError: + logging.warning(f"Warning: Unable to access {directory}. Skipping this path.") + + logging.debug("recursive file list on directory {}".format(directory)) for dirpath, subdirs, filenames in os.walk(directory, followlinks=True, topdown=True): subdirs[:] = [d for d in subdirs if d not in excluded_dir_names] for file_name in filenames: relative_path = os.path.relpath(os.path.join(dirpath, file_name), directory) result.append(relative_path) + for d in subdirs: path = os.path.join(dirpath, d) - dirs[path] = os.path.getmtime(path) + try: + dirs[path] = os.path.getmtime(path) + except FileNotFoundError: + logging.warning(f"Warning: Unable to access {path}. Skipping this path.") + continue + logging.debug("found {} files".format(len(result))) return result, dirs def filter_files_extensions(files, extensions): @@ -164,6 +181,8 @@ def get_full_path(folder_name, filename): full_path = os.path.join(x, filename) if os.path.isfile(full_path): return full_path + elif os.path.islink(full_path): + logging.warning("WARNING path {} exists but doesn't link anywhere, skipping.".format(full_path)) return None @@ -234,12 +253,12 @@ def get_save_image_path(filename_prefix, output_dir, image_width=0, image_height err = "**** ERROR: Saving image outside the output folder is not allowed." + \ "\n full_output_folder: " + os.path.abspath(full_output_folder) + \ "\n output_dir: " + output_dir + \ - "\n commonpath: " + os.path.commonpath((output_dir, os.path.abspath(full_output_folder))) - print(err) + "\n commonpath: " + os.path.commonpath((output_dir, os.path.abspath(full_output_folder))) + logging.error(err) raise Exception(err) try: - counter = max(filter(lambda a: a[1][:-1] == filename and a[1][-1] == "_", map(map_filename, os.listdir(full_output_folder))))[0] + 1 + counter = max(filter(lambda a: os.path.normcase(a[1][:-1]) == os.path.normcase(filename) and a[1][-1] == "_", map(map_filename, os.listdir(full_output_folder))))[0] + 1 except ValueError: counter = 1 except FileNotFoundError: diff --git a/latent_preview.py b/latent_preview.py index 61754751e..54aa233f2 100644 --- a/latent_preview.py +++ b/latent_preview.py @@ -4,11 +4,20 @@ import struct import numpy as np from comfy.cli_args import args, LatentPreviewMethod from comfy.taesd.taesd import TAESD +import comfy.model_management import folder_paths import comfy.utils +import logging MAX_PREVIEW_RESOLUTION = 512 +def preview_to_image(latent_image): + latents_ubyte = (((latent_image + 1.0) / 2.0).clamp(0, 1) # change scale from -1..1 to 0..1 + .mul(0xFF) # to 0..255 + ).to(device="cpu", dtype=torch.uint8, non_blocking=comfy.model_management.device_supports_non_blocking(latent_image.device)) + + return Image.fromarray(latents_ubyte.numpy()) + class LatentPreviewer: def decode_latent_to_preview(self, x0): pass @@ -22,13 +31,8 @@ class TAESDPreviewerImpl(LatentPreviewer): self.taesd = taesd def decode_latent_to_preview(self, x0): - x_sample = self.taesd.decode(x0[:1])[0].detach() - x_sample = torch.clamp((x_sample + 1.0) / 2.0, min=0.0, max=1.0) - x_sample = 255. * np.moveaxis(x_sample.cpu().numpy(), 0, 2) - x_sample = x_sample.astype(np.uint8) - - preview_image = Image.fromarray(x_sample) - return preview_image + x_sample = self.taesd.decode(x0[:1])[0].movedim(0, 2) + return preview_to_image(x_sample) class Latent2RGBPreviewer(LatentPreviewer): @@ -36,14 +40,9 @@ class Latent2RGBPreviewer(LatentPreviewer): self.latent_rgb_factors = torch.tensor(latent_rgb_factors, device="cpu") def decode_latent_to_preview(self, x0): - latent_image = x0[0].permute(1, 2, 0).cpu() @ self.latent_rgb_factors - - latents_ubyte = (((latent_image + 1) / 2) - .clamp(0, 1) # change scale from -1..1 to 0..1 - .mul(0xFF) # to 0..255 - .byte()).cpu() - - return Image.fromarray(latents_ubyte.numpy()) + self.latent_rgb_factors = self.latent_rgb_factors.to(dtype=x0.dtype, device=x0.device) + latent_image = x0[0].permute(1, 2, 0) @ self.latent_rgb_factors + return preview_to_image(latent_image) def get_previewer(device, latent_format): @@ -62,15 +61,13 @@ def get_previewer(device, latent_format): if method == LatentPreviewMethod.Auto: method = LatentPreviewMethod.Latent2RGB - if taesd_decoder_path: - method = LatentPreviewMethod.TAESD if method == LatentPreviewMethod.TAESD: if taesd_decoder_path: taesd = TAESD(None, taesd_decoder_path).to(device) previewer = TAESDPreviewerImpl(taesd) else: - print("Warning: TAESD previews enabled, but could not find models/vae_approx/{}".format(latent_format.taesd_decoder_name)) + logging.warning("Warning: TAESD previews enabled, but could not find models/vae_approx/{}".format(latent_format.taesd_decoder_name)) if previewer is None: if latent_format.latent_rgb_factors is not None: diff --git a/main.py b/main.py index 69d9bce6c..a374f2b12 100644 --- a/main.py +++ b/main.py @@ -54,15 +54,15 @@ import threading import gc from comfy.cli_args import args +import logging if os.name == "nt": - import logging logging.getLogger("xformers").addFilter(lambda record: 'A matching Triton is not available' not in record.getMessage()) if __name__ == "__main__": if args.cuda_device is not None: os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device) - print("Set cuda device to:", args.cuda_device) + logging.info("Set cuda device to: {}".format(args.cuda_device)) if args.deterministic: if 'CUBLAS_WORKSPACE_CONFIG' not in os.environ: @@ -88,7 +88,7 @@ def cuda_malloc_warning(): if b in device_name: cuda_malloc_warning = True if cuda_malloc_warning: - print("\nWARNING: this card most likely does not support cuda-malloc, if you get \"CUDA error\" please run ComfyUI with: --disable-cuda-malloc\n") + logging.warning("\nWARNING: this card most likely does not support cuda-malloc, if you get \"CUDA error\" please run ComfyUI with: --disable-cuda-malloc\n") def prompt_worker(q, server): e = execution.PromptExecutor(server) @@ -121,7 +121,7 @@ def prompt_worker(q, server): current_time = time.perf_counter() execution_time = current_time - execution_start_time - print("Prompt executed in {:.2f} seconds".format(execution_time)) + logging.info("Prompt executed in {:.2f} seconds".format(execution_time)) flags = q.get_flags() free_memory = flags.get("free_memory", False) @@ -139,6 +139,7 @@ def prompt_worker(q, server): if need_gc: current_time = time.perf_counter() if (current_time - last_gc_collect) > gc_collect_interval: + comfy.model_management.cleanup_models() gc.collect() comfy.model_management.soft_empty_cache() last_gc_collect = current_time @@ -182,17 +183,24 @@ def load_extra_path_config(yaml_path): full_path = y if base_path is not None: full_path = os.path.join(base_path, full_path) - print("Adding extra search path", x, full_path) + logging.info("Adding extra search path {} {}".format(x, full_path)) folder_paths.add_model_folder_path(x, full_path) if __name__ == "__main__": if args.temp_directory: temp_dir = os.path.join(os.path.abspath(args.temp_directory), "temp") - print(f"Setting temp directory to: {temp_dir}") + logging.info(f"Setting temp directory to: {temp_dir}") folder_paths.set_temp_directory(temp_dir) cleanup_temp() + if args.windows_standalone_build: + try: + import new_updater + new_updater.update_windows_updater() + except: + pass + loop = asyncio.new_event_loop() asyncio.set_event_loop(loop) server = server.PromptServer(loop) @@ -217,7 +225,7 @@ if __name__ == "__main__": if args.output_directory: output_dir = os.path.abspath(args.output_directory) - print(f"Setting output directory to: {output_dir}") + logging.info(f"Setting output directory to: {output_dir}") folder_paths.set_output_directory(output_dir) #These are the default folders that checkpoints, clip and vae models will be saved to when using CheckpointSave, etc.. nodes @@ -227,7 +235,7 @@ if __name__ == "__main__": if args.input_directory: input_dir = os.path.abspath(args.input_directory) - print(f"Setting input directory to: {input_dir}") + logging.info(f"Setting input directory to: {input_dir}") folder_paths.set_input_directory(input_dir) if args.quick_test_for_ci: @@ -235,16 +243,16 @@ if __name__ == "__main__": call_on_start = None if args.auto_launch: - def startup_server(address, port): + def startup_server(scheme, address, port): import webbrowser if os.name == 'nt' and address == '0.0.0.0': address = '127.0.0.1' - webbrowser.open(f"http://{address}:{port}") + webbrowser.open(f"{scheme}://{address}:{port}") call_on_start = startup_server try: loop.run_until_complete(run(server, address=args.listen, port=args.port, verbose=not args.dont_print_server, call_on_start=call_on_start)) except KeyboardInterrupt: - print("\nStopped server") + logging.info("\nStopped server") cleanup_temp() diff --git a/comfy_extras/chainner_models/__init__.py b/models/photomaker/put_photomaker_models_here similarity index 100% rename from comfy_extras/chainner_models/__init__.py rename to models/photomaker/put_photomaker_models_here diff --git a/new_updater.py b/new_updater.py new file mode 100644 index 000000000..a49e0877c --- /dev/null +++ b/new_updater.py @@ -0,0 +1,35 @@ +import os +import shutil + +base_path = os.path.dirname(os.path.realpath(__file__)) + + +def update_windows_updater(): + top_path = os.path.dirname(base_path) + updater_path = os.path.join(base_path, ".ci/update_windows/update.py") + bat_path = os.path.join(base_path, ".ci/update_windows/update_comfyui.bat") + + dest_updater_path = os.path.join(top_path, "update/update.py") + dest_bat_path = os.path.join(top_path, "update/update_comfyui.bat") + dest_bat_deps_path = os.path.join(top_path, "update/update_comfyui_and_python_dependencies.bat") + + try: + with open(dest_bat_path, 'rb') as f: + contents = f.read() + except: + return + + if not contents.startswith(b"..\\python_embeded\\python.exe .\\update.py"): + return + + shutil.copy(updater_path, dest_updater_path) + try: + with open(dest_bat_deps_path, 'rb') as f: + contents = f.read() + contents = contents.replace(b'..\\python_embeded\\python.exe .\\update.py ..\\ComfyUI\\', b'call update_comfyui.bat nopause') + with open(dest_bat_deps_path, 'wb') as f: + f.write(contents) + except: + pass + shutil.copy(bat_path, dest_bat_path) + print("Updated the windows standalone package updater.") diff --git a/node_helpers.py b/node_helpers.py new file mode 100644 index 000000000..43b9e829f --- /dev/null +++ b/node_helpers.py @@ -0,0 +1,24 @@ +from PIL import ImageFile, UnidentifiedImageError + +def conditioning_set_values(conditioning, values={}): + c = [] + for t in conditioning: + n = [t[0], t[1].copy()] + for k in values: + n[1][k] = values[k] + c.append(n) + + return c + +def pillow(fn, arg): + prev_value = None + try: + x = fn(arg) + except (OSError, UnidentifiedImageError, ValueError): #PIL issues #4472 and #2445, also fixes ComfyUI issue #3416 + prev_value = ImageFile.LOAD_TRUNCATED_IMAGES + ImageFile.LOAD_TRUNCATED_IMAGES = True + x = fn(arg) + finally: + if prev_value is not None: + ImageFile.LOAD_TRUNCATED_IMAGES = prev_value + return x diff --git a/nodes.py b/nodes.py index 6c7317b69..34821ca3f 100644 --- a/nodes.py +++ b/nodes.py @@ -8,15 +8,16 @@ import traceback import math import time import random +import logging -from PIL import Image, ImageOps, ImageSequence +from PIL import Image, ImageOps, ImageSequence, ImageFile from PIL.PngImagePlugin import PngInfo + import numpy as np import safetensors.torch sys.path.insert(0, os.path.join(os.path.dirname(os.path.realpath(__file__)), "comfy")) - import comfy.diffusers_load import comfy.samplers import comfy.sample @@ -33,6 +34,7 @@ import importlib import folder_paths import latent_preview +import node_helpers def before_node_execution(): comfy.model_management.throw_exception_if_processing_interrupted() @@ -40,12 +42,12 @@ def before_node_execution(): def interrupt_processing(value=True): comfy.model_management.interrupt_current_processing(value) -MAX_RESOLUTION=8192 +MAX_RESOLUTION=16384 class CLIPTextEncode: @classmethod def INPUT_TYPES(s): - return {"required": {"text": ("STRING", {"multiline": True}), "clip": ("CLIP", )}} + return {"required": {"text": ("STRING", {"multiline": True, "dynamicPrompts": True}), "clip": ("CLIP", )}} RETURN_TYPES = ("CONDITIONING",) FUNCTION = "encode" @@ -83,7 +85,7 @@ class ConditioningAverage : out = [] if len(conditioning_from) > 1: - print("Warning: ConditioningAverage conditioning_from contains more than 1 cond, only the first one will actually be applied to conditioning_to.") + logging.warning("Warning: ConditioningAverage conditioning_from contains more than 1 cond, only the first one will actually be applied to conditioning_to.") cond_from = conditioning_from[0][0] pooled_output_from = conditioning_from[0][1].get("pooled_output", None) @@ -122,7 +124,7 @@ class ConditioningConcat: out = [] if len(conditioning_from) > 1: - print("Warning: ConditioningConcat conditioning_from contains more than 1 cond, only the first one will actually be applied to conditioning_to.") + logging.warning("Warning: ConditioningConcat conditioning_from contains more than 1 cond, only the first one will actually be applied to conditioning_to.") cond_from = conditioning_from[0][0] @@ -150,13 +152,9 @@ class ConditioningSetArea: CATEGORY = "conditioning" def append(self, conditioning, width, height, x, y, strength): - c = [] - for t in conditioning: - n = [t[0], t[1].copy()] - n[1]['area'] = (height // 8, width // 8, y // 8, x // 8) - n[1]['strength'] = strength - n[1]['set_area_to_bounds'] = False - c.append(n) + c = node_helpers.conditioning_set_values(conditioning, {"area": (height // 8, width // 8, y // 8, x // 8), + "strength": strength, + "set_area_to_bounds": False}) return (c, ) class ConditioningSetAreaPercentage: @@ -175,15 +173,27 @@ class ConditioningSetAreaPercentage: CATEGORY = "conditioning" def append(self, conditioning, width, height, x, y, strength): - c = [] - for t in conditioning: - n = [t[0], t[1].copy()] - n[1]['area'] = ("percentage", height, width, y, x) - n[1]['strength'] = strength - n[1]['set_area_to_bounds'] = False - c.append(n) + c = node_helpers.conditioning_set_values(conditioning, {"area": ("percentage", height, width, y, x), + "strength": strength, + "set_area_to_bounds": False}) return (c, ) +class ConditioningSetAreaStrength: + @classmethod + def INPUT_TYPES(s): + return {"required": {"conditioning": ("CONDITIONING", ), + "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + }} + RETURN_TYPES = ("CONDITIONING",) + FUNCTION = "append" + + CATEGORY = "conditioning" + + def append(self, conditioning, strength): + c = node_helpers.conditioning_set_values(conditioning, {"strength": strength}) + return (c, ) + + class ConditioningSetMask: @classmethod def INPUT_TYPES(s): @@ -198,19 +208,15 @@ class ConditioningSetMask: CATEGORY = "conditioning" def append(self, conditioning, mask, set_cond_area, strength): - c = [] set_area_to_bounds = False if set_cond_area != "default": set_area_to_bounds = True if len(mask.shape) < 3: mask = mask.unsqueeze(0) - for t in conditioning: - n = [t[0], t[1].copy()] - _, h, w = mask.shape - n[1]['mask'] = mask - n[1]['set_area_to_bounds'] = set_area_to_bounds - n[1]['mask_strength'] = strength - c.append(n) + + c = node_helpers.conditioning_set_values(conditioning, {"mask": mask, + "set_area_to_bounds": set_area_to_bounds, + "mask_strength": strength}) return (c, ) class ConditioningZeroOut: @@ -245,13 +251,8 @@ class ConditioningSetTimestepRange: CATEGORY = "advanced/conditioning" def set_range(self, conditioning, start, end): - c = [] - for t in conditioning: - d = t[1].copy() - d['start_percent'] = start - d['end_percent'] = end - n = [t[0], d] - c.append(n) + c = node_helpers.conditioning_set_values(conditioning, {"start_percent": start, + "end_percent": end}) return (c, ) class VAEDecode: @@ -289,18 +290,7 @@ class VAEEncode: CATEGORY = "latent" - @staticmethod - def vae_encode_crop_pixels(pixels): - x = (pixels.shape[1] // 8) * 8 - y = (pixels.shape[2] // 8) * 8 - if pixels.shape[1] != x or pixels.shape[2] != y: - x_offset = (pixels.shape[1] % 8) // 2 - y_offset = (pixels.shape[2] % 8) // 2 - pixels = pixels[:, x_offset:x + x_offset, y_offset:y + y_offset, :] - return pixels - def encode(self, vae, pixels): - pixels = self.vae_encode_crop_pixels(pixels) t = vae.encode(pixels[:,:,:,:3]) return ({"samples":t}, ) @@ -316,7 +306,6 @@ class VAEEncodeTiled: CATEGORY = "_for_testing" def encode(self, vae, pixels, tile_size): - pixels = VAEEncode.vae_encode_crop_pixels(pixels) t = vae.encode_tiled(pixels[:,:,:,:3], tile_x=tile_size, tile_y=tile_size, ) return ({"samples":t}, ) @@ -330,14 +319,14 @@ class VAEEncodeForInpaint: CATEGORY = "latent/inpaint" def encode(self, vae, pixels, mask, grow_mask_by=6): - x = (pixels.shape[1] // 8) * 8 - y = (pixels.shape[2] // 8) * 8 + x = (pixels.shape[1] // vae.downscale_ratio) * vae.downscale_ratio + y = (pixels.shape[2] // vae.downscale_ratio) * vae.downscale_ratio mask = torch.nn.functional.interpolate(mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])), size=(pixels.shape[1], pixels.shape[2]), mode="bilinear") pixels = pixels.clone() if pixels.shape[1] != x or pixels.shape[2] != y: - x_offset = (pixels.shape[1] % 8) // 2 - y_offset = (pixels.shape[2] % 8) // 2 + x_offset = (pixels.shape[1] % vae.downscale_ratio) // 2 + y_offset = (pixels.shape[2] % vae.downscale_ratio) // 2 pixels = pixels[:,x_offset:x + x_offset, y_offset:y + y_offset,:] mask = mask[:,:,x_offset:x + x_offset, y_offset:y + y_offset] @@ -404,13 +393,8 @@ class InpaintModelConditioning: out = [] for conditioning in [positive, negative]: - c = [] - for t in conditioning: - d = t[1].copy() - d["concat_latent_image"] = concat_latent - d["concat_mask"] = mask - n = [t[0], d] - c.append(n) + c = node_helpers.conditioning_set_values(conditioning, {"concat_latent_image": concat_latent, + "concat_mask": mask}) out.append(c) return (out[0], out[1], out_latent) @@ -599,8 +583,8 @@ class LoraLoader: return {"required": { "model": ("MODEL",), "clip": ("CLIP", ), "lora_name": (folder_paths.get_filename_list("loras"), ), - "strength_model": ("FLOAT", {"default": 1.0, "min": -20.0, "max": 20.0, "step": 0.01}), - "strength_clip": ("FLOAT", {"default": 1.0, "min": -20.0, "max": 20.0, "step": 0.01}), + "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}), + "strength_clip": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}), }} RETURN_TYPES = ("MODEL", "CLIP") FUNCTION = "load_lora" @@ -633,7 +617,7 @@ class LoraLoaderModelOnly(LoraLoader): def INPUT_TYPES(s): return {"required": { "model": ("MODEL",), "lora_name": (folder_paths.get_filename_list("loras"), ), - "strength_model": ("FLOAT", {"default": 1.0, "min": -20.0, "max": 20.0, "step": 0.01}), + "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}), }} RETURN_TYPES = ("MODEL",) FUNCTION = "load_lora_model_only" @@ -834,15 +818,20 @@ class CLIPLoader: @classmethod def INPUT_TYPES(s): return {"required": { "clip_name": (folder_paths.get_filename_list("clip"), ), + "type": (["stable_diffusion", "stable_cascade"], ), }} RETURN_TYPES = ("CLIP",) FUNCTION = "load_clip" CATEGORY = "advanced/loaders" - def load_clip(self, clip_name): + def load_clip(self, clip_name, type="stable_diffusion"): + clip_type = comfy.sd.CLIPType.STABLE_DIFFUSION + if type == "stable_cascade": + clip_type = comfy.sd.CLIPType.STABLE_CASCADE + clip_path = folder_paths.get_full_path("clip", clip_name) - clip = comfy.sd.load_clip(ckpt_paths=[clip_path], embedding_directory=folder_paths.get_folder_paths("embeddings")) + clip = comfy.sd.load_clip(ckpt_paths=[clip_path], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type) return (clip,) class DualCLIPLoader: @@ -977,7 +966,7 @@ class GLIGENTextBoxApply: return {"required": {"conditioning_to": ("CONDITIONING", ), "clip": ("CLIP", ), "gligen_textbox_model": ("GLIGEN", ), - "text": ("STRING", {"multiline": True}), + "text": ("STRING", {"multiline": True, "dynamicPrompts": True}), "width": ("INT", {"default": 64, "min": 8, "max": MAX_RESOLUTION, "step": 8}), "height": ("INT", {"default": 64, "min": 8, "max": MAX_RESOLUTION, "step": 8}), "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 8}), @@ -990,7 +979,7 @@ class GLIGENTextBoxApply: def append(self, conditioning_to, clip, gligen_textbox_model, text, width, height, x, y): c = [] - cond, cond_pooled = clip.encode_from_tokens(clip.tokenize(text), return_pooled=True) + cond, cond_pooled = clip.encode_from_tokens(clip.tokenize(text), return_pooled="unprojected") for t in conditioning_to: n = [t[0], t[1].copy()] position_params = [(cond_pooled, height // 8, width // 8, y // 8, x // 8)] @@ -1414,7 +1403,7 @@ class SaveImage: filename_prefix += self.prefix_append full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir, images[0].shape[1], images[0].shape[0]) results = list() - for image in images: + for (batch_number, image) in enumerate(images): i = 255. * image.cpu().numpy() img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8)) metadata = None @@ -1426,7 +1415,8 @@ class SaveImage: for x in extra_pnginfo: metadata.add_text(x, json.dumps(extra_pnginfo[x])) - file = f"{filename}_{counter:05}_.png" + filename_with_batch_num = filename.replace("%batch_num%", str(batch_number)) + file = f"{filename_with_batch_num}_{counter:05}_.png" img.save(os.path.join(full_output_folder, file), pnginfo=metadata, compress_level=self.compress_level) results.append({ "filename": file, @@ -1466,14 +1456,29 @@ class LoadImage: FUNCTION = "load_image" def load_image(self, image): image_path = folder_paths.get_annotated_filepath(image) - img = Image.open(image_path) + + img = node_helpers.pillow(Image.open, image_path) + output_images = [] output_masks = [] + w, h = None, None + + excluded_formats = ['MPO'] + for i in ImageSequence.Iterator(img): - i = ImageOps.exif_transpose(i) + i = node_helpers.pillow(ImageOps.exif_transpose, i) + if i.mode == 'I': i = i.point(lambda i: i * (1 / 255)) image = i.convert("RGB") + + if len(output_images) == 0: + w = image.size[0] + h = image.size[1] + + if image.size[0] != w or image.size[1] != h: + continue + image = np.array(image).astype(np.float32) / 255.0 image = torch.from_numpy(image)[None,] if 'A' in i.getbands(): @@ -1484,7 +1489,7 @@ class LoadImage: output_images.append(image) output_masks.append(mask.unsqueeze(0)) - if len(output_images) > 1: + if len(output_images) > 1 and img.format not in excluded_formats: output_image = torch.cat(output_images, dim=0) output_mask = torch.cat(output_masks, dim=0) else: @@ -1525,8 +1530,8 @@ class LoadImageMask: FUNCTION = "load_image" def load_image(self, image, channel): image_path = folder_paths.get_annotated_filepath(image) - i = Image.open(image_path) - i = ImageOps.exif_transpose(i) + i = node_helpers.pillow(Image.open, image_path) + i = node_helpers.pillow(ImageOps.exif_transpose, i) if i.getbands() != ("R", "G", "B", "A"): if i.mode == 'I': i = i.point(lambda i: i * (1 / 255)) @@ -1754,6 +1759,7 @@ NODE_CLASS_MAPPINGS = { "ConditioningConcat": ConditioningConcat, "ConditioningSetArea": ConditioningSetArea, "ConditioningSetAreaPercentage": ConditioningSetAreaPercentage, + "ConditioningSetAreaStrength": ConditioningSetAreaStrength, "ConditioningSetMask": ConditioningSetMask, "KSamplerAdvanced": KSamplerAdvanced, "SetLatentNoiseMask": SetLatentNoiseMask, @@ -1860,6 +1866,7 @@ def load_custom_node(module_path, ignore=set()): sp = os.path.splitext(module_path) module_name = sp[0] try: + logging.debug("Trying to load custom node {}".format(module_path)) if os.path.isfile(module_path): module_spec = importlib.util.spec_from_file_location(module_name, module_path) module_dir = os.path.split(module_path)[0] @@ -1884,11 +1891,11 @@ def load_custom_node(module_path, ignore=set()): NODE_DISPLAY_NAME_MAPPINGS.update(module.NODE_DISPLAY_NAME_MAPPINGS) return True else: - print(f"Skip {module_path} module for custom nodes due to the lack of NODE_CLASS_MAPPINGS.") + logging.warning(f"Skip {module_path} module for custom nodes due to the lack of NODE_CLASS_MAPPINGS.") return False except Exception as e: - print(traceback.format_exc()) - print(f"Cannot import {module_path} module for custom nodes:", e) + logging.warning(traceback.format_exc()) + logging.warning(f"Cannot import {module_path} module for custom nodes: {e}") return False def load_custom_nodes(): @@ -1909,14 +1916,14 @@ def load_custom_nodes(): node_import_times.append((time.perf_counter() - time_before, module_path, success)) if len(node_import_times) > 0: - print("\nImport times for custom nodes:") + logging.info("\nImport times for custom nodes:") for n in sorted(node_import_times): if n[2]: import_message = "" else: import_message = " (IMPORT FAILED)" - print("{:6.1f} seconds{}:".format(n[0], import_message), n[1]) - print() + logging.info("{:6.1f} seconds{}: {}".format(n[0], import_message, n[1])) + logging.info("") def init_custom_nodes(): extras_dir = os.path.join(os.path.dirname(os.path.realpath(__file__)), "comfy_extras") @@ -1943,9 +1950,34 @@ def init_custom_nodes(): "nodes_perpneg.py", "nodes_stable3d.py", "nodes_sdupscale.py", + "nodes_photomaker.py", + "nodes_cond.py", + "nodes_morphology.py", + "nodes_stable_cascade.py", + "nodes_differential_diffusion.py", + "nodes_ip2p.py", + "nodes_model_merging_model_specific.py", + "nodes_pag.py", + "nodes_align_your_steps.py", + "nodes_attention_multiply.py", + "nodes_advanced_samplers.py", + "nodes_webcam.py", ] + import_failed = [] for node_file in extras_files: - load_custom_node(os.path.join(extras_dir, node_file)) + if not load_custom_node(os.path.join(extras_dir, node_file)): + import_failed.append(node_file) load_custom_nodes() + + if len(import_failed) > 0: + logging.warning("WARNING: some comfy_extras/ nodes did not import correctly. This may be because they are missing some dependencies.\n") + for node in import_failed: + logging.warning("IMPORT FAILED: {}".format(node)) + logging.warning("\nThis issue might be caused by new missing dependencies added the last time you updated ComfyUI.") + if args.windows_standalone_build: + logging.warning("Please run the update script: update/update_comfyui.bat") + else: + logging.warning("Please do a: pip install -r requirements.txt") + logging.warning("") diff --git a/requirements.txt b/requirements.txt index e804618e7..906b96eda 100644 --- a/requirements.txt +++ b/requirements.txt @@ -10,3 +10,7 @@ Pillow scipy tqdm psutil + +#non essential dependencies: +kornia>=0.7.1 +spandrel diff --git a/script_examples/websockets_api_example_ws_images.py b/script_examples/websockets_api_example_ws_images.py new file mode 100644 index 000000000..737488621 --- /dev/null +++ b/script_examples/websockets_api_example_ws_images.py @@ -0,0 +1,159 @@ +#This is an example that uses the websockets api and the SaveImageWebsocket node to get images directly without +#them being saved to disk + +import websocket #NOTE: websocket-client (https://github.com/websocket-client/websocket-client) +import uuid +import json +import urllib.request +import urllib.parse + +server_address = "127.0.0.1:8188" +client_id = str(uuid.uuid4()) + +def queue_prompt(prompt): + p = {"prompt": prompt, "client_id": client_id} + data = json.dumps(p).encode('utf-8') + req = urllib.request.Request("http://{}/prompt".format(server_address), data=data) + return json.loads(urllib.request.urlopen(req).read()) + +def get_image(filename, subfolder, folder_type): + data = {"filename": filename, "subfolder": subfolder, "type": folder_type} + url_values = urllib.parse.urlencode(data) + with urllib.request.urlopen("http://{}/view?{}".format(server_address, url_values)) as response: + return response.read() + +def get_history(prompt_id): + with urllib.request.urlopen("http://{}/history/{}".format(server_address, prompt_id)) as response: + return json.loads(response.read()) + +def get_images(ws, prompt): + prompt_id = queue_prompt(prompt)['prompt_id'] + output_images = {} + current_node = "" + while True: + out = ws.recv() + if isinstance(out, str): + message = json.loads(out) + if message['type'] == 'executing': + data = message['data'] + if data['prompt_id'] == prompt_id: + if data['node'] is None: + break #Execution is done + else: + current_node = data['node'] + else: + if current_node == 'save_image_websocket_node': + images_output = output_images.get(current_node, []) + images_output.append(out[8:]) + output_images[current_node] = images_output + + return output_images + +prompt_text = """ +{ + "3": { + "class_type": "KSampler", + "inputs": { + "cfg": 8, + "denoise": 1, + "latent_image": [ + "5", + 0 + ], + "model": [ + "4", + 0 + ], + "negative": [ + "7", + 0 + ], + "positive": [ + "6", + 0 + ], + "sampler_name": "euler", + "scheduler": "normal", + "seed": 8566257, + "steps": 20 + } + }, + "4": { + "class_type": "CheckpointLoaderSimple", + "inputs": { + "ckpt_name": "v1-5-pruned-emaonly.ckpt" + } + }, + "5": { + "class_type": "EmptyLatentImage", + "inputs": { + "batch_size": 1, + "height": 512, + "width": 512 + } + }, + "6": { + "class_type": "CLIPTextEncode", + "inputs": { + "clip": [ + "4", + 1 + ], + "text": "masterpiece best quality girl" + } + }, + "7": { + "class_type": "CLIPTextEncode", + "inputs": { + "clip": [ + "4", + 1 + ], + "text": "bad hands" + } + }, + "8": { + "class_type": "VAEDecode", + "inputs": { + "samples": [ + "3", + 0 + ], + "vae": [ + "4", + 2 + ] + } + }, + "save_image_websocket_node": { + "class_type": "SaveImageWebsocket", + "inputs": { + "images": [ + "8", + 0 + ] + } + } +} +""" + +prompt = json.loads(prompt_text) +#set the text prompt for our positive CLIPTextEncode +prompt["6"]["inputs"]["text"] = "masterpiece best quality man" + +#set the seed for our KSampler node +prompt["3"]["inputs"]["seed"] = 5 + +ws = websocket.WebSocket() +ws.connect("ws://{}/ws?clientId={}".format(server_address, client_id)) +images = get_images(ws, prompt) + +#Commented out code to display the output images: + +# for node_id in images: +# for image_data in images[node_id]: +# from PIL import Image +# import io +# image = Image.open(io.BytesIO(image_data)) +# image.show() + diff --git a/server.py b/server.py index 1135fff88..bab3b0600 100644 --- a/server.py +++ b/server.py @@ -11,19 +11,14 @@ import urllib import json import glob import struct +import ssl from PIL import Image, ImageOps from PIL.PngImagePlugin import PngInfo from io import BytesIO -try: - import aiohttp - from aiohttp import web -except ImportError: - print("Module 'aiohttp' not installed. Please install it via:") - print("pip install aiohttp") - print("or") - print("pip install -r requirements.txt") - sys.exit() +import aiohttp +from aiohttp import web +import logging import mimetypes from comfy.cli_args import args @@ -40,7 +35,7 @@ async def send_socket_catch_exception(function, message): try: await function(message) except (aiohttp.ClientError, aiohttp.ClientPayloadError, ConnectionResetError) as err: - print("send error:", err) + logging.warning("send error: {}".format(err)) @web.middleware async def cache_control(request: web.Request, handler): @@ -118,7 +113,7 @@ class PromptServer(): async for msg in ws: if msg.type == aiohttp.WSMsgType.ERROR: - print('ws connection closed with exception %s' % ws.exception()) + logging.warning('ws connection closed with exception %s' % ws.exception()) finally: self.sockets.pop(sid, None) return ws @@ -419,8 +414,8 @@ class PromptServer(): try: out[x] = node_info(x) except Exception as e: - print(f"[ERROR] An error occurred while retrieving information for the '{x}' node.", file=sys.stderr) - traceback.print_exc() + logging.error(f"[ERROR] An error occurred while retrieving information for the '{x}' node.") + logging.error(traceback.format_exc()) return web.json_response(out) @routes.get("/object_info/{node_class}") @@ -453,7 +448,7 @@ class PromptServer(): @routes.post("/prompt") async def post_prompt(request): - print("got prompt") + logging.info("got prompt") resp_code = 200 out_string = "" json_data = await request.json() @@ -485,7 +480,7 @@ class PromptServer(): response = {"prompt_id": prompt_id, "number": number, "node_errors": valid[3]} return web.json_response(response) else: - print("invalid prompt:", valid[1]) + logging.warning("invalid prompt: {}".format(valid[1])) return web.json_response({"error": valid[1], "node_errors": valid[3]}, status=400) else: return web.json_response({"error": "no prompt", "node_errors": []}, status=400) @@ -539,11 +534,11 @@ class PromptServer(): for name, dir in nodes.EXTENSION_WEB_DIRS.items(): self.app.add_routes([ - web.static('/extensions/' + urllib.parse.quote(name), dir, follow_symlinks=True), + web.static('/extensions/' + urllib.parse.quote(name), dir), ]) self.app.add_routes([ - web.static('/', self.web_root, follow_symlinks=True), + web.static('/', self.web_root), ]) def get_queue_info(self): @@ -629,16 +624,22 @@ class PromptServer(): async def start(self, address, port, verbose=True, call_on_start=None): runner = web.AppRunner(self.app, access_log=None) await runner.setup() - site = web.TCPSite(runner, address, port) + ssl_ctx = None + scheme = "http" + if args.tls_keyfile and args.tls_certfile: + ssl_ctx = ssl.SSLContext(protocol=ssl.PROTOCOL_TLS_SERVER, verify_mode=ssl.CERT_NONE) + ssl_ctx.load_cert_chain(certfile=args.tls_certfile, + keyfile=args.tls_keyfile) + scheme = "https" + + site = web.TCPSite(runner, address, port, ssl_context=ssl_ctx) await site.start() - if address == '': - address = '0.0.0.0' if verbose: - print("Starting server\n") - print("To see the GUI go to: http://{}:{}".format(address, port)) + logging.info("Starting server\n") + logging.info("To see the GUI go to: {}://{}:{}".format(scheme, address, port)) if call_on_start is not None: - call_on_start(address, port) + call_on_start(scheme, address, port) def add_on_prompt_handler(self, handler): self.on_prompt_handlers.append(handler) @@ -648,7 +649,7 @@ class PromptServer(): try: json_data = handler(json_data) except Exception as e: - print(f"[ERROR] An error occurred during the on_prompt_handler processing") - traceback.print_exc() + logging.warning(f"[ERROR] An error occurred during the on_prompt_handler processing") + logging.warning(traceback.format_exc()) return json_data diff --git a/tests-ui/tests/groupNode.test.js b/tests-ui/tests/groupNode.test.js index e6ebedd91..53a5828d1 100644 --- a/tests-ui/tests/groupNode.test.js +++ b/tests-ui/tests/groupNode.test.js @@ -947,7 +947,7 @@ describe("group node", () => { expect(p1.widgets.value.widget.options?.step).toBe(80); // width/height step * 10 expect(p2.widgets.value.widget.options?.min).toBe(16); // width/height min - expect(p2.widgets.value.widget.options?.max).toBe(8192); // width/height max + expect(p2.widgets.value.widget.options?.max).toBe(16384); // width/height max expect(p2.widgets.value.widget.options?.step).toBe(80); // width/height step * 10 expect(p1.widgets.value.value).toBe(128); diff --git a/tests-ui/utils/ezgraph.js b/tests-ui/utils/ezgraph.js index 8a55246ee..97be7aa72 100644 --- a/tests-ui/utils/ezgraph.js +++ b/tests-ui/utils/ezgraph.js @@ -204,13 +204,17 @@ export class EzWidget { convertToWidget() { if (!this.isConvertedToInput) throw new Error(`Widget ${this.widget.name} cannot be converted as it is already a widget.`); - this.node.menu[`Convert ${this.widget.name} to widget`].call(); + var menu = this.node.menu["Convert Input to Widget"].item.submenu.options; + var index = menu.findIndex(a => a.content == `Convert ${this.widget.name} to widget`); + menu[index].callback.call(); } convertToInput() { if (this.isConvertedToInput) throw new Error(`Widget ${this.widget.name} cannot be converted as it is already an input.`); - this.node.menu[`Convert ${this.widget.name} to input`].call(); + var menu = this.node.menu["Convert Widget to Input"].item.submenu.options; + var index = menu.findIndex(a => a.content == `Convert ${this.widget.name} to input`); + menu[index].callback.call(); } } diff --git a/web/extensions/core/colorPalette.js b/web/extensions/core/colorPalette.js index b8d83613d..02546782f 100644 --- a/web/extensions/core/colorPalette.js +++ b/web/extensions/core/colorPalette.js @@ -20,6 +20,10 @@ const colorPalettes = { "MODEL": "#B39DDB", // light lavender-purple "STYLE_MODEL": "#C2FFAE", // light green-yellow "VAE": "#FF6E6E", // bright red + "NOISE": "#B0B0B0", // gray + "GUIDER": "#66FFFF", // cyan + "SAMPLER": "#ECB4B4", // very soft red + "SIGMAS": "#CDFFCD", // soft lime green "TAESD": "#DCC274", // cheesecake }, "litegraph_base": { diff --git a/web/extensions/core/dynamicPrompts.js b/web/extensions/core/dynamicPrompts.js index 599a9e685..7417361ba 100644 --- a/web/extensions/core/dynamicPrompts.js +++ b/web/extensions/core/dynamicPrompts.js @@ -17,7 +17,7 @@ app.registerExtension({ // Locate dynamic prompt text widgets // Include any widgets with dynamicPrompts set to true, and customtext const widgets = node.widgets.filter( - (n) => (n.type === "customtext" && n.dynamicPrompts !== false) || n.dynamicPrompts + (n) => n.dynamicPrompts ); for (const widget of widgets) { // Override the serialization of the value to resolve dynamic prompts for all widgets supporting it in this node diff --git a/web/extensions/core/groupNode.js b/web/extensions/core/groupNode.js index 18b0e6aa6..0b0763d1d 100644 --- a/web/extensions/core/groupNode.js +++ b/web/extensions/core/groupNode.js @@ -349,7 +349,7 @@ export class GroupNodeConfig { } if (config[0] === "IMAGEUPLOAD") { if (!extra) extra = {}; - extra.widget = `${prefix}${config[1]?.widget ?? "image"}`; + extra.widget = this.oldToNewWidgetMap[node.index]?.[config[1]?.widget ?? "image"] ?? "image"; } if (extra) { @@ -910,6 +910,9 @@ export class GroupNodeHandler { const self = this; const onNodeCreated = this.node.onNodeCreated; this.node.onNodeCreated = function () { + if (!this.widgets) { + return; + } const config = self.groupData.nodeData.config; if (config) { for (const n in config) { @@ -966,6 +969,26 @@ export class GroupNodeHandler { api.removeEventListener("executing", executing); api.removeEventListener("executed", executed); }; + + this.node.refreshComboInNode = (defs) => { + // Update combo widget options + for (const widgetName in this.groupData.newToOldWidgetMap) { + const widget = this.node.widgets.find((w) => w.name === widgetName); + if (widget?.type === "combo") { + const old = this.groupData.newToOldWidgetMap[widgetName]; + const def = defs[old.node.type]; + const input = def?.input?.required?.[old.inputName] ?? def?.input?.optional?.[old.inputName]; + if (!input) continue; + + widget.options.values = input[0]; + + if (old.inputName !== "image" && !widget.options.values.includes(widget.value)) { + widget.value = widget.options.values[0]; + widget.callback(widget.value); + } + } + } + }; } updateInnerWidgets() { @@ -1245,6 +1268,14 @@ const ext = { node[GROUP] = new GroupNodeHandler(node); } }, + async refreshComboInNodes(defs) { + // Re-register group nodes so new ones are created with the correct options + Object.assign(globalDefs, defs); + const nodes = app.graph.extra?.groupNodes; + if (nodes) { + await GroupNodeConfig.registerFromWorkflow(nodes, {}); + } + } }; app.registerExtension(ext); diff --git a/web/extensions/core/groupNodeManage.css b/web/extensions/core/groupNodeManage.css index 5ac89aee3..5470ecb5e 100644 --- a/web/extensions/core/groupNodeManage.css +++ b/web/extensions/core/groupNodeManage.css @@ -48,7 +48,7 @@ list-style: none; } .comfy-group-manage-list-items { - max-height: 70vh; + max-height: calc(100% - 40px); overflow-y: scroll; overflow-x: hidden; } diff --git a/web/extensions/core/keybinds.js b/web/extensions/core/keybinds.js index e1906ee64..55d2e775d 100644 --- a/web/extensions/core/keybinds.js +++ b/web/extensions/core/keybinds.js @@ -36,7 +36,6 @@ app.registerExtension({ s: "#comfy-save-button", o: "#comfy-file-input", Backspace: "#comfy-clear-button", - Delete: "#comfy-clear-button", d: "#comfy-load-default-button", }; diff --git a/web/extensions/core/maskeditor.js b/web/extensions/core/maskeditor.js index bb2f16d42..36f7496e7 100644 --- a/web/extensions/core/maskeditor.js +++ b/web/extensions/core/maskeditor.js @@ -62,7 +62,7 @@ async function uploadMask(filepath, formData) { ClipspaceDialog.invalidatePreview(); } -function prepare_mask(image, maskCanvas, maskCtx) { +function prepare_mask(image, maskCanvas, maskCtx, maskColor) { // paste mask data into alpha channel maskCtx.drawImage(image, 0, 0, maskCanvas.width, maskCanvas.height); const maskData = maskCtx.getImageData(0, 0, maskCanvas.width, maskCanvas.height); @@ -74,9 +74,9 @@ function prepare_mask(image, maskCanvas, maskCtx) { else maskData.data[i+3] = 255; - maskData.data[i] = 0; - maskData.data[i+1] = 0; - maskData.data[i+2] = 0; + maskData.data[i] = maskColor.r; + maskData.data[i+1] = maskColor.g; + maskData.data[i+2] = maskColor.b; } maskCtx.globalCompositeOperation = 'source-over'; @@ -110,6 +110,7 @@ class MaskEditorDialog extends ComfyDialog { createButton(name, callback) { var button = document.createElement("button"); + button.style.pointerEvents = "auto"; button.innerText = name; button.addEventListener("click", callback); return button; @@ -146,6 +147,7 @@ class MaskEditorDialog extends ComfyDialog { divElement.style.display = "flex"; divElement.style.position = "relative"; divElement.style.top = "2px"; + divElement.style.pointerEvents = "auto"; self.brush_slider_input = document.createElement('input'); self.brush_slider_input.setAttribute('type', 'range'); self.brush_slider_input.setAttribute('min', '1'); @@ -162,6 +164,41 @@ class MaskEditorDialog extends ComfyDialog { return divElement; } + createOpacitySlider(self, name, callback) { + const divElement = document.createElement('div'); + divElement.id = "maskeditor-opacity-slider"; + divElement.style.cssFloat = "left"; + divElement.style.fontFamily = "sans-serif"; + divElement.style.marginRight = "4px"; + divElement.style.color = "var(--input-text)"; + divElement.style.backgroundColor = "var(--comfy-input-bg)"; + divElement.style.borderRadius = "8px"; + divElement.style.borderColor = "var(--border-color)"; + divElement.style.borderStyle = "solid"; + divElement.style.fontSize = "15px"; + divElement.style.height = "21px"; + divElement.style.padding = "1px 6px"; + divElement.style.display = "flex"; + divElement.style.position = "relative"; + divElement.style.top = "2px"; + divElement.style.pointerEvents = "auto"; + self.opacity_slider_input = document.createElement('input'); + self.opacity_slider_input.setAttribute('type', 'range'); + self.opacity_slider_input.setAttribute('min', '0.1'); + self.opacity_slider_input.setAttribute('max', '1.0'); + self.opacity_slider_input.setAttribute('step', '0.01') + self.opacity_slider_input.setAttribute('value', '0.7'); + const labelElement = document.createElement("label"); + labelElement.textContent = name; + + divElement.appendChild(labelElement); + divElement.appendChild(self.opacity_slider_input); + + self.opacity_slider_input.addEventListener("input", callback); + + return divElement; + } + setlayout(imgCanvas, maskCanvas) { const self = this; @@ -173,6 +210,7 @@ class MaskEditorDialog extends ComfyDialog { bottom_panel.style.left = "20px"; bottom_panel.style.right = "20px"; bottom_panel.style.height = "50px"; + bottom_panel.style.pointerEvents = "none"; var brush = document.createElement("div"); brush.id = "brush"; @@ -191,14 +229,36 @@ class MaskEditorDialog extends ComfyDialog { this.element.appendChild(bottom_panel); document.body.appendChild(brush); + var clearButton = this.createLeftButton("Clear", () => { + self.maskCtx.clearRect(0, 0, self.maskCanvas.width, self.maskCanvas.height); + }); + this.brush_size_slider = this.createLeftSlider(self, "Thickness", (event) => { self.brush_size = event.target.value; self.updateBrushPreview(self, null, null); }); - var clearButton = this.createLeftButton("Clear", - () => { - self.maskCtx.clearRect(0, 0, self.maskCanvas.width, self.maskCanvas.height); - }); + + this.brush_opacity_slider = this.createOpacitySlider(self, "Opacity", (event) => { + self.brush_opacity = event.target.value; + if (self.brush_color_mode !== "negative") { + self.maskCanvas.style.opacity = self.brush_opacity; + } + }); + + this.colorButton = this.createLeftButton(this.getColorButtonText(), () => { + if (self.brush_color_mode === "black") { + self.brush_color_mode = "white"; + } + else if (self.brush_color_mode === "white") { + self.brush_color_mode = "negative"; + } + else { + self.brush_color_mode = "black"; + } + + self.updateWhenBrushColorModeChanged(); + }); + var cancelButton = this.createRightButton("Cancel", () => { document.removeEventListener("mouseup", MaskEditorDialog.handleMouseUp); document.removeEventListener("keydown", MaskEditorDialog.handleKeyDown); @@ -219,6 +279,8 @@ class MaskEditorDialog extends ComfyDialog { bottom_panel.appendChild(this.saveButton); bottom_panel.appendChild(cancelButton); bottom_panel.appendChild(this.brush_size_slider); + bottom_panel.appendChild(this.brush_opacity_slider); + bottom_panel.appendChild(this.colorButton); imgCanvas.style.position = "absolute"; maskCanvas.style.position = "absolute"; @@ -228,6 +290,10 @@ class MaskEditorDialog extends ComfyDialog { maskCanvas.style.top = imgCanvas.style.top; maskCanvas.style.left = imgCanvas.style.left; + + const maskCanvasStyle = this.getMaskCanvasStyle(); + maskCanvas.style.mixBlendMode = maskCanvasStyle.mixBlendMode; + maskCanvas.style.opacity = maskCanvasStyle.opacity; } async show() { @@ -313,7 +379,7 @@ class MaskEditorDialog extends ComfyDialog { let maskCtx = this.maskCanvas.getContext('2d', {willReadFrequently: true }); imgCtx.drawImage(orig_image, 0, 0, orig_image.width, orig_image.height); - prepare_mask(mask_image, this.maskCanvas, maskCtx); + prepare_mask(mask_image, this.maskCanvas, maskCtx, this.getMaskColor()); } async setImages(imgCanvas) { @@ -439,7 +505,85 @@ class MaskEditorDialog extends ComfyDialog { } } + getMaskCanvasStyle() { + if (this.brush_color_mode === "negative") { + return { + mixBlendMode: "difference", + opacity: "1", + }; + } + else { + return { + mixBlendMode: "initial", + opacity: this.brush_opacity, + }; + } + } + + getMaskColor() { + if (this.brush_color_mode === "black") { + return { r: 0, g: 0, b: 0 }; + } + if (this.brush_color_mode === "white") { + return { r: 255, g: 255, b: 255 }; + } + if (this.brush_color_mode === "negative") { + // negative effect only works with white color + return { r: 255, g: 255, b: 255 }; + } + + return { r: 0, g: 0, b: 0 }; + } + + getMaskFillStyle() { + const maskColor = this.getMaskColor(); + + return "rgb(" + maskColor.r + "," + maskColor.g + "," + maskColor.b + ")"; + } + + getColorButtonText() { + let colorCaption = "unknown"; + + if (this.brush_color_mode === "black") { + colorCaption = "black"; + } + else if (this.brush_color_mode === "white") { + colorCaption = "white"; + } + else if (this.brush_color_mode === "negative") { + colorCaption = "negative"; + } + + return "Color: " + colorCaption; + } + + updateWhenBrushColorModeChanged() { + this.colorButton.innerText = this.getColorButtonText(); + + // update mask canvas css styles + + const maskCanvasStyle = this.getMaskCanvasStyle(); + this.maskCanvas.style.mixBlendMode = maskCanvasStyle.mixBlendMode; + this.maskCanvas.style.opacity = maskCanvasStyle.opacity; + + // update mask canvas rgb colors + + const maskColor = this.getMaskColor(); + + const maskData = this.maskCtx.getImageData(0, 0, this.maskCanvas.width, this.maskCanvas.height); + + for (let i = 0; i < maskData.data.length; i += 4) { + maskData.data[i] = maskColor.r; + maskData.data[i+1] = maskColor.g; + maskData.data[i+2] = maskColor.b; + } + + this.maskCtx.putImageData(maskData, 0, 0); + } + + brush_opacity = 0.7; brush_size = 10; + brush_color_mode = "black"; drawing_mode = false; lastx = -1; lasty = -1; @@ -518,6 +662,19 @@ class MaskEditorDialog extends ComfyDialog { event.preventDefault(); self.pan_move(self, event); } + + let left_button_down = window.TouchEvent && event instanceof TouchEvent || event.buttons == 1; + + if(event.shiftKey && left_button_down) { + self.drawing_mode = false; + + const y = event.clientY; + let delta = (self.zoom_lasty - y)*0.005; + self.zoom_ratio = Math.max(Math.min(10.0, self.last_zoom_ratio - delta), 0.2); + + this.invalidatePanZoom(); + return; + } } pan_move(self, event) { @@ -535,7 +692,7 @@ class MaskEditorDialog extends ComfyDialog { } draw_move(self, event) { - if(event.ctrlKey) { + if(event.ctrlKey || event.shiftKey) { return; } @@ -546,7 +703,10 @@ class MaskEditorDialog extends ComfyDialog { self.updateBrushPreview(self); - if (window.TouchEvent && event instanceof TouchEvent || event.buttons == 1) { + let left_button_down = window.TouchEvent && event instanceof TouchEvent || event.buttons == 1; + let right_button_down = [2, 5, 32].includes(event.buttons); + + if (!event.altKey && left_button_down) { var diff = performance.now() - self.lasttime; const maskRect = self.maskCanvas.getBoundingClientRect(); @@ -581,7 +741,7 @@ class MaskEditorDialog extends ComfyDialog { if(diff > 20 && !this.drawing_mode) requestAnimationFrame(() => { self.maskCtx.beginPath(); - self.maskCtx.fillStyle = "rgb(0,0,0)"; + self.maskCtx.fillStyle = this.getMaskFillStyle(); self.maskCtx.globalCompositeOperation = "source-over"; self.maskCtx.arc(x, y, brush_size, 0, Math.PI * 2, false); self.maskCtx.fill(); @@ -591,7 +751,7 @@ class MaskEditorDialog extends ComfyDialog { else requestAnimationFrame(() => { self.maskCtx.beginPath(); - self.maskCtx.fillStyle = "rgb(0,0,0)"; + self.maskCtx.fillStyle = this.getMaskFillStyle(); self.maskCtx.globalCompositeOperation = "source-over"; var dx = x - self.lastx; @@ -613,7 +773,7 @@ class MaskEditorDialog extends ComfyDialog { self.lasttime = performance.now(); } - else if(event.buttons == 2 || event.buttons == 5 || event.buttons == 32) { + else if((event.altKey && left_button_down) || right_button_down) { const maskRect = self.maskCanvas.getBoundingClientRect(); const x = (event.offsetX || event.targetTouches[0].clientX - maskRect.left) / self.zoom_ratio; const y = (event.offsetY || event.targetTouches[0].clientY - maskRect.top) / self.zoom_ratio; @@ -687,13 +847,20 @@ class MaskEditorDialog extends ComfyDialog { self.drawing_mode = true; event.preventDefault(); + + if(event.shiftKey) { + self.zoom_lasty = event.clientY; + self.last_zoom_ratio = self.zoom_ratio; + return; + } + const maskRect = self.maskCanvas.getBoundingClientRect(); const x = (event.offsetX || event.targetTouches[0].clientX - maskRect.left) / self.zoom_ratio; const y = (event.offsetY || event.targetTouches[0].clientY - maskRect.top) / self.zoom_ratio; self.maskCtx.beginPath(); - if (event.button == 0) { - self.maskCtx.fillStyle = "rgb(0,0,0)"; + if (!event.altKey && event.button == 0) { + self.maskCtx.fillStyle = this.getMaskFillStyle(); self.maskCtx.globalCompositeOperation = "source-over"; } else { self.maskCtx.globalCompositeOperation = "destination-out"; diff --git a/web/extensions/core/simpleTouchSupport.js b/web/extensions/core/simpleTouchSupport.js new file mode 100644 index 000000000..041fc2c4c --- /dev/null +++ b/web/extensions/core/simpleTouchSupport.js @@ -0,0 +1,102 @@ +import { app } from "../../scripts/app.js"; + +let touchZooming; +let touchCount = 0; + +app.registerExtension({ + name: "Comfy.SimpleTouchSupport", + setup() { + let zoomPos; + let touchTime; + let lastTouch; + + function getMultiTouchPos(e) { + return Math.hypot(e.touches[0].clientX - e.touches[1].clientX, e.touches[0].clientY - e.touches[1].clientY); + } + + app.canvasEl.addEventListener( + "touchstart", + (e) => { + touchCount++; + lastTouch = null; + if (e.touches?.length === 1) { + // Store start time for press+hold for context menu + touchTime = new Date(); + lastTouch = e.touches[0]; + } else { + touchTime = null; + if (e.touches?.length === 2) { + // Store center pos for zoom + zoomPos = getMultiTouchPos(e); + app.canvas.pointer_is_down = false; + } + } + }, + true + ); + + app.canvasEl.addEventListener("touchend", (e) => { + touchZooming = false; + touchCount = e.touches?.length ?? touchCount - 1; + if (touchTime && !e.touches?.length) { + if (new Date() - touchTime > 600) { + try { + // hack to get litegraph to use this event + e.constructor = CustomEvent; + } catch (error) {} + e.clientX = lastTouch.clientX; + e.clientY = lastTouch.clientY; + + app.canvas.pointer_is_down = true; + app.canvas._mousedown_callback(e); + } + touchTime = null; + } + }); + + app.canvasEl.addEventListener( + "touchmove", + (e) => { + touchTime = null; + if (e.touches?.length === 2) { + app.canvas.pointer_is_down = false; + touchZooming = true; + LiteGraph.closeAllContextMenus(); + app.canvas.search_box?.close(); + const newZoomPos = getMultiTouchPos(e); + + const midX = (e.touches[0].clientX + e.touches[1].clientX) / 2; + const midY = (e.touches[0].clientY + e.touches[1].clientY) / 2; + + let scale = app.canvas.ds.scale; + const diff = zoomPos - newZoomPos; + if (diff > 0.5) { + scale *= 1 / 1.07; + } else if (diff < -0.5) { + scale *= 1.07; + } + app.canvas.ds.changeScale(scale, [midX, midY]); + app.canvas.setDirty(true, true); + zoomPos = newZoomPos; + } + }, + true + ); + }, +}); + +const processMouseDown = LGraphCanvas.prototype.processMouseDown; +LGraphCanvas.prototype.processMouseDown = function (e) { + if (touchZooming || touchCount) { + return; + } + return processMouseDown.apply(this, arguments); +}; + +const processMouseMove = LGraphCanvas.prototype.processMouseMove; +LGraphCanvas.prototype.processMouseMove = function (e) { + if (touchZooming || touchCount > 1) { + return; + } + return processMouseMove.apply(this, arguments); +}; diff --git a/web/extensions/core/snapToGrid.js b/web/extensions/core/snapToGrid.js index dc534d6ed..aac017748 100644 --- a/web/extensions/core/snapToGrid.js +++ b/web/extensions/core/snapToGrid.js @@ -2,6 +2,13 @@ import { app } from "../../scripts/app.js"; // Shift + drag/resize to snap to grid +/** Rounds a Vector2 in-place to the current CANVAS_GRID_SIZE. */ +function roundVectorToGrid(vec) { + vec[0] = LiteGraph.CANVAS_GRID_SIZE * Math.round(vec[0] / LiteGraph.CANVAS_GRID_SIZE); + vec[1] = LiteGraph.CANVAS_GRID_SIZE * Math.round(vec[1] / LiteGraph.CANVAS_GRID_SIZE); + return vec; +} + app.registerExtension({ name: "Comfy.SnapToGrid", init() { @@ -43,10 +50,7 @@ app.registerExtension({ const onResize = node.onResize; node.onResize = function () { if (app.shiftDown) { - const w = LiteGraph.CANVAS_GRID_SIZE * Math.round(node.size[0] / LiteGraph.CANVAS_GRID_SIZE); - const h = LiteGraph.CANVAS_GRID_SIZE * Math.round(node.size[1] / LiteGraph.CANVAS_GRID_SIZE); - node.size[0] = w; - node.size[1] = h; + roundVectorToGrid(node.size); } return onResize?.apply(this, arguments); }; @@ -57,9 +61,7 @@ app.registerExtension({ const origDrawNode = LGraphCanvas.prototype.drawNode; LGraphCanvas.prototype.drawNode = function (node, ctx) { if (app.shiftDown && this.node_dragged && node.id in this.selected_nodes) { - const x = LiteGraph.CANVAS_GRID_SIZE * Math.round(node.pos[0] / LiteGraph.CANVAS_GRID_SIZE); - const y = LiteGraph.CANVAS_GRID_SIZE * Math.round(node.pos[1] / LiteGraph.CANVAS_GRID_SIZE); - + const [x, y] = roundVectorToGrid([...node.pos]); const shiftX = x - node.pos[0]; let shiftY = y - node.pos[1]; @@ -85,5 +87,85 @@ app.registerExtension({ return origDrawNode.apply(this, arguments); }; + + + + /** + * The currently moving, selected group only. Set after the `selected_group` has actually started + * moving. + */ + let selectedAndMovingGroup = null; + + /** + * Handles moving a group; tracking when a group has been moved (to show the ghost in `drawGroups` + * below) as well as handle the last move call from LiteGraph's `processMouseUp`. + */ + const groupMove = LGraphGroup.prototype.move; + LGraphGroup.prototype.move = function(deltax, deltay, ignore_nodes) { + const v = groupMove.apply(this, arguments); + // When we've started moving, set `selectedAndMovingGroup` as LiteGraph sets `selected_group` + // too eagerly and we don't want to behave like we're moving until we get a delta. + if (!selectedAndMovingGroup && app.canvas.selected_group === this && (deltax || deltay)) { + selectedAndMovingGroup = this; + } + + // LiteGraph will call group.move both on mouse-move as well as mouse-up though we only want + // to snap on a mouse-up which we can determine by checking if `app.canvas.last_mouse_dragging` + // has been set to `false`. Essentially, this check here is the equivilant to calling an + // `LGraphGroup.prototype.onNodeMoved` if it had existed. + if (app.canvas.last_mouse_dragging === false && app.shiftDown) { + // After moving a group (while app.shiftDown), snap all the child nodes and, finally, + // align the group itself. + this.recomputeInsideNodes(); + for (const node of this._nodes) { + node.alignToGrid(); + } + LGraphNode.prototype.alignToGrid.apply(this); + } + return v; + }; + + /** + * Handles drawing a group when, snapping the size when one is actively being resized tracking and/or + * drawing a ghost box when one is actively being moved. This mimics the node snapping behavior for + * both. + */ + const drawGroups = LGraphCanvas.prototype.drawGroups; + LGraphCanvas.prototype.drawGroups = function (canvas, ctx) { + if (this.selected_group && app.shiftDown) { + if (this.selected_group_resizing) { + roundVectorToGrid(this.selected_group.size); + } else if (selectedAndMovingGroup) { + const [x, y] = roundVectorToGrid([...selectedAndMovingGroup.pos]); + const f = ctx.fillStyle; + const s = ctx.strokeStyle; + ctx.fillStyle = "rgba(100, 100, 100, 0.33)"; + ctx.strokeStyle = "rgba(100, 100, 100, 0.66)"; + ctx.rect(x, y, ...selectedAndMovingGroup.size); + ctx.fill(); + ctx.stroke(); + ctx.fillStyle = f; + ctx.strokeStyle = s; + } + } else if (!this.selected_group) { + selectedAndMovingGroup = null; + } + return drawGroups.apply(this, arguments); + }; + + + /** Handles adding a group in a snapping-enabled state. */ + const onGroupAdd = LGraphCanvas.onGroupAdd; + LGraphCanvas.onGroupAdd = function() { + const v = onGroupAdd.apply(app.canvas, arguments); + if (app.shiftDown) { + const lastGroup = app.graph._groups[app.graph._groups.length - 1]; + if (lastGroup) { + roundVectorToGrid(lastGroup.pos); + roundVectorToGrid(lastGroup.size); + } + } + return v; + }; }, }); diff --git a/web/extensions/core/undoRedo.js b/web/extensions/core/undoRedo.js index ff976c74c..900eed2a7 100644 --- a/web/extensions/core/undoRedo.js +++ b/web/extensions/core/undoRedo.js @@ -1,4 +1,5 @@ import { app } from "../../scripts/app.js"; +import { api } from "../../scripts/api.js" const MAX_HISTORY = 50; @@ -15,6 +16,7 @@ function checkState() { } activeState = clone(currentState); redo.length = 0; + api.dispatchEvent(new CustomEvent("graphChanged", { detail: activeState })); } } @@ -92,7 +94,7 @@ const undoRedo = async (e) => { }; const bindInput = (activeEl) => { - if (activeEl?.tagName !== "CANVAS" && activeEl?.tagName !== "BODY") { + if (activeEl && activeEl.tagName !== "CANVAS" && activeEl.tagName !== "BODY") { for (const evt of ["change", "input", "blur"]) { if (`on${evt}` in activeEl) { const listener = () => { @@ -111,12 +113,16 @@ window.addEventListener( "keydown", (e) => { requestAnimationFrame(async () => { - const activeEl = document.activeElement; - if (activeEl?.tagName === "INPUT" || activeEl?.type === "textarea") { - // Ignore events on inputs, they have their native history - return; + let activeEl; + // If we are auto queue in change mode then we do want to trigger on inputs + if (!app.ui.autoQueueEnabled || app.ui.autoQueueMode === "instant") { + activeEl = document.activeElement; + if (activeEl?.tagName === "INPUT" || activeEl?.type === "textarea") { + // Ignore events on inputs, they have their native history + return; + } } - + keyIgnored = e.key === "Control" || e.key === "Shift" || e.key === "Alt" || e.key === "Meta"; if (keyIgnored) return; @@ -143,6 +149,11 @@ window.addEventListener("mouseup", () => { checkState(); }); +// Handle prompt queue event for dynamic widget changes +api.addEventListener("promptQueued", () => { + checkState(); +}); + // Handle litegraph clicks const processMouseUp = LGraphCanvas.prototype.processMouseUp; LGraphCanvas.prototype.processMouseUp = function (e) { @@ -156,3 +167,11 @@ LGraphCanvas.prototype.processMouseDown = function (e) { checkState(); return v; }; + +// Handle litegraph context menu for COMBO widgets +const close = LiteGraph.ContextMenu.prototype.close; +LiteGraph.ContextMenu.prototype.close = function(e) { + const v = close.apply(this, arguments); + checkState(); + return v; +} \ No newline at end of file diff --git a/web/extensions/core/webcamCapture.js b/web/extensions/core/webcamCapture.js new file mode 100644 index 000000000..dd5725bd4 --- /dev/null +++ b/web/extensions/core/webcamCapture.js @@ -0,0 +1,126 @@ +import { app } from "../../scripts/app.js"; +import { api } from "../../scripts/api.js"; + +const WEBCAM_READY = Symbol(); + +app.registerExtension({ + name: "Comfy.WebcamCapture", + getCustomWidgets(app) { + return { + WEBCAM(node, inputName) { + let res; + node[WEBCAM_READY] = new Promise((resolve) => (res = resolve)); + + const container = document.createElement("div"); + container.style.background = "rgba(0,0,0,0.25)"; + container.style.textAlign = "center"; + + const video = document.createElement("video"); + video.style.height = video.style.width = "100%"; + + const loadVideo = async () => { + try { + const stream = await navigator.mediaDevices.getUserMedia({ video: true, audio: false }); + container.replaceChildren(video); + + setTimeout(() => res(video), 500); // Fallback as loadedmetadata doesnt fire sometimes? + video.addEventListener("loadedmetadata", () => res(video), false); + video.srcObject = stream; + video.play(); + } catch (error) { + const label = document.createElement("div"); + label.style.color = "red"; + label.style.overflow = "auto"; + label.style.maxHeight = "100%"; + label.style.whiteSpace = "pre-wrap"; + + if (window.isSecureContext) { + label.textContent = "Unable to load webcam, please ensure access is granted:\n" + error.message; + } else { + label.textContent = "Unable to load webcam. A secure context is required, if you are not accessing ComfyUI on localhost (127.0.0.1) you will have to enable TLS (https)\n\n" + error.message; + } + + container.replaceChildren(label); + } + }; + + loadVideo(); + + return { widget: node.addDOMWidget(inputName, "WEBCAM", container) }; + }, + }; + }, + nodeCreated(node) { + if ((node.type, node.constructor.comfyClass !== "WebcamCapture")) return; + + let video; + const camera = node.widgets.find((w) => w.name === "image"); + const w = node.widgets.find((w) => w.name === "width"); + const h = node.widgets.find((w) => w.name === "height"); + const captureOnQueue = node.widgets.find((w) => w.name === "capture_on_queue"); + + const canvas = document.createElement("canvas"); + + const capture = () => { + canvas.width = w.value; + canvas.height = h.value; + const ctx = canvas.getContext("2d"); + ctx.drawImage(video, 0, 0, w.value, h.value); + const data = canvas.toDataURL("image/png"); + + const img = new Image(); + img.onload = () => { + node.imgs = [img]; + app.graph.setDirtyCanvas(true); + requestAnimationFrame(() => { + node.setSizeForImage?.(); + }); + }; + img.src = data; + }; + + const btn = node.addWidget("button", "waiting for camera...", "capture", capture); + btn.disabled = true; + btn.serializeValue = () => undefined; + + camera.serializeValue = async () => { + if (captureOnQueue.value) { + capture(); + } else if (!node.imgs?.length) { + const err = `No webcam image captured`; + alert(err); + throw new Error(err); + } + + // Upload image to temp storage + const blob = await new Promise((r) => canvas.toBlob(r)); + const name = `${+new Date()}.png`; + const file = new File([blob], name); + const body = new FormData(); + body.append("image", file); + body.append("subfolder", "webcam"); + body.append("type", "temp"); + const resp = await api.fetchApi("/upload/image", { + method: "POST", + body, + }); + if (resp.status !== 200) { + const err = `Error uploading camera image: ${resp.status} - ${resp.statusText}`; + alert(err); + throw new Error(err); + } + return `webcam/${name} [temp]`; + }; + + node[WEBCAM_READY].then((v) => { + video = v; + // If width isnt specified then use video output resolution + if (!w.value) { + w.value = video.videoWidth || 640; + h.value = video.videoHeight || 480; + } + btn.disabled = false; + btn.label = "capture"; + }); + }, +}); diff --git a/web/extensions/core/widgetInputs.js b/web/extensions/core/widgetInputs.js index 3f1c1f8c1..f1a1d22cd 100644 --- a/web/extensions/core/widgetInputs.js +++ b/web/extensions/core/widgetInputs.js @@ -22,6 +22,7 @@ function isConvertableWidget(widget, config) { } function hideWidget(node, widget, suffix = "") { + if (widget.type?.startsWith(CONVERTED_TYPE)) return; widget.origType = widget.type; widget.origComputeSize = widget.computeSize; widget.origSerializeValue = widget.serializeValue; @@ -255,11 +256,27 @@ export function mergeIfValid(output, config2, forceUpdate, recreateWidget, confi return { customConfig }; } +let useConversionSubmenusSetting; app.registerExtension({ name: "Comfy.WidgetInputs", + init() { + useConversionSubmenusSetting = app.ui.settings.addSetting({ + id: "Comfy.NodeInputConversionSubmenus", + name: "Node widget/input conversion sub-menus", + tooltip: "In the node context menu, place the entries that convert between input/widget in sub-menus.", + type: "boolean", + defaultValue: true, + }); + }, async beforeRegisterNodeDef(nodeType, nodeData, app) { // Add menu options to conver to/from widgets const origGetExtraMenuOptions = nodeType.prototype.getExtraMenuOptions; + nodeType.prototype.convertWidgetToInput = function (widget) { + const config = getConfig.call(this, widget.name) ?? [widget.type, widget.options || {}]; + if (!isConvertableWidget(widget, config)) return false; + convertToInput(this, widget, config); + return true; + }; nodeType.prototype.getExtraMenuOptions = function (_, options) { const r = origGetExtraMenuOptions ? origGetExtraMenuOptions.apply(this, arguments) : undefined; @@ -285,12 +302,31 @@ app.registerExtension({ } } } + + //Convert.. main menu if (toInput.length) { - options.push(...toInput, null); + if (useConversionSubmenusSetting.value) { + options.push({ + content: "Convert Widget to Input", + submenu: { + options: toInput, + }, + }); + } else { + options.push(...toInput, null); + } } - if (toWidget.length) { - options.push(...toWidget, null); + if (useConversionSubmenusSetting.value) { + options.push({ + content: "Convert Input to Widget", + submenu: { + options: toWidget, + }, + }); + } else { + options.push(...toWidget, null); + } } } diff --git a/web/lib/litegraph.core.js b/web/lib/litegraph.core.js index 434c4a83b..427a62b59 100644 --- a/web/lib/litegraph.core.js +++ b/web/lib/litegraph.core.js @@ -7247,7 +7247,7 @@ LGraphNode.prototype.executeAction = function(action) //create links for (var i = 0; i < clipboard_info.links.length; ++i) { var link_info = clipboard_info.links[i]; - var origin_node; + var origin_node = undefined; var origin_node_relative_id = link_info[0]; if (origin_node_relative_id != null) { origin_node = nodes[origin_node_relative_id]; @@ -11496,7 +11496,7 @@ LGraphNode.prototype.executeAction = function(action) } timeout_close = setTimeout(function() { dialog.close(); - }, 500); + }, typeof options.hide_on_mouse_leave === "number" ? options.hide_on_mouse_leave : 500); }); // if filtering, check focus changed to comboboxes and prevent closing if (options.do_type_filter){ @@ -11549,7 +11549,7 @@ LGraphNode.prototype.executeAction = function(action) dialog.close(); } else if (e.keyCode == 13) { if (selected) { - select(selected.innerHTML); + select(unescape(selected.dataset["type"])); } else if (first) { select(first); } else { @@ -11910,7 +11910,7 @@ LGraphNode.prototype.executeAction = function(action) var ctor = LiteGraph.registered_node_types[ type ]; if(filter && ctor.filter != filter ) return false; - if ((!options.show_all_if_empty || str) && type.toLowerCase().indexOf(str) === -1) + if ((!options.show_all_if_empty || str) && type.toLowerCase().indexOf(str) === -1 && (!ctor.title || ctor.title.toLowerCase().indexOf(str) === -1)) return false; // filter by slot IN, OUT types @@ -11964,7 +11964,18 @@ LGraphNode.prototype.executeAction = function(action) if (!first) { first = type; } - help.innerText = type; + + const nodeType = LiteGraph.registered_node_types[type]; + if (nodeType?.title) { + help.innerText = nodeType?.title; + const typeEl = document.createElement("span"); + typeEl.className = "litegraph lite-search-item-type"; + typeEl.textContent = type; + help.append(typeEl); + } else { + help.innerText = type; + } + help.dataset["type"] = escape(type); help.className = "litegraph lite-search-item"; if (className) { diff --git a/web/lib/litegraph.css b/web/lib/litegraph.css index 918858f41..5524e24ba 100644 --- a/web/lib/litegraph.css +++ b/web/lib/litegraph.css @@ -184,6 +184,7 @@ color: white; padding-left: 10px; margin-right: 5px; + max-width: 300px; } .litegraph.litesearchbox .name { @@ -227,6 +228,18 @@ color: black; } +.litegraph.lite-search-item-type { + display: inline-block; + background: rgba(0,0,0,0.2); + margin-left: 5px; + font-size: 14px; + padding: 2px 5px; + position: relative; + top: -2px; + opacity: 0.8; + border-radius: 4px; + } + /* DIALOGs ******/ .litegraph .dialog { diff --git a/web/scripts/api.js b/web/scripts/api.js index 3a9bcc87a..8c8155be6 100644 --- a/web/scripts/api.js +++ b/web/scripts/api.js @@ -5,6 +5,7 @@ class ComfyApi extends EventTarget { super(); this.api_host = location.host; this.api_base = location.pathname.split('/').slice(0, -1).join('/'); + this.initialClientId = sessionStorage.getItem("clientId"); } apiURL(route) { @@ -118,7 +119,8 @@ class ComfyApi extends EventTarget { case "status": if (msg.data.sid) { this.clientId = msg.data.sid; - window.name = this.clientId; + window.name = this.clientId; // use window name so it isnt reused when duplicating tabs + sessionStorage.setItem("clientId", this.clientId); // store in session storage so duplicate tab can load correct workflow } this.dispatchEvent(new CustomEvent("status", { detail: msg.data.status })); break; diff --git a/web/scripts/app.js b/web/scripts/app.js index d131045d7..f96d197a8 100644 --- a/web/scripts/app.js +++ b/web/scripts/app.js @@ -262,6 +262,36 @@ export class ComfyApp { }) ); } + + #addRestoreWorkflowView() { + const serialize = LGraph.prototype.serialize; + const self = this; + LGraph.prototype.serialize = function() { + const workflow = serialize.apply(this, arguments); + + // Store the drag & scale info in the serialized workflow if the setting is enabled + if (self.enableWorkflowViewRestore.value) { + if (!workflow.extra) { + workflow.extra = {}; + } + workflow.extra.ds = { + scale: self.canvas.ds.scale, + offset: self.canvas.ds.offset, + }; + } else if (workflow.extra?.ds) { + // Clear any old view data + delete workflow.extra.ds; + } + + return workflow; + } + this.enableWorkflowViewRestore = this.ui.settings.addSetting({ + id: "Comfy.EnableWorkflowViewRestore", + name: "Save and restore canvas position and zoom level in workflows", + type: "boolean", + defaultValue: true + }); + } /** * Adds special context menu handling for nodes @@ -269,6 +299,71 @@ export class ComfyApp { * @param {*} node The node to add the menu handler */ #addNodeContextMenuHandler(node) { + function getCopyImageOption(img) { + if (typeof window.ClipboardItem === "undefined") return []; + return [ + { + content: "Copy Image", + callback: async () => { + const url = new URL(img.src); + url.searchParams.delete("preview"); + + const writeImage = async (blob) => { + await navigator.clipboard.write([ + new ClipboardItem({ + [blob.type]: blob, + }), + ]); + }; + + try { + const data = await fetch(url); + const blob = await data.blob(); + try { + await writeImage(blob); + } catch (error) { + // Chrome seems to only support PNG on write, convert and try again + if (blob.type !== "image/png") { + const canvas = $el("canvas", { + width: img.naturalWidth, + height: img.naturalHeight, + }); + const ctx = canvas.getContext("2d"); + let image; + if (typeof window.createImageBitmap === "undefined") { + image = new Image(); + const p = new Promise((resolve, reject) => { + image.onload = resolve; + image.onerror = reject; + }).finally(() => { + URL.revokeObjectURL(image.src); + }); + image.src = URL.createObjectURL(blob); + await p; + } else { + image = await createImageBitmap(blob); + } + try { + ctx.drawImage(image, 0, 0); + canvas.toBlob(writeImage, "image/png"); + } finally { + if (typeof image.close === "function") { + image.close(); + } + } + + return; + } + throw error; + } + } catch (error) { + alert("Error copying image: " + (error.message ?? error)); + } + }, + }, + ]; + } + node.prototype.getExtraMenuOptions = function (_, options) { if (this.imgs) { // If this node has images then we add an open in new tab item @@ -286,16 +381,17 @@ export class ComfyApp { content: "Open Image", callback: () => { let url = new URL(img.src); - url.searchParams.delete('preview'); - window.open(url, "_blank") + url.searchParams.delete("preview"); + window.open(url, "_blank"); }, }, + ...getCopyImageOption(img), { content: "Save Image", callback: () => { const a = document.createElement("a"); let url = new URL(img.src); - url.searchParams.delete('preview'); + url.searchParams.delete("preview"); a.href = url; a.setAttribute("download", new URLSearchParams(url.search).get("filename")); document.body.append(a); @@ -308,33 +404,41 @@ export class ComfyApp { } options.push({ - content: "Bypass", - callback: (obj) => { if (this.mode === 4) this.mode = 0; else this.mode = 4; this.graph.change(); } - }); + content: "Bypass", + callback: (obj) => { + if (this.mode === 4) this.mode = 0; + else this.mode = 4; + this.graph.change(); + }, + }); // prevent conflict of clipspace content - if(!ComfyApp.clipspace_return_node) { + if (!ComfyApp.clipspace_return_node) { options.push({ - content: "Copy (Clipspace)", - callback: (obj) => { ComfyApp.copyToClipspace(this); } - }); + content: "Copy (Clipspace)", + callback: (obj) => { + ComfyApp.copyToClipspace(this); + }, + }); - if(ComfyApp.clipspace != null) { + if (ComfyApp.clipspace != null) { options.push({ - content: "Paste (Clipspace)", - callback: () => { ComfyApp.pasteFromClipspace(this); } - }); + content: "Paste (Clipspace)", + callback: () => { + ComfyApp.pasteFromClipspace(this); + }, + }); } - if(ComfyApp.isImageNode(this)) { + if (ComfyApp.isImageNode(this)) { options.push({ - content: "Open in MaskEditor", - callback: (obj) => { - ComfyApp.copyToClipspace(this); - ComfyApp.clipspace_return_node = this; - ComfyApp.open_maskeditor(); - } - }); + content: "Open in MaskEditor", + callback: (obj) => { + ComfyApp.copyToClipspace(this); + ComfyApp.clipspace_return_node = this; + ComfyApp.open_maskeditor(); + }, + }); } } }; @@ -879,6 +983,12 @@ export class ComfyApp { const origProcessMouseDown = LGraphCanvas.prototype.processMouseDown; LGraphCanvas.prototype.processMouseDown = function(e) { + // prepare for ctrl+shift drag: zoom start + if(e.ctrlKey && e.shiftKey && e.buttons) { + self.zoom_drag_start = [e.x, e.y, this.ds.scale]; + return; + } + const res = origProcessMouseDown.apply(this, arguments); this.selected_group_moving = false; @@ -899,6 +1009,26 @@ export class ComfyApp { const origProcessMouseMove = LGraphCanvas.prototype.processMouseMove; LGraphCanvas.prototype.processMouseMove = function(e) { + // handle ctrl+shift drag + if(e.ctrlKey && e.shiftKey && self.zoom_drag_start) { + // stop canvas zoom action + if(!e.buttons) { + self.zoom_drag_start = null; + return; + } + + // calculate delta + let deltaY = e.y - self.zoom_drag_start[1]; + let startScale = self.zoom_drag_start[2]; + + let scale = startScale - deltaY/100; + + this.ds.changeScale(scale, [this.ds.element.width/2, this.ds.element.height/2]); + this.graph.change(); + + return; + } + const orig_selected_group = this.selected_group; if (this.selected_group && !this.selected_group_resizing && !this.selected_group_moving) { @@ -985,6 +1115,20 @@ export class ComfyApp { // Trigger onPaste return true; } + + if((e.key === '+') && e.altKey) { + block_default = true; + let scale = this.ds.scale * 1.1; + this.ds.changeScale(scale, [this.ds.element.width/2, this.ds.element.height/2]); + this.graph.change(); + } + + if((e.key === '-') && e.altKey) { + block_default = true; + let scale = this.ds.scale * 1 / 1.1; + this.ds.changeScale(scale, [this.ds.element.width/2, this.ds.element.height/2]); + this.graph.change(); + } } this.graph.change(); @@ -1391,6 +1535,7 @@ export class ComfyApp { this.#addProcessKeyHandler(); this.#addConfigureHandler(); this.#addApiUpdateHandlers(); + this.#addRestoreWorkflowView(); this.graph = new LGraph(); @@ -1425,12 +1570,17 @@ export class ComfyApp { // Load previous workflow let restored = false; try { - const json = localStorage.getItem("workflow"); - if (json) { - const workflow = JSON.parse(json); - await this.loadGraphData(workflow); - restored = true; - } + const loadWorkflow = async (json) => { + if (json) { + const workflow = JSON.parse(json); + await this.loadGraphData(workflow); + return true; + } + }; + const clientId = api.initialClientId ?? api.clientId; + restored = + (clientId && (await loadWorkflow(sessionStorage.getItem(`workflow:${clientId}`)))) || + (await loadWorkflow(localStorage.getItem("workflow"))); } catch (err) { console.error("Error loading previous workflow", err); } @@ -1441,7 +1591,13 @@ export class ComfyApp { } // Save current workflow automatically - setInterval(() => localStorage.setItem("workflow", JSON.stringify(this.graph.serialize())), 1000); + setInterval(() => { + const workflow = JSON.stringify(this.graph.serialize()); + localStorage.setItem("workflow", workflow); + if (api.clientId) { + sessionStorage.setItem(`workflow:${api.clientId}`, workflow); + } + }, 1000); this.#addDrawNodeHandler(); this.#addDrawGroupsHandler(); @@ -1644,7 +1800,7 @@ export class ComfyApp { * @param {*} graphData A serialized graph object * @param { boolean } clean If the graph state, e.g. images, should be cleared */ - async loadGraphData(graphData, clean = true) { + async loadGraphData(graphData, clean = true, restore_view = true) { if (clean !== false) { this.clean(); } @@ -1680,6 +1836,10 @@ export class ComfyApp { try { this.graph.configure(graphData); + if (restore_view && this.enableWorkflowViewRestore.value && graphData.extra?.ds) { + this.canvas.ds.offset = graphData.extra.ds.offset; + this.canvas.ds.scale = graphData.extra.ds.scale; + } } catch (error) { let errorHint = []; // Try extracting filename to see if it was caused by an extension script @@ -1994,6 +2154,15 @@ export class ComfyApp { } finally { this.#processingQueue = false; } + api.dispatchEvent(new CustomEvent("promptQueued", { detail: { number, batchCount } })); + } + + showErrorOnFileLoad(file) { + this.ui.dialog.show( + $el("div", [ + $el("p", {textContent: `Unable to find workflow in ${file.name}`}) + ]).outerHTML + ); } /** @@ -2003,25 +2172,27 @@ export class ComfyApp { async handleFile(file) { if (file.type === "image/png") { const pngInfo = await getPngMetadata(file); - if (pngInfo) { - if (pngInfo.workflow) { - await this.loadGraphData(JSON.parse(pngInfo.workflow)); - } else if (pngInfo.prompt) { - this.loadApiJson(JSON.parse(pngInfo.prompt)); - } else if (pngInfo.parameters) { - importA1111(this.graph, pngInfo.parameters); - } + if (pngInfo?.workflow) { + await this.loadGraphData(JSON.parse(pngInfo.workflow)); + } else if (pngInfo?.prompt) { + this.loadApiJson(JSON.parse(pngInfo.prompt)); + } else if (pngInfo?.parameters) { + importA1111(this.graph, pngInfo.parameters); + } else { + this.showErrorOnFileLoad(file); } } else if (file.type === "image/webp") { const pngInfo = await getWebpMetadata(file); - if (pngInfo) { - if (pngInfo.workflow) { - this.loadGraphData(JSON.parse(pngInfo.workflow)); - } else if (pngInfo.Workflow) { - this.loadGraphData(JSON.parse(pngInfo.Workflow)); // Support loading workflows from that webp custom node. - } else if (pngInfo.prompt) { - this.loadApiJson(JSON.parse(pngInfo.prompt)); - } + // Support loading workflows from that webp custom node. + const workflow = pngInfo?.workflow || pngInfo?.Workflow; + const prompt = pngInfo?.prompt || pngInfo?.Prompt; + + if (workflow) { + this.loadGraphData(JSON.parse(workflow)); + } else if (prompt) { + this.loadApiJson(JSON.parse(prompt)); + } else { + this.showErrorOnFileLoad(file); } } else if (file.type === "application/json" || file.name?.endsWith(".json")) { const reader = new FileReader(); @@ -2042,7 +2213,11 @@ export class ComfyApp { await this.loadGraphData(JSON.parse(info.workflow)); } else if (info.prompt) { this.loadApiJson(JSON.parse(info.prompt)); + } else { + this.showErrorOnFileLoad(file); } + } else { + this.showErrorOnFileLoad(file); } } @@ -2063,6 +2238,7 @@ export class ComfyApp { const data = apiData[id]; const node = LiteGraph.createNode(data.class_type); node.id = isNaN(+id) ? id : +id; + node.title = data._meta?.title ?? node.title graph.add(node); } @@ -2074,8 +2250,17 @@ export class ComfyApp { if (value instanceof Array) { const [fromId, fromSlot] = value; const fromNode = app.graph.getNodeById(fromId); - const toSlot = node.inputs?.findIndex((inp) => inp.name === input); - if (toSlot !== -1) { + let toSlot = node.inputs?.findIndex((inp) => inp.name === input); + if (toSlot == null || toSlot === -1) { + try { + // Target has no matching input, most likely a converted widget + const widget = node.widgets?.find((w) => w.name === input); + if (widget && node.convertWidgetToInput?.(widget)) { + toSlot = node.inputs?.length - 1; + } + } catch (error) {} + } + if (toSlot != null || toSlot !== -1) { fromNode.connect(fromSlot, node, toSlot); } } else { @@ -2137,6 +2322,14 @@ export class ComfyApp { } } } + + await this.#invokeExtensionsAsync("refreshComboInNodes", defs); + } + + resetView() { + app.canvas.ds.scale = 1; + app.canvas.ds.offset = [0, 0] + app.graph.setDirtyCanvas(true, true); } /** diff --git a/web/scripts/domWidget.js b/web/scripts/domWidget.js index d5eeebdbd..b7f437ad2 100644 --- a/web/scripts/domWidget.js +++ b/web/scripts/domWidget.js @@ -11,9 +11,10 @@ function intersect(a, b) { else return null; } -function getClipPath(node, element, elRect) { +function getClipPath(node, element) { const selectedNode = Object.values(app.canvas.selected_nodes)[0]; if (selectedNode && selectedNode !== node) { + const elRect = element.getBoundingClientRect(); const MARGIN = 7; const scale = app.canvas.ds.scale; @@ -269,7 +270,7 @@ LGraphNode.prototype.addDOMWidget = function (name, type, element, options) { }); if (enableDomClipping) { - element.style.clipPath = getClipPath(node, element, elRect); + element.style.clipPath = getClipPath(node, element); element.style.willChange = "clip-path"; } diff --git a/web/scripts/logging.js b/web/scripts/logging.js index c73462e1e..875dd970b 100644 --- a/web/scripts/logging.js +++ b/web/scripts/logging.js @@ -269,6 +269,9 @@ export class ComfyLogging { id: settingId, name: settingId, defaultValue: true, + onChange: (value) => { + this.enabled = value; + }, type: (name, setter, value) => { return $el("tr", [ $el("td", [ @@ -283,7 +286,7 @@ export class ComfyLogging { type: "checkbox", checked: value, onchange: (event) => { - setter((this.enabled = event.target.checked)); + setter(event.target.checked); }, }), $el("button", { diff --git a/web/scripts/pnginfo.js b/web/scripts/pnginfo.js index 83a4ebc86..7132fb60f 100644 --- a/web/scripts/pnginfo.js +++ b/web/scripts/pnginfo.js @@ -24,7 +24,7 @@ export function getPngMetadata(file) { const length = dataView.getUint32(offset); // Get the chunk type const type = String.fromCharCode(...pngData.slice(offset + 4, offset + 8)); - if (type === "tEXt" || type == "comf") { + if (type === "tEXt" || type == "comf" || type === "iTXt") { // Get the keyword let keyword_end = offset + 8; while (pngData[keyword_end] !== 0) { @@ -33,7 +33,7 @@ export function getPngMetadata(file) { const keyword = String.fromCharCode(...pngData.slice(offset + 8, keyword_end)); // Get the text const contentArraySegment = pngData.slice(keyword_end + 1, offset + 8 + length); - const contentJson = Array.from(contentArraySegment).map(s=>String.fromCharCode(s)).join('') + const contentJson = new TextDecoder("utf-8").decode(contentArraySegment); txt_chunks[keyword] = contentJson; } @@ -170,9 +170,12 @@ export async function importA1111(graph, parameters) { const opts = parameters .substr(p) .split("\n")[1] - .split(",") + .match(new RegExp("\\s*([^:]+:\\s*([^\"\\{].*?|\".*?\"|\\{.*?\\}))\\s*(,|$)", "g")) .reduce((p, n) => { const s = n.split(":"); + if (s[1].endsWith(',')) { + s[1] = s[1].substr(0, s[1].length -1); + } p[s[0].trim().toLowerCase()] = s[1].trim(); return p; }, {}); @@ -191,6 +194,7 @@ export async function importA1111(graph, parameters) { const vaeLoaderNode = LiteGraph.createNode("VAELoader"); const saveNode = LiteGraph.createNode("SaveImage"); let hrSamplerNode = null; + let hrSteps = null; const ceil64 = (v) => Math.ceil(v / 64) * 64; @@ -290,6 +294,9 @@ export async function importA1111(graph, parameters) { model(v) { setWidgetValue(ckptNode, "ckpt_name", v, true); }, + "vae"(v) { + setWidgetValue(vaeLoaderNode, "vae_name", v, true); + }, "cfg scale"(v) { setWidgetValue(samplerNode, "cfg", +v); }, @@ -316,6 +323,7 @@ export async function importA1111(graph, parameters) { const h = ceil64(+wxh[1]); const hrUp = popOpt("hires upscale"); const hrSz = popOpt("hires resize"); + hrSteps = popOpt("hires steps"); let hrMethod = popOpt("hires upscaler"); setWidgetValue(imageNode, "width", w); @@ -398,7 +406,7 @@ export async function importA1111(graph, parameters) { } if (hrSamplerNode) { - setWidgetValue(hrSamplerNode, "steps", getWidget(samplerNode, "steps").value); + setWidgetValue(hrSamplerNode, "steps", hrSteps? +hrSteps : getWidget(samplerNode, "steps").value); setWidgetValue(hrSamplerNode, "cfg", getWidget(samplerNode, "cfg").value); setWidgetValue(hrSamplerNode, "scheduler", getWidget(samplerNode, "scheduler").value); setWidgetValue(hrSamplerNode, "sampler_name", getWidget(samplerNode, "sampler_name").value); @@ -415,7 +423,7 @@ export async function importA1111(graph, parameters) { graph.arrange(); - for (const opt of ["model hash", "ensd"]) { + for (const opt of ["model hash", "ensd", "version", "vae hash", "ti hashes", "lora hashes", "hashes"]) { delete opts[opt]; } diff --git a/web/scripts/ui.js b/web/scripts/ui.js index d07d69dc8..72e43d357 100644 --- a/web/scripts/ui.js +++ b/web/scripts/ui.js @@ -1,5 +1,6 @@ import { api } from "./api.js"; import { ComfyDialog as _ComfyDialog } from "./ui/dialog.js"; +import { toggleSwitch } from "./ui/toggleSwitch.js"; import { ComfySettingsDialog } from "./ui/settings.js"; export const ComfyDialog = _ComfyDialog; @@ -89,12 +90,15 @@ function dragElement(dragEl, settings) { }).observe(dragEl); function ensureInBounds() { - if (dragEl.classList.contains("comfy-menu-manual-pos")) { + try { newPosX = Math.min(document.body.clientWidth - dragEl.clientWidth, Math.max(0, dragEl.offsetLeft)); newPosY = Math.min(document.body.clientHeight - dragEl.clientHeight, Math.max(0, dragEl.offsetTop)); positionElement(); } + catch(exception){ + // robust + } } function positionElement() { @@ -224,7 +228,7 @@ class ComfyList { $el("button", { textContent: "Load", onclick: async () => { - await app.loadGraphData(item.prompt[3].extra_pnginfo.workflow); + await app.loadGraphData(item.prompt[3].extra_pnginfo.workflow, true, false); if (item.outputs) { app.nodeOutputs = item.outputs; } @@ -368,18 +372,67 @@ export class ComfyUI { }, }); - this.menuContainer = $el("div.comfy-menu", {parent: document.body}, [ - $el("div.drag-handle", { + const autoQueueModeEl = toggleSwitch( + "autoQueueMode", + [ + { text: "instant", tooltip: "A new prompt will be queued as soon as the queue reaches 0" }, + { text: "change", tooltip: "A new prompt will be queued when the queue is at 0 and the graph is/has changed" }, + ], + { + onChange: (value) => { + this.autoQueueMode = value.item.value; + }, + } + ); + autoQueueModeEl.style.display = "none"; + + api.addEventListener("graphChanged", () => { + if (this.autoQueueMode === "change" && this.autoQueueEnabled === true) { + if (this.lastQueueSize === 0) { + this.graphHasChanged = false; + app.queuePrompt(0, this.batchCount); + } else { + this.graphHasChanged = true; + } + } + }); + + this.menuHamburger = $el( + "div.comfy-menu-hamburger", + { + parent: document.body, + onclick: () => { + this.menuContainer.style.display = "block"; + this.menuHamburger.style.display = "none"; + }, + }, + [$el("div"), $el("div"), $el("div")] + ); + + this.menuContainer = $el("div.comfy-menu", { parent: document.body }, [ + $el("div.drag-handle.comfy-menu-header", { style: { overflow: "hidden", position: "relative", width: "100%", cursor: "default" } - }, [ + }, [ $el("span.drag-handle"), - $el("span", {$: (q) => (this.queueSize = q)}), - $el("button.comfy-settings-btn", {textContent: "⚙️", onclick: () => this.settings.show()}), + $el("span.comfy-menu-queue-size", { $: (q) => (this.queueSize = q) }), + $el("div.comfy-menu-actions", [ + $el("button.comfy-settings-btn", { + textContent: "⚙️", + onclick: () => this.settings.show(), + }), + $el("button.comfy-close-menu-btn", { + textContent: "\u00d7", + onclick: () => { + this.menuContainer.style.display = "none"; + this.menuHamburger.style.display = "flex"; + }, + }), + ]), ]), $el("button.comfy-queue-btn", { id: "queue-button", @@ -394,6 +447,7 @@ export class ComfyUI { document.getElementById("extraOptions").style.display = i.srcElement.checked ? "block" : "none"; this.batchCount = i.srcElement.checked ? document.getElementById("batchCountInputRange").value : 1; document.getElementById("autoQueueCheckbox").checked = false; + this.autoQueueEnabled = false; }, }), ]), @@ -425,20 +479,22 @@ export class ComfyUI { }, }), ]), - $el("div",[ $el("label",{ for:"autoQueueCheckbox", innerHTML: "Auto Queue" - // textContent: "Auto Queue" }), $el("input", { id: "autoQueueCheckbox", type: "checkbox", checked: false, title: "Automatically queue prompt when the queue size hits 0", - + onchange: (e) => { + this.autoQueueEnabled = e.target.checked; + autoQueueModeEl.style.display = this.autoQueueEnabled ? "" : "none"; + } }), + autoQueueModeEl ]) ]), $el("div.comfy-menu-btns", [ @@ -541,16 +597,23 @@ export class ComfyUI { if (!confirmClear.value || confirm("Clear workflow?")) { app.clean(); app.graph.clear(); + app.resetView(); } } }), $el("button", { id: "comfy-load-default-button", textContent: "Load Default", onclick: async () => { if (!confirmClear.value || confirm("Load default workflow?")) { + app.resetView(); await app.loadGraphData() } } }), + $el("button", { + id: "comfy-reset-view-button", textContent: "Reset View", onclick: async () => { + app.resetView(); + } + }), ]); const devMode = this.settings.addSetting({ @@ -572,10 +635,13 @@ export class ComfyUI { if ( this.lastQueueSize != 0 && status.exec_info.queue_remaining == 0 && - document.getElementById("autoQueueCheckbox").checked && - ! app.lastExecutionError + this.autoQueueEnabled && + (this.autoQueueMode === "instant" || this.graphHasChanged) && + !app.lastExecutionError ) { app.queuePrompt(0, this.batchCount); + status.exec_info.queue_remaining += this.batchCount; + this.graphHasChanged = false; } this.lastQueueSize = status.exec_info.queue_remaining; } diff --git a/web/scripts/ui/settings.js b/web/scripts/ui/settings.js index 1cdba5cfe..9e9d13af0 100644 --- a/web/scripts/ui/settings.js +++ b/web/scripts/ui/settings.js @@ -16,7 +16,17 @@ export class ComfySettingsDialog extends ComfyDialog { }, [ $el("table.comfy-modal-content.comfy-table", [ - $el("caption", { textContent: "Settings" }), + $el( + "caption", + { textContent: "Settings" }, + $el("button.comfy-btn", { + type: "button", + textContent: "\u00d7", + onclick: () => { + this.element.close(); + }, + }) + ), $el("tbody", { $: (tbody) => (this.textElement = tbody) }), $el("button", { type: "button", diff --git a/web/scripts/ui/toggleSwitch.js b/web/scripts/ui/toggleSwitch.js new file mode 100644 index 000000000..59597ef90 --- /dev/null +++ b/web/scripts/ui/toggleSwitch.js @@ -0,0 +1,60 @@ +import { $el } from "../ui.js"; + +/** + * @typedef { { text: string, value?: string, tooltip?: string } } ToggleSwitchItem + */ +/** + * Creates a toggle switch element + * @param { string } name + * @param { Array void } [opts.onChange] + */ +export function toggleSwitch(name, items, { onChange } = {}) { + let selectedIndex; + let elements; + + function updateSelected(index) { + if (selectedIndex != null) { + elements[selectedIndex].classList.remove("comfy-toggle-selected"); + } + onChange?.({ item: items[index], prev: selectedIndex == null ? undefined : items[selectedIndex] }); + selectedIndex = index; + elements[selectedIndex].classList.add("comfy-toggle-selected"); + } + + elements = items.map((item, i) => { + if (typeof item === "string") item = { text: item }; + if (!item.value) item.value = item.text; + + const toggle = $el( + "label", + { + textContent: item.text, + title: item.tooltip ?? "", + }, + $el("input", { + name, + type: "radio", + value: item.value ?? item.text, + checked: item.selected, + onchange: () => { + updateSelected(i); + }, + }) + ); + if (item.selected) { + updateSelected(i); + } + return toggle; + }); + + const container = $el("div.comfy-toggle-switch", elements); + + if (selectedIndex == null) { + elements[0].children[0].checked = true; + updateSelected(0); + } + + return container; +} diff --git a/web/scripts/widgets.js b/web/scripts/widgets.js index 0529b1d80..6a6899705 100644 --- a/web/scripts/widgets.js +++ b/web/scripts/widgets.js @@ -81,6 +81,9 @@ export function addValueControlWidgets(node, targetWidget, defaultValue = "rando const isCombo = targetWidget.type === "combo"; let comboFilter; + if (isCombo) { + valueControl.options.values.push("increment-wrap"); + } if (isCombo && options.addFilterList !== false) { comboFilter = node.addWidget( "string", @@ -128,6 +131,12 @@ export function addValueControlWidgets(node, targetWidget, defaultValue = "rando case "increment": current_index += 1; break; + case "increment-wrap": + current_index += 1; + if ( current_index >= current_length ) { + current_index = 0; + } + break; case "decrement": current_index -= 1; break; @@ -220,7 +229,11 @@ function createIntWidget(node, inputName, inputData, app, isSeedInput) { val, function (v) { const s = this.options.step / 10; - this.value = Math.round(v / s) * s; + let sh = this.options.min % s; + if (isNaN(sh)) { + sh = 0; + } + this.value = Math.round((v - sh) / s) * s + sh; }, config ), @@ -295,10 +308,12 @@ export const ComfyWidgets = { let disable_rounding = app.ui.settings.getSettingValue("Comfy.DisableFloatRounding") if (precision == 0) precision = undefined; const { val, config } = getNumberDefaults(inputData, 0.5, precision, !disable_rounding); - return { widget: node.addWidget(widgetType, inputName, val, + return { widget: node.addWidget(widgetType, inputName, val, function (v) { if (config.round) { - this.value = Math.round(v/config.round)*config.round; + this.value = Math.round((v + Number.EPSILON)/config.round)*config.round; + if (this.value > config.max) this.value = config.max; + if (this.value < config.min) this.value = config.min; } else { this.value = v; } diff --git a/web/style.css b/web/style.css index 5c1133495..cf7a8b9ea 100644 --- a/web/style.css +++ b/web/style.css @@ -82,6 +82,24 @@ body { margin: 3px 3px 3px 4px; } +.comfy-menu-hamburger { + position: fixed; + top: 10px; + z-index: 9999; + right: 10px; + width: 30px; + display: none; + gap: 8px; + flex-direction: column; + cursor: pointer; +} +.comfy-menu-hamburger div { + height: 3px; + width: 100%; + border-radius: 20px; + background-color: white; +} + .comfy-menu { font-size: 15px; position: absolute; @@ -101,6 +119,44 @@ body { box-shadow: 3px 3px 8px rgba(0, 0, 0, 0.4); } +.comfy-menu-header { + display: flex; +} + +.comfy-menu-actions { + display: flex; + gap: 3px; + align-items: center; + height: 20px; + position: relative; + top: -1px; + font-size: 22px; +} + +.comfy-menu .comfy-menu-actions button { + background-color: rgba(0, 0, 0, 0); + padding: 0; + border: none; + cursor: pointer; + font-size: inherit; +} + +.comfy-menu .comfy-menu-actions .comfy-settings-btn { + font-size: 0.6em; +} + +button.comfy-close-menu-btn { + font-size: 1em; + line-height: 12px; + color: #ccc; + position: relative; + top: -1px; +} + +.comfy-menu-queue-size { + flex: auto; +} + .comfy-menu button, .comfy-modal button { font-size: 20px; @@ -139,17 +195,12 @@ body { .comfy-menu-btns button:hover, .comfy-menu .comfy-list button:hover, .comfy-modal button:hover, -.comfy-settings-btn:hover { +.comfy-menu-actions button:hover { filter: brightness(1.2); + will-change: transform; cursor: pointer; } -.comfy-menu span.drag-handle { - position: absolute; - top: 0; - left: 0; -} - span.drag-handle { width: 10px; height: 20px; @@ -214,15 +265,6 @@ span.drag-handle::after { font-size: 12px; } -button.comfy-settings-btn { - background-color: rgba(0, 0, 0, 0); - font-size: 12px; - padding: 0; - position: absolute; - right: 0; - border: none; -} - button.comfy-queue-btn { margin: 6px 0 !important; } @@ -268,7 +310,19 @@ button.comfy-queue-btn { } .comfy-menu span.drag-handle { - visibility: hidden + display: none; + } + + .comfy-menu-queue-size { + flex: unset; + } + + .comfy-menu-header { + justify-content: space-between; + } + .comfy-menu-actions { + gap: 10px; + font-size: 28px; } } @@ -319,7 +373,7 @@ dialog::backdrop { text-align: right; } -#comfy-settings-dialog button { +#comfy-settings-dialog tbody button, #comfy-settings-dialog table > button { background-color: var(--bg-color); border: 1px var(--border-color) solid; border-radius: 0; @@ -342,12 +396,33 @@ dialog::backdrop { } .comfy-table caption { + position: sticky; + top: 0; background-color: var(--bg-color); color: var(--input-text); font-size: 1rem; font-weight: bold; padding: 8px; text-align: center; + border-bottom: 1px solid var(--border-color); +} + +.comfy-table caption .comfy-btn { + position: absolute; + top: -2px; + right: 0; + bottom: 0; + cursor: pointer; + border: none; + height: 100%; + border-radius: 0; + aspect-ratio: 1/1; + user-select: none; + font-size: 20px; +} + +.comfy-table caption .comfy-btn:focus { + outline: none; } .comfy-table tr:nth-child(even) { @@ -388,11 +463,13 @@ dialog::backdrop { z-index: 9999 !important; background-color: var(--comfy-menu-bg) !important; filter: brightness(95%); + will-change: transform; } .litegraph.litecontextmenu .litemenu-entry:hover:not(.disabled):not(.separator) { background-color: var(--comfy-menu-bg) !important; filter: brightness(155%); + will-change: transform; color: var(--input-text); } @@ -453,10 +530,30 @@ dialog::backdrop { color: var(--input-text); background-color: var(--comfy-input-bg); filter: brightness(80%); + will-change: transform; padding-left: 0.2em; } .litegraph.lite-search-item.generic_type { color: var(--input-text); filter: brightness(50%); + will-change: transform; +} + +@media only screen and (max-width: 450px) { + #comfy-settings-dialog .comfy-table tbody { + display: grid; + } + #comfy-settings-dialog .comfy-table tr { + display: grid; + } + #comfy-settings-dialog tr > td:first-child { + text-align: center; + border-bottom: none; + padding-bottom: 0; + } + #comfy-settings-dialog tr > td:not(:first-child) { + text-align: center; + border-top: none; + } }