Merge remote-tracking branch 'upstream/master'

.github/workflows/stable-release.yml

@@ -2,9 +2,28 @@
 name: "Release Stable Version"
 
 on:
-  push:
+  workflow_dispatch:
-    tags:
+    inputs:
-      - 'v*'
+      git_tag:
+        description: 'Git tag'
+        required: true
+        type: string
+      cu:
+        description: 'CUDA version'
+        required: true
+        type: string
+        default: "121"
+      python_minor:
+        description: 'Python minor version'
+        required: true
+        type: string
+        default: "11"
+      python_patch:
+        description: 'Python patch version'
+        required: true
+        type: string
+        default: "9"
+
 
 jobs:
   package_comfy_windows:
@@ -13,69 +32,44 @@ jobs:
       packages: "write"
       pull-requests: "read"
     runs-on: windows-latest
-    strategy:
-      matrix:
-        python_version: [3.11.8]
-        cuda_version: [121]
     steps:
-      - name: Calculate Minor Version
-        shell: bash
-        run: |
-          # Extract the minor version from the Python version
-          MINOR_VERSION=$(echo "${{ matrix.python_version }}" | cut -d'.' -f2)
-          echo "MINOR_VERSION=$MINOR_VERSION" >> $GITHUB_ENV
-      - name: Setup Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python_version }}
-        
       - uses: actions/checkout@v4
         with:
+          ref: ${{ inputs.git_tag }}
           fetch-depth: 0
           persist-credentials: false
+      - uses: actions/cache/restore@v4
+        id: cache
+        with:
+          path: |
+            cu${{ inputs.cu }}_python_deps.tar
+            update_comfyui_and_python_dependencies.bat
+          key: ${{ runner.os }}-build-cu${{ inputs.cu }}-${{ inputs.python_minor }}
       - shell: bash
         run: |
-          echo "@echo off
+          mv cu${{ inputs.cu }}_python_deps.tar ../
-          call update_comfyui.bat nopause
-          echo -
-          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 -
-          pause
-          ..\python_embeded\python.exe -s -m pip install --upgrade torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu${{ matrix.cuda_version }} -r ../ComfyUI/requirements.txt pygit2
-          pause" > update_comfyui_and_python_dependencies.bat
-
-          python -m pip wheel --no-cache-dir torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu${{ matrix.cuda_version }} -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 cu${{ matrix.cuda_version }}_python_deps
-          mv cu${{ matrix.cuda_version }}_python_deps ../
           mv update_comfyui_and_python_dependencies.bat ../
           cd ..
+          tar xf cu${{ inputs.cu }}_python_deps.tar
           pwd
           ls
 
+      - shell: bash
+        run: |
+          cd ..
           cp -r ComfyUI ComfyUI_copy
-          curl https://www.python.org/ftp/python/${{ matrix.python_version }}/python-${{ matrix.python_version }}-embed-amd64.zip -o python_embeded.zip
+          curl https://www.python.org/ftp/python/3.${{ inputs.python_minor }}.${{ inputs.python_patch }}/python-3.${{ inputs.python_minor }}.${{ inputs.python_patch }}-embed-amd64.zip -o python_embeded.zip
           unzip python_embeded.zip -d python_embeded
           cd python_embeded
           echo ${{ env.MINOR_VERSION }}
-          echo 'import site' >> ./python3${{ env.MINOR_VERSION }}._pth
+          echo 'import site' >> ./python3${{ inputs.python_minor }}._pth
           curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
           ./python.exe get-pip.py
-          ./python.exe --version
+          ./python.exe -s -m pip install ../cu${{ inputs.cu }}_python_deps/*
-          echo "Pip version:"
+            sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
-          ./python.exe -m pip --version
-
-          set PATH=$PWD/Scripts:$PATH
-          echo $PATH
-          ./python.exe -s -m pip install ../cu${{ matrix.cuda_version }}_python_deps/*
-          sed -i '1i../ComfyUI' ./python3${{ env.MINOR_VERSION }}._pth
             cd ..
 
-          git clone https://github.com/comfyanonymous/taesd
+          git clone --depth 1 https://github.com/comfyanonymous/taesd
           cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
 
           mkdir ComfyUI_windows_portable
@@ -104,7 +98,7 @@ jobs:
         with:
           repo_token: ${{ secrets.GITHUB_TOKEN }}
           file: ComfyUI_windows_portable_nvidia.7z
-          tag: ${{ github.ref }}
+          tag: ${{ inputs.git_tag }}
           overwrite: true
           prerelease: true
           make_latest: false

.github/workflows/test-browser.yml

@@ -32,7 +32,7 @@ jobs:
         node-version: lts/*
     - uses: actions/setup-python@v4
       with:
-        python-version: '3.10'
+        python-version: '3.8'
     - name: Install requirements
       run: |
         python -m pip install --upgrade pip

.github/workflows/windows_release_nightly_pytorch.yml

@@ -55,7 +55,7 @@ jobs:
             sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
             cd ..
 
-            git clone https://github.com/comfyanonymous/taesd
+            git clone --depth 1 https://github.com/comfyanonymous/taesd
             cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
 
             mkdir ComfyUI_windows_portable_nightly_pytorch

.github/workflows/windows_release_package.yml

@@ -66,7 +66,7 @@ jobs:
             sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
             cd ..
 
-            git clone https://github.com/comfyanonymous/taesd
+            git clone --depth 1 https://github.com/comfyanonymous/taesd
             cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
 
             mkdir ComfyUI_windows_portable

README.md

@@ -165,20 +165,6 @@ You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS ve
 
 ```pip install torch-directml``` Then you can launch ComfyUI with: ```python main.py --directml```
 
-### I already have another UI for Stable Diffusion installed do I really have to install all of these dependencies?
-
-You don't. If you have another UI installed and working with its own python venv you can use that venv to run ComfyUI. You can open up your favorite terminal and activate it:
-
-```source path_to_other_sd_gui/venv/bin/activate```
-
-or on Windows:
-
-With Powershell: ```"path_to_other_sd_gui\venv\Scripts\Activate.ps1"```
-
-With cmd.exe: ```"path_to_other_sd_gui\venv\Scripts\activate.bat"```
-
-And then you can use that terminal to run ComfyUI without installing any dependencies. Note that the venv folder might be called something else depending on the SD UI.
-
 # Running
 
 ```python main.py```

comfy/ldm/aura/mmdit.py

@@ -9,6 +9,7 @@ import torch.nn.functional as F
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.ops
+import comfy.ldm.common_dit
 
 def modulate(x, shift, scale):
     return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
@@ -407,10 +408,7 @@ class MMDiT(nn.Module):
 
     def patchify(self, x):
         B, C, H, W = x.size()
-        pad_h = (self.patch_size - H % self.patch_size) % self.patch_size
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
-        pad_w = (self.patch_size - W % self.patch_size) % self.patch_size
-
-        x = torch.nn.functional.pad(x, (0, pad_w, 0, pad_h), mode='circular')
         x = x.view(
             B,
             C,

comfy/ldm/common_dit.py

@@ -0,0 +1,8 @@
+import torch
+
+def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
+    if padding_mode == "circular" and torch.jit.is_tracing() or torch.jit.is_scripting():
+        padding_mode = "reflect"
+    pad_h = (patch_size[0] - img.shape[-2] % patch_size[0]) % patch_size[0]
+    pad_w = (patch_size[1] - img.shape[-1] % patch_size[1]) % patch_size[1]
+    return torch.nn.functional.pad(img, (0, pad_w, 0, pad_h), mode=padding_mode)

comfy/ldm/flux/math.py

@@ -14,7 +14,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor) -> Tensor:
 
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
-    if comfy.model_management.is_device_mps(pos.device):
+    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu():
         device = torch.device("cpu")
     else:
         device = pos.device

comfy/ldm/flux/model.py

@@ -15,6 +15,7 @@ from .layers import (
 )
 
 from einops import rearrange, repeat
+import comfy.ldm.common_dit
 
 @dataclass
 class FluxParams:
@@ -42,7 +43,7 @@ class Flux(nn.Module):
         self.dtype = dtype
         params = FluxParams(**kwargs)
         self.params = params
-        self.in_channels = params.in_channels
+        self.in_channels = params.in_channels * 2 * 2
         self.out_channels = self.in_channels
         if params.hidden_size % params.num_heads != 0:
             raise ValueError(
@@ -125,10 +126,7 @@ class Flux(nn.Module):
     def forward(self, x, timestep, context, y, guidance, **kwargs):
         bs, c, h, w = x.shape
         patch_size = 2
-        pad_h = (patch_size - h % 2) % patch_size
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
-        pad_w = (patch_size - w % 2) % patch_size
-
-        x = torch.nn.functional.pad(x, (0, pad_w, 0, pad_h), mode='circular')
 
         img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
 

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -9,6 +9,7 @@ from .. import attention
 from einops import rearrange, repeat
 from .util import timestep_embedding
 import comfy.ops
+import comfy.ldm.common_dit
 
 def default(x, y):
     if x is not None:
@@ -111,9 +112,7 @@ class PatchEmbed(nn.Module):
         #             f"Input width ({W}) should be divisible by patch size ({self.patch_size[1]})."
         #         )
         if self.dynamic_img_pad:
-            pad_h = (self.patch_size[0] - H % self.patch_size[0]) % self.patch_size[0]
+            x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size, padding_mode=self.padding_mode)
-            pad_w = (self.patch_size[1] - W % self.patch_size[1]) % self.patch_size[1]
-            x = torch.nn.functional.pad(x, (0, pad_w, 0, pad_h), mode=self.padding_mode)
         x = self.proj(x)
         if self.flatten:
             x = x.flatten(2).transpose(1, 2)  # NCHW -> NLC

comfy/lora.py

@@ -288,4 +288,12 @@ def model_lora_keys_unet(model, key_map={}):
                 key_lora = k[len("diffusion_model."):-len(".weight")]
                 key_map["base_model.model.{}".format(key_lora)] = k #official hunyuan lora format
 
+    if isinstance(model, comfy.model_base.Flux): #Diffusers lora Flux
+        diffusers_keys = comfy.utils.flux_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
+        for k in diffusers_keys:
+            if k.endswith(".weight"):
+                to = diffusers_keys[k]
+                key_lora = "transformer.{}".format(k[:-len(".weight")]) #simpletrainer and probably regular diffusers flux lora format
+                key_map[key_lora] = to
+
     return key_map

comfy/model_base.py

@@ -84,6 +84,7 @@ class BaseModel(torch.nn.Module):
             if comfy.model_management.force_channels_last():
                 self.diffusion_model.to(memory_format=torch.channels_last)
                 logging.debug("using channels last mode for diffusion model")
+            logging.info("model weight dtype {}, manual cast: {}".format(self.get_dtype(), self.manual_cast_dtype))
         self.model_type = model_type
         self.model_sampling = model_sampling(model_config, model_type)
 
@@ -94,6 +95,7 @@ class BaseModel(torch.nn.Module):
         self.concat_keys = ()
         logging.info("model_type {}".format(model_type.name))
         logging.debug("adm {}".format(self.adm_channels))
+        self.memory_usage_factor = model_config.memory_usage_factor
 
     def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         sigma = t
@@ -252,11 +254,11 @@ class BaseModel(torch.nn.Module):
                 dtype = self.manual_cast_dtype
             #TODO: this needs to be tweaked
             area = input_shape[0] * math.prod(input_shape[2:])
-            return (area * comfy.model_management.dtype_size(dtype) / 50) * (1024 * 1024)
+            return (area * comfy.model_management.dtype_size(dtype) * 0.01 * self.memory_usage_factor) * (1024 * 1024)
         else:
             #TODO: this formula might be too aggressive since I tweaked the sub-quad and split algorithms to use less memory.
             area = input_shape[0] * math.prod(input_shape[2:])
-            return (area * 0.3) * (1024 * 1024)
+            return (area * 0.15 * self.memory_usage_factor) * (1024 * 1024)
 
 
 def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0, seed=None):
@@ -354,6 +356,7 @@ class SDXL(BaseModel):
         flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1)
         return torch.cat((clip_pooled.to(flat.device), flat), dim=1)
 
+
 class SVD_img2vid(BaseModel):
     def __init__(self, model_config, model_type=ModelType.V_PREDICTION_EDM, device=None):
         super().__init__(model_config, model_type, device=device)
@@ -594,17 +597,6 @@ class SD3(BaseModel):
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
 
-    def memory_required(self, input_shape):
-        if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
-            dtype = self.get_dtype()
-            if self.manual_cast_dtype is not None:
-                dtype = self.manual_cast_dtype
-            #TODO: this probably needs to be tweaked
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * comfy.model_management.dtype_size(dtype) * 0.012) * (1024 * 1024)
-        else:
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * 0.3) * (1024 * 1024)
 
 class AuraFlow(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
@@ -702,15 +694,3 @@ class Flux(BaseModel):
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([kwargs.get("guidance", 3.5)]))
         return out
-
-    def memory_required(self, input_shape):
-        if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
-            dtype = self.get_dtype()
-            if self.manual_cast_dtype is not None:
-                dtype = self.manual_cast_dtype
-            #TODO: this probably needs to be tweaked
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * comfy.model_management.dtype_size(dtype) * 0.020) * (1024 * 1024)
-        else:
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * 0.3) * (1024 * 1024)

comfy/model_detection.py

@@ -131,7 +131,7 @@ def detect_unet_config(state_dict, key_prefix):
     if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys: #Flux
         dit_config = {}
         dit_config["image_model"] = "flux"
-        dit_config["in_channels"] = 64
+        dit_config["in_channels"] = 16
         dit_config["vec_in_dim"] = 768
         dit_config["context_in_dim"] = 4096
         dit_config["hidden_size"] = 3072

comfy/model_management.py

@@ -352,6 +352,7 @@ def unload_model_clones(model, unload_weights_only=True, force_unload=True):
 def free_memory(memory_required, device, keep_loaded=[]):
     unloaded_model = []
     can_unload = []
+    unloaded_models = []
 
     for i in range(len(current_loaded_models) -1, -1, -1):
         shift_model = current_loaded_models[i]
@@ -369,7 +370,7 @@ def free_memory(memory_required, device, keep_loaded=[]):
         unloaded_model.append(i)
 
     for i in sorted(unloaded_model, reverse=True):
-        current_loaded_models.pop(i)
+        unloaded_models.append(current_loaded_models.pop(i))
 
     if len(unloaded_model) > 0:
         soft_empty_cache()
@@ -378,6 +379,7 @@ def free_memory(memory_required, device, keep_loaded=[]):
             mem_free_total, mem_free_torch = get_free_memory(device, torch_free_too=True)
             if mem_free_torch > mem_free_total * 0.25:
                 soft_empty_cache()
+    return unloaded_models
 
 def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None):
     global vram_state
@@ -421,6 +423,12 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
         for d in devs:
             if d != torch.device("cpu"):
                 free_memory(extra_mem, d, models_already_loaded)
+                free_mem = get_free_memory(d)
+                if free_mem < minimum_memory_required:
+                    logging.info("Unloading models for lowram load.") #TODO: partial model unloading when this case happens, also handle the opposite case where models can be unlowvramed.
+                    models_to_load = free_memory(minimum_memory_required, d)
+                    logging.info("{} models unloaded.".format(len(models_to_load)))
+        if len(models_to_load) == 0:
             return
 
     logging.info(f"Loading {len(models_to_load)} new model{'s' if len(models_to_load) > 1 else ''}")
@@ -450,7 +458,7 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
         if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):
             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 - minimum_memory_required)))
+            lowvram_model_memory = max(64 * (1024 * 1024), (current_free_mem - minimum_memory_required), min(current_free_mem * 0.4, current_free_mem - minimum_inference_memory()))
             if model_size <= lowvram_model_memory: #only switch to lowvram if really necessary
                 lowvram_model_memory = 0
 
@@ -527,6 +535,9 @@ def unet_inital_load_device(parameters, dtype):
     else:
         return cpu_dev
 
+def maximum_vram_for_weights(device=None):
+    return (get_total_memory(device) * 0.88 - minimum_inference_memory())
+
 def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, torch.bfloat16, torch.float32]):
     if args.bf16_unet:
         return torch.bfloat16
@@ -536,6 +547,21 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         return torch.float8_e4m3fn
     if args.fp8_e5m2_unet:
         return torch.float8_e5m2
+
+    fp8_dtype = None
+    try:
+        for dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
+            if dtype in supported_dtypes:
+                fp8_dtype = dtype
+                break
+    except:
+        pass
+
+    if fp8_dtype is not None:
+        free_model_memory = maximum_vram_for_weights(device)
+        if model_params * 2 > free_model_memory:
+            return fp8_dtype
+
     if should_use_fp16(device=device, model_params=model_params, manual_cast=True):
         if torch.float16 in supported_dtypes:
             return torch.float16
@@ -871,7 +897,7 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
             fp16_works = True
 
     if fp16_works or manual_cast:
-        free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory())
+        free_model_memory = maximum_vram_for_weights(device)
         if (not prioritize_performance) or model_params * 4 > free_model_memory:
             return True
 
@@ -910,17 +936,14 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     if is_intel_xpu():
         return True
 
-    if device is None:
+    props = torch.cuda.get_device_properties("cuda")
-        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())
+        free_model_memory = maximum_vram_for_weights(device)
         if (not prioritize_performance) or model_params * 4 > free_model_memory:
             return True
 

comfy/samplers.py

@@ -171,7 +171,7 @@ def calc_cond_batch(model, conds, x_in, timestep, model_options):
         for i in range(1, len(to_batch_temp) + 1):
             batch_amount = to_batch_temp[:len(to_batch_temp)//i]
             input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]
-            if model.memory_required(input_shape) < free_memory:
+            if model.memory_required(input_shape) * 1.5 < free_memory:
                 to_batch = batch_amount
                 break
 

comfy/sd.py

@@ -510,13 +510,18 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
 
     diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd)
     parameters = comfy.utils.calculate_parameters(sd, diffusion_model_prefix)
+    weight_dtype = comfy.utils.weight_dtype(sd, diffusion_model_prefix)
     load_device = model_management.get_torch_device()
 
     model_config = model_detection.model_config_from_unet(sd, diffusion_model_prefix)
     if model_config is None:
         raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
 
-    unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=model_config.supported_inference_dtypes)
+    unet_weight_dtype = list(model_config.supported_inference_dtypes)
+    if weight_dtype is not None:
+        unet_weight_dtype.append(weight_dtype)
+
+    unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=unet_weight_dtype)
     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)
 

comfy/supported_models.py

@@ -31,6 +31,7 @@ class SD15(supported_models_base.BASE):
     }
 
     latent_format = latent_formats.SD15
+    memory_usage_factor = 1.0
 
     def process_clip_state_dict(self, state_dict):
         k = list(state_dict.keys())
@@ -77,6 +78,7 @@ class SD20(supported_models_base.BASE):
     }
 
     latent_format = latent_formats.SD15
+    memory_usage_factor = 1.0
 
     def model_type(self, state_dict, prefix=""):
         if self.unet_config["in_channels"] == 4: #SD2.0 inpainting models are not v prediction
@@ -140,6 +142,7 @@ class SDXLRefiner(supported_models_base.BASE):
     }
 
     latent_format = latent_formats.SDXL
+    memory_usage_factor = 1.0
 
     def get_model(self, state_dict, prefix="", device=None):
         return model_base.SDXLRefiner(self, device=device)
@@ -178,6 +181,8 @@ class SDXL(supported_models_base.BASE):
 
     latent_format = latent_formats.SDXL
 
+    memory_usage_factor = 0.7
+
     def model_type(self, state_dict, prefix=""):
         if 'edm_mean' in state_dict and 'edm_std' in state_dict: #Playground V2.5
             self.latent_format = latent_formats.SDXL_Playground_2_5()
@@ -505,6 +510,9 @@ class SD3(supported_models_base.BASE):
 
     unet_extra_config = {}
     latent_format = latent_formats.SD3
+
+    memory_usage_factor = 1.2
+
     text_encoder_key_prefix = ["text_encoders."]
 
     def get_model(self, state_dict, prefix="", device=None):
@@ -631,6 +639,9 @@ class Flux(supported_models_base.BASE):
 
     unet_extra_config = {}
     latent_format = latent_formats.Flux
+
+    memory_usage_factor = 2.8
+
     supported_inference_dtypes = [torch.bfloat16, torch.float32]
 
     vae_key_prefix = ["vae."]
@@ -641,7 +652,11 @@ class Flux(supported_models_base.BASE):
         return out
 
     def clip_target(self, state_dict={}):
-        return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.FluxClipModel)
+        pref = self.text_encoder_key_prefix[0]
+        t5_key = "{}t5xxl.transformer.encoder.final_layer_norm.weight".format(pref)
+        if t5_key in state_dict:
+            dtype_t5 = state_dict[t5_key].dtype
+        return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.flux_clip(dtype_t5=dtype_t5))
 
 class FluxSchnell(Flux):
     unet_config = {

comfy/supported_models_base.py

@@ -27,6 +27,8 @@ class BASE:
     text_encoder_key_prefix = ["cond_stage_model."]
     supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
 
+    memory_usage_factor = 2.0
+
     manual_cast_dtype = None
 
     @classmethod

comfy/text_encoders/flux.py

@@ -52,9 +52,9 @@ class FluxClipModel(torch.nn.Module):
 
     def encode_token_weights(self, token_weight_pairs):
         token_weight_pairs_l = token_weight_pairs["l"]
-        token_weight_pars_t5 = token_weight_pairs["t5xxl"]
+        token_weight_pairs_t5 = token_weight_pairs["t5xxl"]
 
-        t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pars_t5)
+        t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pairs_t5)
         l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs_l)
         return t5_out, l_pooled
 

comfy/text_encoders/sd3_clip.py

@@ -81,7 +81,7 @@ class SD3ClipModel(torch.nn.Module):
     def encode_token_weights(self, token_weight_pairs):
         token_weight_pairs_l = token_weight_pairs["l"]
         token_weight_pairs_g = token_weight_pairs["g"]
-        token_weight_pars_t5 = token_weight_pairs["t5xxl"]
+        token_weight_pairs_t5 = token_weight_pairs["t5xxl"]
         lg_out = None
         pooled = None
         out = None
@@ -108,7 +108,7 @@ class SD3ClipModel(torch.nn.Module):
             pooled = torch.cat((l_pooled, g_pooled), dim=-1)
 
         if self.t5xxl is not None:
-            t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pars_t5)
+            t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pairs_t5)
             if lg_out is not None:
                 out = torch.cat([lg_out, t5_out], dim=-2)
             else:

comfy/utils.py

@@ -40,9 +40,22 @@ def calculate_parameters(sd, prefix=""):
     params = 0
     for k in sd.keys():
         if k.startswith(prefix):
-            params += sd[k].nelement()
+            w = sd[k]
+            params += w.nelement()
     return params
 
+def weight_dtype(sd, prefix=""):
+    dtypes = {}
+    for k in sd.keys():
+        if k.startswith(prefix):
+            w = sd[k]
+            dtypes[w.dtype] = dtypes.get(w.dtype, 0) + 1
+
+    if len(dtypes) == 0:
+        return None
+
+    return max(dtypes, key=dtypes.get)
+
 def state_dict_key_replace(state_dict, keys_to_replace):
     for x in keys_to_replace:
         if x in state_dict:
@@ -402,6 +415,59 @@ def auraflow_to_diffusers(mmdit_config, output_prefix=""):
 
     return key_map
 
+def flux_to_diffusers(mmdit_config, output_prefix=""):
+    n_double_layers = mmdit_config.get("depth", 0)
+    n_single_layers = mmdit_config.get("depth_single_blocks", 0)
+    hidden_size = mmdit_config.get("hidden_size", 0)
+
+    key_map = {}
+    for index in range(n_double_layers):
+        prefix_from = "transformer_blocks.{}".format(index)
+        prefix_to = "{}double_blocks.{}".format(output_prefix, index)
+
+        for end in ("weight", "bias"):
+            k = "{}.attn.".format(prefix_from)
+            qkv = "{}.img_attn.qkv.{}".format(prefix_to, end)
+            key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
+            key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
+            key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
+
+        block_map = {"attn.to_out.0.weight": "img_attn.proj.weight",
+                     "attn.to_out.0.bias": "img_attn.proj.bias",
+                    }
+
+        for k in block_map:
+            key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k])
+
+    for index in range(n_single_layers):
+        prefix_from = "single_transformer_blocks.{}".format(index)
+        prefix_to = "{}single_blocks.{}".format(output_prefix, index)
+
+        for end in ("weight", "bias"):
+            k = "{}.attn.".format(prefix_from)
+            qkv = "{}.linear1.{}".format(prefix_to, end)
+            key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
+            key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
+            key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
+            key_map["{}proj_mlp.{}".format(k, end)] = (qkv, (0, hidden_size * 3, hidden_size))
+
+        block_map = {#TODO
+                    }
+
+        for k in block_map:
+            key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k])
+
+    MAP_BASIC = { #TODO
+    }
+
+    for k in MAP_BASIC:
+        if len(k) > 2:
+            key_map[k[1]] = ("{}{}".format(output_prefix, k[0]), None, k[2])
+        else:
+            key_map[k[1]] = "{}{}".format(output_prefix, k[0])
+
+    return key_map
+
 def repeat_to_batch_size(tensor, batch_size, dim=0):
     if tensor.shape[dim] > batch_size:
         return tensor.narrow(dim, 0, batch_size)

comfy_extras/nodes_audio.py

@@ -7,6 +7,7 @@ import io
 import json
 import struct
 import random
+import hashlib
 from comfy.cli_args import args
 
 class EmptyLatentAudio:

comfy_extras/nodes_flux.py

@@ -1,3 +1,4 @@
+import node_helpers
 
 class CLIPTextEncodeFlux:
     @classmethod
@@ -11,7 +12,7 @@ class CLIPTextEncodeFlux:
     RETURN_TYPES = ("CONDITIONING",)
     FUNCTION = "encode"
 
-    CATEGORY = "advanced/conditioning"
+    CATEGORY = "advanced/conditioning/flux"
 
     def encode(self, clip, clip_l, t5xxl, guidance):
         tokens = clip.tokenize(clip_l)
@@ -22,6 +23,25 @@ class CLIPTextEncodeFlux:
         output["guidance"] = guidance
         return ([[cond, output]], )
 
+class FluxGuidance:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "conditioning": ("CONDITIONING", ),
+            "guidance": ("FLOAT", {"default": 3.5, "min": 0.0, "max": 100.0, "step": 0.1}),
+            }}
+
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "append"
+
+    CATEGORY = "advanced/conditioning/flux"
+
+    def append(self, conditioning, guidance):
+        c = node_helpers.conditioning_set_values(conditioning, {"guidance": guidance})
+        return (c, )
+
+
 NODE_CLASS_MAPPINGS = {
     "CLIPTextEncodeFlux": CLIPTextEncodeFlux,
+    "FluxGuidance": FluxGuidance,
 }

comfy_extras/nodes_model_advanced.py

@@ -2,6 +2,7 @@ import folder_paths
 import comfy.sd
 import comfy.model_sampling
 import comfy.latent_formats
+import nodes
 import torch
 
 class LCM(comfy.model_sampling.EPS):
@@ -170,6 +171,42 @@ class ModelSamplingAuraFlow(ModelSamplingSD3):
     def patch_aura(self, model, shift):
         return self.patch(model, shift, multiplier=1.0)
 
+class ModelSamplingFlux:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "max_shift": ("FLOAT", {"default": 1.15, "min": 0.0, "max": 100.0, "step":0.01}),
+                              "base_shift": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 100.0, "step":0.01}),
+                              "width": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}),
+                              "height": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}),
+                              }}
+
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "advanced/model"
+
+    def patch(self, model, max_shift, base_shift, width, height):
+        m = model.clone()
+
+        x1 = 256
+        x2 = 4096
+        mm = (max_shift - base_shift) / (x2 - x1)
+        b = base_shift - mm * x1
+        shift = (width * height / (8 * 8 * 2 * 2)) * mm + b
+
+        sampling_base = comfy.model_sampling.ModelSamplingFlux
+        sampling_type = comfy.model_sampling.CONST
+
+        class ModelSamplingAdvanced(sampling_base, sampling_type):
+            pass
+
+        model_sampling = ModelSamplingAdvanced(model.model.model_config)
+        model_sampling.set_parameters(shift=shift)
+        m.add_object_patch("model_sampling", model_sampling)
+        return (m, )
+
+
 class ModelSamplingContinuousEDM:
     @classmethod
     def INPUT_TYPES(s):
@@ -284,5 +321,6 @@ NODE_CLASS_MAPPINGS = {
     "ModelSamplingStableCascade": ModelSamplingStableCascade,
     "ModelSamplingSD3": ModelSamplingSD3,
     "ModelSamplingAuraFlow": ModelSamplingAuraFlow,
+    "ModelSamplingFlux": ModelSamplingFlux,
     "RescaleCFG": RescaleCFG,
 }

comfy_extras/nodes_model_merging.py

@@ -264,6 +264,7 @@ class CLIPSave:
 
         metadata = {}
         if not args.disable_metadata:
+            metadata["format"] = "pt"
             metadata["prompt"] = prompt_info
             if extra_pnginfo is not None:
                 for x in extra_pnginfo:

comfy_extras/nodes_model_merging_model_specific.py

@@ -75,9 +75,36 @@ class ModelMergeSD3_2B(comfy_extras.nodes_model_merging.ModelMergeBlocks):
 
         return {"required": arg_dict}
 
+class ModelMergeFlux1(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["img_in."] = argument
+        arg_dict["time_in."] = argument
+        arg_dict["guidance_in"] = argument
+        arg_dict["vector_in."] = argument
+        arg_dict["txt_in."] = argument
+
+        for i in range(19):
+            arg_dict["double_blocks.{}.".format(i)] = argument
+
+        for i in range(38):
+            arg_dict["single_blocks.{}.".format(i)] = argument
+
+        arg_dict["final_layer."] = argument
+
+        return {"required": arg_dict}
+
 NODE_CLASS_MAPPINGS = {
     "ModelMergeSD1": ModelMergeSD1,
     "ModelMergeSD2": ModelMergeSD1, #SD1 and SD2 have the same blocks
     "ModelMergeSDXL": ModelMergeSDXL,
     "ModelMergeSD3_2B": ModelMergeSD3_2B,
+    "ModelMergeFlux1": ModelMergeFlux1,
 }

comfy_extras/nodes_pag.py

@@ -12,7 +12,7 @@ class PerturbedAttentionGuidance:
         return {
             "required": {
                 "model": ("MODEL",),
-                "scale": ("FLOAT", {"default": 3.0, "min": 0.0, "max": 100.0, "step": 0.1, "round": 0.01}),
+                "scale": ("FLOAT", {"default": 3.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": 0.01}),
             }
         }
 

comfy_extras/nodes_sag.py

@@ -96,7 +96,7 @@ class SelfAttentionGuidance:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "model": ("MODEL",),
-                             "scale": ("FLOAT", {"default": 0.5, "min": -2.0, "max": 5.0, "step": 0.1}),
+                             "scale": ("FLOAT", {"default": 0.5, "min": -2.0, "max": 5.0, "step": 0.01}),
                              "blur_sigma": ("FLOAT", {"default": 2.0, "min": 0.0, "max": 10.0, "step": 0.1}),
                               }}
     RETURN_TYPES = ("MODEL",)

comfy_extras/nodes_sd3.py

@@ -27,8 +27,8 @@ class EmptySD3LatentImage:
 
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "width": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}),
+        return {"required": { "width": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                              "height": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}),
+                              "height": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
                               "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096})}}
     RETURN_TYPES = ("LATENT",)
     FUNCTION = "generate"

execution.py

@@ -188,6 +188,11 @@ def recursive_execute(server, prompt, outputs, current_item, extra_data, execute
             "current_inputs": input_data_formatted,
             "current_outputs": output_data_formatted
         }
+
+        if isinstance(ex, comfy.model_management.OOM_EXCEPTION):
+            logging.error("Got an OOM, unloading all loaded models.")
+            comfy.model_management.unload_all_models()
+
         return (False, error_details, ex)
 
     executed.add(unique_id)

server.py

@@ -127,7 +127,11 @@ class PromptServer():
 
         @routes.get("/")
         async def get_root(request):
-            return web.FileResponse(os.path.join(self.web_root, "index.html"))
+            response = web.FileResponse(os.path.join(self.web_root, "index.html"))
+            response.headers['Cache-Control'] = 'no-cache'
+            response.headers["Pragma"] = "no-cache"
+            response.headers["Expires"] = "0"
+            return response
 
         @routes.get("/embeddings")
         def get_embeddings(self):

web/scripts/pnginfo.js

@@ -190,9 +190,10 @@ function parseVorbisComment(dataView) {
 		const comment = getString(dataView, offset, commentLength);
 		offset += commentLength;
 
-		const [key, value] = comment.split('=');
+		const ind = comment.indexOf('=')
+		const key = comment.substring(0, ind);
 
-		comments[key] = value;
+		comments[key] = comment.substring(ind+1);
 	}
 
 	return comments;