diff --git a/.ci/nightly/update_windows/update_comfyui_and_python_dependencies.bat b/.ci/nightly/update_windows/update_comfyui_and_python_dependencies.bat deleted file mode 100755 index 94f5d1023..000000000 --- a/.ci/nightly/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 --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cu121 -r ../ComfyUI/requirements.txt pygit2 -pause diff --git a/.ci/nightly/windows_base_files/run_nvidia_gpu.bat b/.ci/nightly/windows_base_files/run_nvidia_gpu.bat deleted file mode 100755 index 8ee2f3402..000000000 --- a/.ci/nightly/windows_base_files/run_nvidia_gpu.bat +++ /dev/null @@ -1,2 +0,0 @@ -.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --use-pytorch-cross-attention -pause diff --git a/.github/workflows/windows_release_nightly_pytorch.yml b/.github/workflows/windows_release_nightly_pytorch.yml index b793f7fe2..90e09d27a 100644 --- a/.github/workflows/windows_release_nightly_pytorch.yml +++ b/.github/workflows/windows_release_nightly_pytorch.yml @@ -2,6 +2,24 @@ name: "Windows Release Nightly pytorch" on: workflow_dispatch: + inputs: + cu: + description: 'cuda version' + required: true + type: string + default: "121" + + python_minor: + description: 'python minor version' + required: true + type: string + default: "12" + + python_patch: + description: 'python patch version' + required: true + type: string + default: "1" # push: # branches: # - master @@ -20,21 +38,21 @@ jobs: persist-credentials: false - uses: actions/setup-python@v4 with: - python-version: '3.11.6' + python-version: 3.${{ inputs.python_minor }}.${{ inputs.python_patch }} - shell: bash run: | cd .. cp -r ComfyUI ComfyUI_copy - curl https://www.python.org/ftp/python/3.11.6/python-3.11.6-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 'import site' >> ./python311._pth + echo 'import site' >> ./python3${{ inputs.python_minor }}._pth curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py ./python.exe get-pip.py - python -m pip wheel torch torchvision torchaudio aiohttp==3.8.5 --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu121 -r ../ComfyUI/requirements.txt pygit2 -w ../temp_wheel_dir + python -m pip wheel torch torchvision torchaudio --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu${{ inputs.cu }} -r ../ComfyUI/requirements.txt pygit2 -w ../temp_wheel_dir ls ../temp_wheel_dir ./python.exe -s -m pip install --pre ../temp_wheel_dir/* - sed -i '1i../ComfyUI' ./python311._pth + sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth cd .. git clone https://github.com/comfyanonymous/taesd @@ -49,9 +67,10 @@ jobs: mkdir update cp -r ComfyUI/.ci/update_windows/* ./update/ cp -r ComfyUI/.ci/windows_base_files/* ./ - cp -r ComfyUI/.ci/nightly/update_windows/* ./update/ - cp -r ComfyUI/.ci/nightly/windows_base_files/* ./ + echo "..\python_embeded\python.exe .\update.py ..\ComfyUI\\ + ..\python_embeded\python.exe -s -m pip install --upgrade --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cu${{ inputs.cu }} -r ../ComfyUI/requirements.txt pygit2 + pause" > ./update/update_comfyui_and_python_dependencies.bat cd .. "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma -mx=8 -mfb=64 -md=32m -ms=on -mf=BCJ2 ComfyUI_windows_portable_nightly_pytorch.7z ComfyUI_windows_portable_nightly_pytorch diff --git a/README.md b/README.md index 450a012bb..167214c05 100644 --- a/README.md +++ b/README.md @@ -93,23 +93,27 @@ Put your SD checkpoints (the huge ckpt/safetensors files) in: models/checkpoints Put your VAE in: models/vae -Note: pytorch does not support python 3.12 yet so make sure your python version is 3.11 or earlier. +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``` -This is the command to install the nightly with ROCm 5.7 that might have some performance improvements: +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: ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm5.7``` ### NVIDIA -Nvidia users should install pytorch using this command: +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: + +```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121``` + #### Troubleshooting If you get the "Torch not compiled with CUDA enabled" error, uninstall torch with: diff --git a/comfy/cldm/cldm.py b/comfy/cldm/cldm.py index bbe5891e6..5eee5a51c 100644 --- a/comfy/cldm/cldm.py +++ b/comfy/cldm/cldm.py @@ -53,7 +53,7 @@ class ControlNet(nn.Module): transformer_depth_middle=None, transformer_depth_output=None, device=None, - operations=comfy.ops, + operations=comfy.ops.disable_weight_init, **kwargs, ): super().__init__() @@ -141,24 +141,24 @@ class ControlNet(nn.Module): ) ] ) - self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels, operations=operations)]) + self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels, operations=operations, dtype=self.dtype, device=device)]) self.input_hint_block = TimestepEmbedSequential( - operations.conv_nd(dims, hint_channels, 16, 3, padding=1), + operations.conv_nd(dims, hint_channels, 16, 3, padding=1, dtype=self.dtype, device=device), nn.SiLU(), - operations.conv_nd(dims, 16, 16, 3, padding=1), + operations.conv_nd(dims, 16, 16, 3, padding=1, dtype=self.dtype, device=device), nn.SiLU(), - operations.conv_nd(dims, 16, 32, 3, padding=1, stride=2), + operations.conv_nd(dims, 16, 32, 3, padding=1, stride=2, dtype=self.dtype, device=device), nn.SiLU(), - operations.conv_nd(dims, 32, 32, 3, padding=1), + operations.conv_nd(dims, 32, 32, 3, padding=1, dtype=self.dtype, device=device), nn.SiLU(), - operations.conv_nd(dims, 32, 96, 3, padding=1, stride=2), + operations.conv_nd(dims, 32, 96, 3, padding=1, stride=2, dtype=self.dtype, device=device), nn.SiLU(), - operations.conv_nd(dims, 96, 96, 3, padding=1), + operations.conv_nd(dims, 96, 96, 3, padding=1, dtype=self.dtype, device=device), nn.SiLU(), - operations.conv_nd(dims, 96, 256, 3, padding=1, stride=2), + operations.conv_nd(dims, 96, 256, 3, padding=1, stride=2, dtype=self.dtype, device=device), nn.SiLU(), - zero_module(operations.conv_nd(dims, 256, model_channels, 3, padding=1)) + operations.conv_nd(dims, 256, model_channels, 3, padding=1, dtype=self.dtype, device=device) ) self._feature_size = model_channels @@ -206,7 +206,7 @@ class ControlNet(nn.Module): ) ) self.input_blocks.append(TimestepEmbedSequential(*layers)) - self.zero_convs.append(self.make_zero_conv(ch, operations=operations)) + self.zero_convs.append(self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device)) self._feature_size += ch input_block_chans.append(ch) if level != len(channel_mult) - 1: @@ -234,7 +234,7 @@ class ControlNet(nn.Module): ) ch = out_ch input_block_chans.append(ch) - self.zero_convs.append(self.make_zero_conv(ch, operations=operations)) + self.zero_convs.append(self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device)) ds *= 2 self._feature_size += ch @@ -276,11 +276,11 @@ class ControlNet(nn.Module): operations=operations )] self.middle_block = TimestepEmbedSequential(*mid_block) - self.middle_block_out = self.make_zero_conv(ch, operations=operations) + self.middle_block_out = self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device) self._feature_size += ch - def make_zero_conv(self, channels, operations=None): - return TimestepEmbedSequential(zero_module(operations.conv_nd(self.dims, channels, channels, 1, padding=0))) + def make_zero_conv(self, channels, operations=None, dtype=None, device=None): + return TimestepEmbedSequential(operations.conv_nd(self.dims, channels, channels, 1, padding=0, dtype=dtype, device=device)) def forward(self, x, hint, timesteps, context, y=None, **kwargs): t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 58d034802..d9c8668f4 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -57,6 +57,7 @@ fp_group.add_argument("--force-fp16", action="store_true", help="Force fp16.") fpunet_group = parser.add_mutually_exclusive_group() fpunet_group.add_argument("--bf16-unet", action="store_true", help="Run the UNET in bf16. This should only be used for testing stuff.") +fpunet_group.add_argument("--fp16-unet", action="store_true", help="Store unet weights in fp16.") fpunet_group.add_argument("--fp8_e4m3fn-unet", action="store_true", help="Store unet weights in fp8_e4m3fn.") fpunet_group.add_argument("--fp8_e5m2-unet", action="store_true", help="Store unet weights in fp8_e5m2.") diff --git a/comfy/clip_model.py b/comfy/clip_model.py new file mode 100644 index 000000000..850b5fdbe --- /dev/null +++ b/comfy/clip_model.py @@ -0,0 +1,188 @@ +import torch +from comfy.ldm.modules.attention import optimized_attention_for_device + +class CLIPAttention(torch.nn.Module): + def __init__(self, embed_dim, heads, dtype, device, operations): + super().__init__() + + self.heads = heads + self.q_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device) + self.k_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device) + self.v_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device) + + self.out_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device) + + def forward(self, x, mask=None, optimized_attention=None): + q = self.q_proj(x) + k = self.k_proj(x) + v = self.v_proj(x) + + out = optimized_attention(q, k, v, self.heads, mask) + return self.out_proj(out) + +ACTIVATIONS = {"quick_gelu": lambda a: a * torch.sigmoid(1.702 * a), + "gelu": torch.nn.functional.gelu, +} + +class CLIPMLP(torch.nn.Module): + def __init__(self, embed_dim, intermediate_size, activation, dtype, device, operations): + super().__init__() + self.fc1 = operations.Linear(embed_dim, intermediate_size, bias=True, dtype=dtype, device=device) + self.activation = ACTIVATIONS[activation] + self.fc2 = operations.Linear(intermediate_size, embed_dim, bias=True, dtype=dtype, device=device) + + def forward(self, x): + x = self.fc1(x) + x = self.activation(x) + x = self.fc2(x) + return x + +class CLIPLayer(torch.nn.Module): + def __init__(self, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations): + super().__init__() + self.layer_norm1 = operations.LayerNorm(embed_dim, dtype=dtype, device=device) + self.self_attn = CLIPAttention(embed_dim, heads, dtype, device, operations) + self.layer_norm2 = operations.LayerNorm(embed_dim, dtype=dtype, device=device) + self.mlp = CLIPMLP(embed_dim, intermediate_size, intermediate_activation, dtype, device, operations) + + def forward(self, x, mask=None, optimized_attention=None): + x += self.self_attn(self.layer_norm1(x), mask, optimized_attention) + x += self.mlp(self.layer_norm2(x)) + return x + + +class CLIPEncoder(torch.nn.Module): + def __init__(self, num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations): + super().__init__() + self.layers = torch.nn.ModuleList([CLIPLayer(embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations) for i in range(num_layers)]) + + def forward(self, x, mask=None, intermediate_output=None): + optimized_attention = optimized_attention_for_device(x.device, mask=mask is not None) + + if intermediate_output is not None: + if intermediate_output < 0: + intermediate_output = len(self.layers) + intermediate_output + + intermediate = None + for i, l in enumerate(self.layers): + x = l(x, mask, optimized_attention) + if i == intermediate_output: + intermediate = x.clone() + return x, intermediate + +class CLIPEmbeddings(torch.nn.Module): + def __init__(self, embed_dim, vocab_size=49408, num_positions=77, dtype=None, device=None): + super().__init__() + self.token_embedding = torch.nn.Embedding(vocab_size, embed_dim, dtype=dtype, device=device) + self.position_embedding = torch.nn.Embedding(num_positions, embed_dim, dtype=dtype, device=device) + + def forward(self, input_tokens): + return self.token_embedding(input_tokens) + self.position_embedding.weight + + +class CLIPTextModel_(torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + num_layers = config_dict["num_hidden_layers"] + embed_dim = config_dict["hidden_size"] + heads = config_dict["num_attention_heads"] + intermediate_size = config_dict["intermediate_size"] + intermediate_activation = config_dict["hidden_act"] + + super().__init__() + self.embeddings = CLIPEmbeddings(embed_dim, dtype=torch.float32, device=device) + self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations) + self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device) + + def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True): + 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 = 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) + if mask is not None: + mask += causal_mask + else: + mask = causal_mask + + x, i = self.encoder(x, mask=mask, intermediate_output=intermediate_output) + x = self.final_layer_norm(x) + if i is not None and final_layer_norm_intermediate: + i = self.final_layer_norm(i) + + pooled_output = x[torch.arange(x.shape[0], device=x.device), input_tokens.to(dtype=torch.int, device=x.device).argmax(dim=-1),] + return x, i, pooled_output + +class CLIPTextModel(torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + self.num_layers = config_dict["num_hidden_layers"] + self.text_model = CLIPTextModel_(config_dict, dtype, device, operations) + self.dtype = dtype + + def get_input_embeddings(self): + return self.text_model.embeddings.token_embedding + + def set_input_embeddings(self, embeddings): + self.text_model.embeddings.token_embedding = embeddings + + def forward(self, *args, **kwargs): + return self.text_model(*args, **kwargs) + +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): + super().__init__() + self.class_embedding = torch.nn.Parameter(torch.empty(embed_dim, dtype=dtype, device=device)) + + self.patch_embedding = operations.Conv2d( + in_channels=num_channels, + out_channels=embed_dim, + kernel_size=patch_size, + stride=patch_size, + bias=False, + dtype=dtype, + device=device + ) + + num_patches = (image_size // patch_size) ** 2 + num_positions = num_patches + 1 + self.position_embedding = torch.nn.Embedding(num_positions, embed_dim, dtype=dtype, device=device) + + def forward(self, pixel_values): + embeds = self.patch_embedding(pixel_values).flatten(2).transpose(1, 2) + return torch.cat([self.class_embedding.expand(pixel_values.shape[0], 1, -1), embeds], dim=1) + self.position_embedding.weight + + +class CLIPVision(torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + num_layers = config_dict["num_hidden_layers"] + embed_dim = config_dict["hidden_size"] + heads = config_dict["num_attention_heads"] + intermediate_size = config_dict["intermediate_size"] + intermediate_activation = config_dict["hidden_act"] + + self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], dtype=torch.float32, device=device, operations=operations) + self.pre_layrnorm = operations.LayerNorm(embed_dim) + self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations) + self.post_layernorm = operations.LayerNorm(embed_dim) + + def forward(self, pixel_values, attention_mask=None, intermediate_output=None): + x = self.embeddings(pixel_values) + x = self.pre_layrnorm(x) + #TODO: attention_mask? + x, i = self.encoder(x, mask=None, intermediate_output=intermediate_output) + pooled_output = self.post_layernorm(x[:, 0, :]) + return x, i, pooled_output + +class CLIPVisionModelProjection(torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + self.vision_model = CLIPVision(config_dict, dtype, device, operations) + self.visual_projection = operations.Linear(config_dict["hidden_size"], config_dict["projection_dim"], bias=False) + + def forward(self, *args, **kwargs): + x = self.vision_model(*args, **kwargs) + out = self.visual_projection(x[2]) + return (x[0], x[1], out) diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py index 9e2e03d72..ba8a3a8d5 100644 --- a/comfy/clip_vision.py +++ b/comfy/clip_vision.py @@ -1,13 +1,20 @@ -from transformers import CLIPVisionModelWithProjection, CLIPVisionConfig, modeling_utils from .utils import load_torch_file, transformers_convert, common_upscale import os import torch import contextlib +import json import comfy.ops import comfy.model_patcher import comfy.model_management import comfy.utils +import comfy.clip_model + +class Output: + def __getitem__(self, key): + return getattr(self, key) + def __setitem__(self, key, item): + setattr(self, key, item) def clip_preprocess(image, size=224): mean = torch.tensor([ 0.48145466,0.4578275,0.40821073], device=image.device, dtype=image.dtype) @@ -22,17 +29,16 @@ def clip_preprocess(image, size=224): class ClipVisionModel(): def __init__(self, json_config): - config = CLIPVisionConfig.from_json_file(json_config) + with open(json_config) as f: + config = json.load(f) + self.load_device = comfy.model_management.text_encoder_device() offload_device = comfy.model_management.text_encoder_offload_device() self.dtype = torch.float32 if comfy.model_management.should_use_fp16(self.load_device, prioritize_performance=False): self.dtype = torch.float16 - with comfy.ops.use_comfy_ops(offload_device, self.dtype): - with modeling_utils.no_init_weights(): - self.model = CLIPVisionModelWithProjection(config) - self.model.to(self.dtype) + self.model = comfy.clip_model.CLIPVisionModelProjection(config, self.dtype, offload_device, comfy.ops.disable_weight_init) self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) def load_sd(self, sd): @@ -48,17 +54,12 @@ class ClipVisionModel(): precision_scope = lambda a, b: contextlib.nullcontext(a) with precision_scope(comfy.model_management.get_autocast_device(self.load_device), torch.float32): - outputs = self.model(pixel_values=pixel_values, output_hidden_states=True) - - for k in outputs: - t = outputs[k] - if t is not None: - if k == 'hidden_states': - outputs["penultimate_hidden_states"] = t[-2].cpu() - outputs["hidden_states"] = None - else: - outputs[k] = t.cpu() + out = self.model(pixel_values=pixel_values, intermediate_output=-2) + outputs = Output() + outputs["last_hidden_state"] = out[0].to(comfy.model_management.intermediate_device()) + outputs["image_embeds"] = out[2].to(comfy.model_management.intermediate_device()) + outputs["penultimate_hidden_states"] = out[1].to(comfy.model_management.intermediate_device()) return outputs def convert_to_transformers(sd, prefix): diff --git a/comfy/controlnet.py b/comfy/controlnet.py index 6d37aa74f..110b5c7c2 100644 --- a/comfy/controlnet.py +++ b/comfy/controlnet.py @@ -36,13 +36,13 @@ class ControlBase: self.cond_hint = None self.strength = 1.0 self.timestep_percent_range = (0.0, 1.0) + self.global_average_pooling = False self.timestep_range = None if device is None: device = comfy.model_management.get_torch_device() self.device = device self.previous_controlnet = None - self.global_average_pooling = False def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0)): self.cond_hint_original = cond_hint @@ -77,6 +77,7 @@ class ControlBase: c.cond_hint_original = self.cond_hint_original c.strength = self.strength c.timestep_percent_range = self.timestep_percent_range + c.global_average_pooling = self.global_average_pooling def inference_memory_requirements(self, dtype): if self.previous_controlnet is not None: @@ -129,12 +130,14 @@ class ControlBase: return out class ControlNet(ControlBase): - def __init__(self, control_model, global_average_pooling=False, device=None): + def __init__(self, control_model, global_average_pooling=False, device=None, load_device=None, manual_cast_dtype=None): super().__init__(device) self.control_model = control_model - self.control_model_wrapped = comfy.model_patcher.ModelPatcher(self.control_model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device()) + 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()) self.global_average_pooling = global_average_pooling self.model_sampling_current = None + self.manual_cast_dtype = manual_cast_dtype def get_control(self, x_noisy, t, cond, batched_number): control_prev = None @@ -149,11 +152,8 @@ class ControlNet(ControlBase): return None dtype = self.control_model.dtype - if comfy.model_management.supports_dtype(self.device, dtype): - precision_scope = lambda a: contextlib.nullcontext(a) - else: - precision_scope = torch.autocast - dtype = torch.float32 + if self.manual_cast_dtype is not None: + 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]: @@ -171,12 +171,11 @@ class ControlNet(ControlBase): timestep = self.model_sampling_current.timestep(t) x_noisy = self.model_sampling_current.calculate_input(t, x_noisy) - with precision_scope(comfy.model_management.get_autocast_device(self.device)): - control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.float(), context=context.to(dtype), y=y) + control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.float(), context=context.to(dtype), y=y) 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) + c = ControlNet(self.control_model, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype) self.copy_to(c) return c @@ -207,10 +206,11 @@ class ControlLoraOps: self.bias = None def forward(self, input): + weight, bias = comfy.ops.cast_bias_weight(self, input) if self.up is not None: - return torch.nn.functional.linear(input, self.weight.to(input.dtype).to(input.device) + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), self.bias) + return torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias) else: - return torch.nn.functional.linear(input, self.weight.to(input.device), self.bias) + return torch.nn.functional.linear(input, weight, bias) class Conv2d(torch.nn.Module): def __init__( @@ -246,25 +246,11 @@ class ControlLoraOps: def forward(self, input): + weight, bias = comfy.ops.cast_bias_weight(self, input) if self.up is not None: - return torch.nn.functional.conv2d(input, self.weight.to(input.dtype).to(input.device) + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), self.bias, self.stride, self.padding, self.dilation, self.groups) + return torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups) else: - return torch.nn.functional.conv2d(input, self.weight.to(input.device), self.bias, self.stride, self.padding, self.dilation, self.groups) - - def conv_nd(self, dims, *args, **kwargs): - if dims == 2: - return self.Conv2d(*args, **kwargs) - else: - raise ValueError(f"unsupported dimensions: {dims}") - - class Conv3d(comfy.ops.Conv3d): - pass - - class GroupNorm(comfy.ops.GroupNorm): - pass - - class LayerNorm(comfy.ops.LayerNorm): - pass + return torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups) class ControlLora(ControlNet): @@ -278,10 +264,19 @@ class ControlLora(ControlNet): controlnet_config = model.model_config.unet_config.copy() controlnet_config.pop("out_channels") controlnet_config["hint_channels"] = self.control_weights["input_hint_block.0.weight"].shape[1] - controlnet_config["operations"] = ControlLoraOps() - self.control_model = comfy.cldm.cldm.ControlNet(**controlnet_config) + self.manual_cast_dtype = model.manual_cast_dtype dtype = model.get_dtype() - self.control_model.to(dtype) + if self.manual_cast_dtype is None: + class control_lora_ops(ControlLoraOps, comfy.ops.disable_weight_init): + pass + else: + class control_lora_ops(ControlLoraOps, comfy.ops.manual_cast): + pass + dtype = self.manual_cast_dtype + + controlnet_config["operations"] = control_lora_ops + controlnet_config["dtype"] = dtype + self.control_model = comfy.cldm.cldm.ControlNet(**controlnet_config) self.control_model.to(comfy.model_management.get_torch_device()) diffusion_model = model.diffusion_model sd = diffusion_model.state_dict() @@ -385,6 +380,10 @@ def load_controlnet(ckpt_path, model=None): 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 + load_device = comfy.model_management.get_torch_device() + 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.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) @@ -413,14 +412,12 @@ def load_controlnet(ckpt_path, model=None): missing, unexpected = control_model.load_state_dict(controlnet_data, strict=False) print(missing, unexpected) - control_model = control_model.to(unet_dtype) - global_average_pooling = False filename = os.path.splitext(ckpt_path)[0] if filename.endswith("_shuffle") or filename.endswith("_shuffle_fp16"): #TODO: smarter way of enabling global_average_pooling global_average_pooling = True - control = ControlNet(control_model, global_average_pooling=global_average_pooling) + control = ControlNet(control_model, global_average_pooling=global_average_pooling, load_device=load_device, manual_cast_dtype=manual_cast_dtype) return control class T2IAdapter(ControlBase): diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index c2b85a691..8d86aa53d 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -19,6 +19,7 @@ if model_management.xformers_enabled(): from comfy.cli_args import args import comfy.ops +ops = comfy.ops.disable_weight_init # CrossAttn precision handling if args.dont_upcast_attention: @@ -55,7 +56,7 @@ def init_(tensor): # feedforward class GEGLU(nn.Module): - def __init__(self, dim_in, dim_out, dtype=None, device=None, operations=comfy.ops): + def __init__(self, dim_in, dim_out, dtype=None, device=None, operations=ops): super().__init__() self.proj = operations.Linear(dim_in, dim_out * 2, dtype=dtype, device=device) @@ -65,7 +66,7 @@ class GEGLU(nn.Module): class FeedForward(nn.Module): - def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=comfy.ops): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=ops): super().__init__() inner_dim = int(dim * mult) dim_out = default(dim_out, dim) @@ -112,10 +113,13 @@ def attention_basic(q, k, v, heads, mask=None): del q, k if exists(mask): - mask = rearrange(mask, 'b ... -> b (...)') - max_neg_value = -torch.finfo(sim.dtype).max - mask = repeat(mask, 'b j -> (b h) () j', h=h) - sim.masked_fill_(~mask, max_neg_value) + if mask.dtype == torch.bool: + mask = rearrange(mask, 'b ... -> b (...)') #TODO: check if this bool part matches pytorch attention + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b j -> (b h) () j', h=h) + sim.masked_fill_(~mask, max_neg_value) + else: + sim += mask # attention, what we cannot get enough of sim = sim.softmax(dim=-1) @@ -340,8 +344,20 @@ else: if model_management.pytorch_attention_enabled(): optimized_attention_masked = attention_pytorch +def optimized_attention_for_device(device, mask=False): + if device == torch.device("cpu"): #TODO + if model_management.pytorch_attention_enabled(): + return attention_pytorch + else: + return attention_basic + if mask: + return optimized_attention_masked + + return optimized_attention + + class CrossAttention(nn.Module): - def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., dtype=None, device=None, operations=comfy.ops): + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., dtype=None, device=None, operations=ops): super().__init__() inner_dim = dim_head * heads context_dim = default(context_dim, query_dim) @@ -374,7 +390,7 @@ class CrossAttention(nn.Module): 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=comfy.ops): + disable_self_attn=False, disable_temporal_crossattention=False, switch_temporal_ca_to_sa=False, dtype=None, device=None, operations=ops): super().__init__() self.ff_in = ff_in or inner_dim is not None @@ -384,7 +400,7 @@ class BasicTransformerBlock(nn.Module): self.is_res = inner_dim == dim if self.ff_in: - self.norm_in = nn.LayerNorm(dim, dtype=dtype, device=device) + self.norm_in = operations.LayerNorm(dim, dtype=dtype, device=device) self.ff_in = FeedForward(dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations) self.disable_self_attn = disable_self_attn @@ -543,7 +559,7 @@ 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=comfy.ops): + use_checkpoint=True, dtype=None, device=None, operations=ops): super().__init__() if exists(context_dim) and not isinstance(context_dim, list): context_dim = [context_dim] * depth @@ -617,7 +633,7 @@ class SpatialVideoTransformer(SpatialTransformer): disable_self_attn=False, disable_temporal_crossattention=False, max_time_embed_period: int = 10000, - dtype=None, device=None, operations=comfy.ops + dtype=None, device=None, operations=ops ): super().__init__( in_channels, diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.py index f23417fd2..fce29cb85 100644 --- a/comfy/ldm/modules/diffusionmodules/model.py +++ b/comfy/ldm/modules/diffusionmodules/model.py @@ -8,6 +8,7 @@ from typing import Optional, Any from comfy import model_management import comfy.ops +ops = comfy.ops.disable_weight_init if model_management.xformers_enabled_vae(): import xformers @@ -48,7 +49,7 @@ class Upsample(nn.Module): super().__init__() self.with_conv = with_conv if self.with_conv: - self.conv = comfy.ops.Conv2d(in_channels, + self.conv = ops.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, @@ -78,7 +79,7 @@ class Downsample(nn.Module): self.with_conv = with_conv if self.with_conv: # no asymmetric padding in torch conv, must do it ourselves - self.conv = comfy.ops.Conv2d(in_channels, + self.conv = ops.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, @@ -105,30 +106,30 @@ class ResnetBlock(nn.Module): self.swish = torch.nn.SiLU(inplace=True) self.norm1 = Normalize(in_channels) - self.conv1 = comfy.ops.Conv2d(in_channels, + self.conv1 = ops.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) if temb_channels > 0: - self.temb_proj = comfy.ops.Linear(temb_channels, + self.temb_proj = ops.Linear(temb_channels, out_channels) self.norm2 = Normalize(out_channels) self.dropout = torch.nn.Dropout(dropout, inplace=True) - self.conv2 = comfy.ops.Conv2d(out_channels, + self.conv2 = ops.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 = comfy.ops.Conv2d(in_channels, + self.conv_shortcut = ops.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) else: - self.nin_shortcut = comfy.ops.Conv2d(in_channels, + self.nin_shortcut = ops.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, @@ -245,22 +246,22 @@ class AttnBlock(nn.Module): self.in_channels = in_channels self.norm = Normalize(in_channels) - self.q = comfy.ops.Conv2d(in_channels, + self.q = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) - self.k = comfy.ops.Conv2d(in_channels, + self.k = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) - self.v = comfy.ops.Conv2d(in_channels, + self.v = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) - self.proj_out = comfy.ops.Conv2d(in_channels, + self.proj_out = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, @@ -312,14 +313,14 @@ class Model(nn.Module): # timestep embedding self.temb = nn.Module() self.temb.dense = nn.ModuleList([ - comfy.ops.Linear(self.ch, + ops.Linear(self.ch, self.temb_ch), - comfy.ops.Linear(self.temb_ch, + ops.Linear(self.temb_ch, self.temb_ch), ]) # downsampling - self.conv_in = comfy.ops.Conv2d(in_channels, + self.conv_in = ops.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, @@ -388,7 +389,7 @@ class Model(nn.Module): # end self.norm_out = Normalize(block_in) - self.conv_out = comfy.ops.Conv2d(block_in, + self.conv_out = ops.Conv2d(block_in, out_ch, kernel_size=3, stride=1, @@ -461,7 +462,7 @@ class Encoder(nn.Module): self.in_channels = in_channels # downsampling - self.conv_in = comfy.ops.Conv2d(in_channels, + self.conv_in = ops.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, @@ -506,7 +507,7 @@ class Encoder(nn.Module): # end self.norm_out = Normalize(block_in) - self.conv_out = comfy.ops.Conv2d(block_in, + self.conv_out = ops.Conv2d(block_in, 2*z_channels if double_z else z_channels, kernel_size=3, stride=1, @@ -541,7 +542,7 @@ class Decoder(nn.Module): def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, - conv_out_op=comfy.ops.Conv2d, + conv_out_op=ops.Conv2d, resnet_op=ResnetBlock, attn_op=AttnBlock, **ignorekwargs): @@ -565,7 +566,7 @@ class Decoder(nn.Module): self.z_shape, np.prod(self.z_shape))) # z to block_in - self.conv_in = comfy.ops.Conv2d(z_channels, + self.conv_in = ops.Conv2d(z_channels, block_in, kernel_size=3, stride=1, diff --git a/comfy/ldm/modules/diffusionmodules/openaimodel.py b/comfy/ldm/modules/diffusionmodules/openaimodel.py index 12efd833c..057dd16b2 100644 --- a/comfy/ldm/modules/diffusionmodules/openaimodel.py +++ b/comfy/ldm/modules/diffusionmodules/openaimodel.py @@ -12,13 +12,13 @@ from .util import ( checkpoint, avg_pool_nd, zero_module, - normalization, timestep_embedding, AlphaBlender, ) from ..attention import SpatialTransformer, SpatialVideoTransformer, default from comfy.ldm.util import exists import comfy.ops +ops = comfy.ops.disable_weight_init class TimestepBlock(nn.Module): """ @@ -70,7 +70,7 @@ class Upsample(nn.Module): upsampling occurs in the inner-two dimensions. """ - def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=comfy.ops): + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=ops): super().__init__() self.channels = channels self.out_channels = out_channels or channels @@ -106,7 +106,7 @@ class Downsample(nn.Module): downsampling occurs in the inner-two dimensions. """ - def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=comfy.ops): + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=ops): super().__init__() self.channels = channels self.out_channels = out_channels or channels @@ -159,7 +159,7 @@ class ResBlock(TimestepBlock): skip_t_emb=False, dtype=None, device=None, - operations=comfy.ops + operations=ops ): super().__init__() self.channels = channels @@ -284,7 +284,7 @@ class VideoResBlock(ResBlock): down: bool = False, dtype=None, device=None, - operations=comfy.ops + operations=ops ): super().__init__( channels, @@ -434,7 +434,7 @@ class UNetModel(nn.Module): disable_temporal_crossattention=False, max_ddpm_temb_period=10000, device=None, - operations=comfy.ops, + operations=ops, ): super().__init__() assert use_spatial_transformer == True, "use_spatial_transformer has to be true" @@ -581,7 +581,7 @@ class UNetModel(nn.Module): up=False, dtype=None, device=None, - operations=comfy.ops + operations=ops ): if self.use_temporal_resblocks: return VideoResBlock( diff --git a/comfy/ldm/modules/diffusionmodules/util.py b/comfy/ldm/modules/diffusionmodules/util.py index 704bbe574..68175b62a 100644 --- a/comfy/ldm/modules/diffusionmodules/util.py +++ b/comfy/ldm/modules/diffusionmodules/util.py @@ -16,7 +16,6 @@ import numpy as np from einops import repeat, rearrange from comfy.ldm.util import instantiate_from_config -import comfy.ops class AlphaBlender(nn.Module): strategies = ["learned", "fixed", "learned_with_images"] @@ -273,46 +272,6 @@ def mean_flat(tensor): return tensor.mean(dim=list(range(1, len(tensor.shape)))) -def normalization(channels, dtype=None): - """ - Make a standard normalization layer. - :param channels: number of input channels. - :return: an nn.Module for normalization. - """ - return GroupNorm32(32, channels, dtype=dtype) - - -# PyTorch 1.7 has SiLU, but we support PyTorch 1.5. -class SiLU(nn.Module): - def forward(self, x): - return x * torch.sigmoid(x) - - -class GroupNorm32(nn.GroupNorm): - def forward(self, x): - return super().forward(x.float()).type(x.dtype) - - -def conv_nd(dims, *args, **kwargs): - """ - Create a 1D, 2D, or 3D convolution module. - """ - if dims == 1: - return nn.Conv1d(*args, **kwargs) - elif dims == 2: - return comfy.ops.Conv2d(*args, **kwargs) - elif dims == 3: - return nn.Conv3d(*args, **kwargs) - raise ValueError(f"unsupported dimensions: {dims}") - - -def linear(*args, **kwargs): - """ - Create a linear module. - """ - return comfy.ops.Linear(*args, **kwargs) - - def avg_pool_nd(dims, *args, **kwargs): """ Create a 1D, 2D, or 3D average pooling module. diff --git a/comfy/ldm/modules/temporal_ae.py b/comfy/ldm/modules/temporal_ae.py index 11ae049f3..7ea68dc9e 100644 --- a/comfy/ldm/modules/temporal_ae.py +++ b/comfy/ldm/modules/temporal_ae.py @@ -5,6 +5,7 @@ import torch from einops import rearrange, repeat import comfy.ops +ops = comfy.ops.disable_weight_init from .diffusionmodules.model import ( AttnBlock, @@ -130,9 +131,9 @@ class AttnVideoBlock(AttnBlock): time_embed_dim = self.in_channels * 4 self.video_time_embed = torch.nn.Sequential( - comfy.ops.Linear(self.in_channels, time_embed_dim), + ops.Linear(self.in_channels, time_embed_dim), torch.nn.SiLU(), - comfy.ops.Linear(time_embed_dim, self.in_channels), + ops.Linear(time_embed_dim, self.in_channels), ) self.merge_strategy = merge_strategy diff --git a/comfy/lora.py b/comfy/lora.py index 29c59d893..5e4009b47 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -43,7 +43,7 @@ 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[A_name], lora[B_name], alpha, mid) + patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid)) loaded_keys.add(A_name) loaded_keys.add(B_name) @@ -64,7 +64,7 @@ def load_lora(lora, to_load): loaded_keys.add(hada_t1_name) loaded_keys.add(hada_t2_name) - patch_dict[to_load[x]] = (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)) loaded_keys.add(hada_w1_a_name) loaded_keys.add(hada_w1_b_name) loaded_keys.add(hada_w2_a_name) @@ -116,8 +116,19 @@ 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_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)) + #glora + a1_name = "{}.a1.weight".format(x) + a2_name = "{}.a2.weight".format(x) + 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)) + loaded_keys.add(a1_name) + loaded_keys.add(a2_name) + loaded_keys.add(b1_name) + loaded_keys.add(b2_name) w_norm_name = "{}.w_norm".format(x) b_norm_name = "{}.b_norm".format(x) @@ -126,21 +137,21 @@ def load_lora(lora, to_load): if w_norm is not None: loaded_keys.add(w_norm_name) - patch_dict[to_load[x]] = (w_norm,) + patch_dict[to_load[x]] = ("diff", (w_norm,)) if b_norm is not None: loaded_keys.add(b_norm_name) - patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = (b_norm,) + patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = ("diff", (b_norm,)) diff_name = "{}.diff".format(x) diff_weight = lora.get(diff_name, None) if diff_weight is not None: - patch_dict[to_load[x]] = (diff_weight,) + patch_dict[to_load[x]] = ("diff", (diff_weight,)) loaded_keys.add(diff_name) diff_bias_name = "{}.diff_b".format(x) diff_bias = lora.get(diff_bias_name, None) if diff_bias is not None: - patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = (diff_bias,) + patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = ("diff", (diff_bias,)) loaded_keys.add(diff_bias_name) for x in lora.keys(): diff --git a/comfy/model_base.py b/comfy/model_base.py index 5bfcc391d..a7582b330 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -4,6 +4,7 @@ from comfy.ldm.modules.encoders.noise_aug_modules import CLIPEmbeddingNoiseAugme from comfy.ldm.modules.diffusionmodules.openaimodel import Timestep import comfy.model_management import comfy.conds +import comfy.ops from enum import Enum import contextlib from . import utils @@ -41,9 +42,14 @@ class BaseModel(torch.nn.Module): unet_config = model_config.unet_config self.latent_format = model_config.latent_format self.model_config = model_config + self.manual_cast_dtype = model_config.manual_cast_dtype if not unet_config.get("disable_unet_model_creation", False): - self.diffusion_model = UNetModel(**unet_config, device=device) + if self.manual_cast_dtype is not None: + operations = comfy.ops.manual_cast + else: + operations = comfy.ops.disable_weight_init + self.diffusion_model = UNetModel(**unet_config, device=device, operations=operations) self.model_type = model_type self.model_sampling = model_sampling(model_config, model_type) @@ -63,11 +69,8 @@ class BaseModel(torch.nn.Module): context = c_crossattn dtype = self.get_dtype() - if comfy.model_management.supports_dtype(xc.device, dtype): - precision_scope = lambda a: contextlib.nullcontext(a) - else: - precision_scope = torch.autocast - dtype = torch.float32 + if self.manual_cast_dtype is not None: + dtype = self.manual_cast_dtype xc = xc.to(dtype) t = self.model_sampling.timestep(t).float() @@ -79,9 +82,7 @@ class BaseModel(torch.nn.Module): extra = extra.to(dtype) extra_conds[o] = extra - with precision_scope(comfy.model_management.get_autocast_device(xc.device)): - model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float() - + model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float() return self.model_sampling.calculate_denoised(sigma, model_output, x) def get_dtype(self): @@ -176,9 +177,12 @@ class BaseModel(torch.nn.Module): 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 needs to be tweaked area = input_shape[0] * input_shape[2] * input_shape[3] - return (area * comfy.model_management.dtype_size(self.get_dtype()) / 50) * (1024 * 1024) + return (area * comfy.model_management.dtype_size(dtype) / 50) * (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] * input_shape[2] * input_shape[3] diff --git a/comfy/model_detection.py b/comfy/model_detection.py index c682c3e1a..e3af422a3 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -289,7 +289,13 @@ 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 = {'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, 1, 1, 2, 2], 'transformer_depth_output': [0, 0, 0, 1, 1, 1, 2, 2, 2], + '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] for unet_config in supported_models: matches = True diff --git a/comfy/model_management.py b/comfy/model_management.py index 94d596969..b6a9471bf 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -466,6 +466,8 @@ def unet_inital_load_device(parameters, dtype): def unet_dtype(device=None, model_params=0): if args.bf16_unet: return torch.bfloat16 + if args.fp16_unet: + return torch.float16 if args.fp8_e4m3fn_unet: return torch.float8_e4m3fn if args.fp8_e5m2_unet: @@ -474,6 +476,20 @@ def unet_dtype(device=None, model_params=0): return torch.float16 return torch.float32 +# None means no manual cast +def unet_manual_cast(weight_dtype, inference_device): + if weight_dtype == torch.float32: + return None + + fp16_supported = comfy.model_management.should_use_fp16(inference_device, prioritize_performance=False) + if fp16_supported and weight_dtype == torch.float16: + return None + + if fp16_supported: + return torch.float16 + else: + return torch.float32 + def text_encoder_offload_device(): if args.gpu_only: return get_torch_device() @@ -503,11 +519,20 @@ def text_encoder_dtype(device=None): elif args.fp32_text_enc: return torch.float32 + if is_device_cpu(device): + return torch.float16 + if should_use_fp16(device, prioritize_performance=False): return torch.float16 else: return torch.float32 +def intermediate_device(): + if args.gpu_only: + return get_torch_device() + else: + return torch.device("cpu") + def vae_device(): return get_torch_device() @@ -529,7 +554,7 @@ def get_autocast_device(dev): def supports_dtype(device, dtype): #TODO if dtype == torch.float32: return True - if torch.device("cpu") == device: + if is_device_cpu(device): return False if dtype == torch.float16: return True @@ -547,15 +572,19 @@ def cast_to_device(tensor, device, dtype, copy=False): elif is_intel_xpu(): device_supports_cast = True + non_blocking = True + if is_device_mps(device): + non_blocking = False #pytorch bug? mps doesn't support non blocking + if device_supports_cast: if copy: if tensor.device == device: - return tensor.to(dtype, copy=copy) - return tensor.to(device, copy=copy).to(dtype) + return tensor.to(dtype, copy=copy, non_blocking=non_blocking) + return tensor.to(device, copy=copy, non_blocking=non_blocking).to(dtype, non_blocking=non_blocking) else: - return tensor.to(device).to(dtype) + return tensor.to(device, non_blocking=non_blocking).to(dtype, non_blocking=non_blocking) else: - return tensor.to(dtype).to(device, copy=copy) + return tensor.to(device, dtype, copy=copy, non_blocking=non_blocking) def xformers_enabled(): global directml_enabled diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index a3cffc3be..e0acdc961 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -61,6 +61,9 @@ class ModelPatcher: else: self.model_options["sampler_cfg_function"] = sampler_cfg_function + def set_model_sampler_post_cfg_function(self, post_cfg_function): + self.model_options["sampler_post_cfg_function"] = self.model_options.get("sampler_post_cfg_function", []) + [post_cfg_function] + def set_model_unet_function_wrapper(self, unet_wrapper_function): self.model_options["model_function_wrapper"] = unet_wrapper_function @@ -70,13 +73,17 @@ class ModelPatcher: to["patches"] = {} to["patches"][name] = to["patches"].get(name, []) + [patch] - def set_model_patch_replace(self, patch, name, block_name, number): + 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] = {} - to["patches_replace"][name][(block_name, number)] = patch + if transformer_index is not None: + block = (block_name, number, transformer_index) + else: + block = (block_name, number) + to["patches_replace"][name][block] = patch def set_model_attn1_patch(self, patch): self.set_model_patch(patch, "attn1_patch") @@ -84,11 +91,11 @@ class ModelPatcher: def set_model_attn2_patch(self, patch): self.set_model_patch(patch, "attn2_patch") - def set_model_attn1_replace(self, patch, block_name, number): - self.set_model_patch_replace(patch, "attn1", block_name, number) + def set_model_attn1_replace(self, patch, block_name, number, transformer_index=None): + self.set_model_patch_replace(patch, "attn1", block_name, number, transformer_index) - def set_model_attn2_replace(self, patch, block_name, number): - self.set_model_patch_replace(patch, "attn2", block_name, number) + def set_model_attn2_replace(self, patch, block_name, number, transformer_index=None): + self.set_model_patch_replace(patch, "attn2", block_name, number, transformer_index) def set_model_attn1_output_patch(self, patch): self.set_model_patch(patch, "attn1_output_patch") @@ -217,13 +224,19 @@ class ModelPatcher: v = (self.calculate_weight(v[1:], v[0].clone(), key), ) if len(v) == 1: + patch_type = "diff" + elif len(v) == 2: + patch_type = v[0] + v = v[1] + + if patch_type == "diff": w1 = v[0] if alpha != 0.0: if w1.shape != weight.shape: print("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) - elif len(v) == 4: #lora/locon + 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) if v[2] is not None: @@ -237,7 +250,7 @@ class ModelPatcher: weight += (alpha * torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1))).reshape(weight.shape).type(weight.dtype) except Exception as e: print("ERROR", key, e) - elif len(v) == 8: #lokr + elif patch_type == "lokr": w1 = v[0] w2 = v[1] w1_a = v[3] @@ -276,7 +289,7 @@ class ModelPatcher: weight += alpha * torch.kron(w1, w2).reshape(weight.shape).type(weight.dtype) except Exception as e: print("ERROR", key, e) - else: #loha + elif patch_type == "loha": w1a = v[0] w1b = v[1] if v[2] is not None: @@ -305,6 +318,18 @@ class ModelPatcher: weight += (alpha * m1 * m2).reshape(weight.shape).type(weight.dtype) except Exception as e: print("ERROR", key, e) + elif patch_type == "glora": + if v[4] is not None: + alpha *= v[4] / v[0].shape[0] + + 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) + else: + print("patch type not recognized", patch_type, key) return weight diff --git a/comfy/model_sampling.py b/comfy/model_sampling.py index 69c8b1f01..cc8745c10 100644 --- a/comfy/model_sampling.py +++ b/comfy/model_sampling.py @@ -22,10 +22,17 @@ class V_PREDICTION(EPS): class ModelSamplingDiscrete(torch.nn.Module): def __init__(self, model_config=None): super().__init__() - beta_schedule = "linear" + if model_config is not None: - beta_schedule = model_config.sampling_settings.get("beta_schedule", beta_schedule) - self._register_schedule(given_betas=None, beta_schedule=beta_schedule, timesteps=1000, linear_start=0.00085, linear_end=0.012, cosine_s=8e-3) + sampling_settings = model_config.sampling_settings + else: + sampling_settings = {} + + beta_schedule = sampling_settings.get("beta_schedule", "linear") + linear_start = sampling_settings.get("linear_start", 0.00085) + linear_end = sampling_settings.get("linear_end", 0.012) + + self._register_schedule(given_betas=None, beta_schedule=beta_schedule, timesteps=1000, linear_start=linear_start, linear_end=linear_end, cosine_s=8e-3) self.sigma_data = 1.0 def _register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000, diff --git a/comfy/ops.py b/comfy/ops.py index deb849d63..08c633847 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -1,48 +1,65 @@ import torch from contextlib import contextmanager -class Linear(torch.nn.Linear): - def reset_parameters(self): - return None +class disable_weight_init: + class Linear(torch.nn.Linear): + def reset_parameters(self): + return None -class Conv2d(torch.nn.Conv2d): - def reset_parameters(self): - return None + class Conv2d(torch.nn.Conv2d): + def reset_parameters(self): + return None -class Conv3d(torch.nn.Conv3d): - def reset_parameters(self): - return None + class Conv3d(torch.nn.Conv3d): + def reset_parameters(self): + return None -class GroupNorm(torch.nn.GroupNorm): - def reset_parameters(self): - return None + class GroupNorm(torch.nn.GroupNorm): + def reset_parameters(self): + return None -class LayerNorm(torch.nn.LayerNorm): - def reset_parameters(self): - return None + class LayerNorm(torch.nn.LayerNorm): + def reset_parameters(self): + return None -def conv_nd(dims, *args, **kwargs): - if dims == 2: - return Conv2d(*args, **kwargs) - elif dims == 3: - return Conv3d(*args, **kwargs) - else: - raise ValueError(f"unsupported dimensions: {dims}") + @classmethod + def conv_nd(s, dims, *args, **kwargs): + if dims == 2: + return s.Conv2d(*args, **kwargs) + elif dims == 3: + return s.Conv3d(*args, **kwargs) + else: + raise ValueError(f"unsupported dimensions: {dims}") -@contextmanager -def use_comfy_ops(device=None, dtype=None): # Kind of an ugly hack but I can't think of a better way - old_torch_nn_linear = torch.nn.Linear - force_device = device - force_dtype = dtype - def linear_with_dtype(in_features: int, out_features: int, bias: bool = True, device=None, dtype=None): - if force_device is not None: - device = force_device - if force_dtype is not None: - dtype = force_dtype - return Linear(in_features, out_features, bias=bias, device=device, dtype=dtype) +def cast_bias_weight(s, input): + bias = None + if s.bias is not None: + bias = s.bias.to(device=input.device, dtype=input.dtype) + weight = s.weight.to(device=input.device, dtype=input.dtype) + return weight, bias - torch.nn.Linear = linear_with_dtype - try: - yield - finally: - torch.nn.Linear = old_torch_nn_linear +class manual_cast(disable_weight_init): + class Linear(disable_weight_init.Linear): + def forward(self, input): + weight, bias = cast_bias_weight(self, input) + return torch.nn.functional.linear(input, weight, bias) + + class Conv2d(disable_weight_init.Conv2d): + def forward(self, input): + weight, bias = cast_bias_weight(self, input) + return self._conv_forward(input, weight, bias) + + class Conv3d(disable_weight_init.Conv3d): + def forward(self, input): + weight, bias = cast_bias_weight(self, input) + return self._conv_forward(input, weight, bias) + + class GroupNorm(disable_weight_init.GroupNorm): + def forward(self, input): + weight, bias = cast_bias_weight(self, input) + return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps) + + class LayerNorm(disable_weight_init.LayerNorm): + def forward(self, input): + weight, bias = cast_bias_weight(self, input) + return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps) diff --git a/comfy/sample.py b/comfy/sample.py index bcbed3343..eadd6dcc8 100644 --- a/comfy/sample.py +++ b/comfy/sample.py @@ -98,7 +98,7 @@ def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative 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 = samples.cpu() + 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"))) @@ -111,7 +111,7 @@ def sample_custom(model, noise, cfg, sampler, sigmas, positive, negative, latent 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 = samples.cpu() + 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/samplers.py b/comfy/samplers.py index 1d012a514..39bc3774a 100644 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -2,259 +2,266 @@ 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 +def get_area_and_mult(conds, x_in, timestep_in): + area = (x_in.shape[2], x_in.shape[3], 0, 0) + strength = 1.0 + + if 'timestep_start' in conds: + timestep_start = conds['timestep_start'] + if timestep_in[0] > timestep_start: + return None + if 'timestep_end' in conds: + timestep_end = conds['timestep_end'] + if timestep_in[0] < timestep_end: + return None + if 'area' in conds: + area = conds['area'] + if 'strength' in conds: + strength = conds['strength'] + + input_x = x_in[:,:,area[2]:area[0] + area[2],area[3]:area[1] + area[3]] + if 'mask' in conds: + # Scale the mask to the size of the input + # The mask should have been resized as we began the sampling process + mask_strength = 1.0 + if "mask_strength" in conds: + mask_strength = conds["mask_strength"] + 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.unsqueeze(1).repeat(input_x.shape[0] // mask.shape[0], input_x.shape[1], 1, 1) + else: + mask = torch.ones_like(input_x) + mult = mask * strength + + if 'mask' not in conds: + rr = 8 + if area[2] != 0: + for t in range(rr): + mult[:,:,t:1+t,:] *= ((1.0/rr) * (t + 1)) + if (area[0] + area[2]) < x_in.shape[2]: + for t in range(rr): + mult[:,:,area[0] - 1 - t:area[0] - t,:] *= ((1.0/rr) * (t + 1)) + if area[3] != 0: + for t in range(rr): + mult[:,:,:,t:1+t] *= ((1.0/rr) * (t + 1)) + if (area[1] + area[3]) < x_in.shape[3]: + for t in range(rr): + mult[:,:,:,area[1] - 1 - t:area[1] - t] *= ((1.0/rr) * (t + 1)) + + conditioning = {} + model_conds = conds["model_conds"] + for c in model_conds: + conditioning[c] = model_conds[c].process_cond(batch_size=x_in.shape[0], device=x_in.device, area=area) + + control = conds.get('control', None) + + patches = None + if 'gligen' in conds: + gligen = conds['gligen'] + patches = {} + gligen_type = gligen[0] + gligen_model = gligen[1] + if gligen_type == "position": + gligen_patch = gligen_model.model.set_position(input_x.shape, gligen[2], input_x.device) + else: + gligen_patch = gligen_model.model.set_empty(input_x.shape, input_x.device) + + patches['middle_patch'] = [gligen_patch] + + cond_obj = collections.namedtuple('cond_obj', ['input_x', 'mult', 'conditioning', 'area', 'control', 'patches']) + return cond_obj(input_x, mult, conditioning, area, control, patches) + +def cond_equal_size(c1, c2): + if c1 is c2: + return True + if c1.keys() != c2.keys(): + return False + for k in c1: + if not c1[k].can_concat(c2[k]): + return False + return True + +def can_concat_cond(c1, c2): + if c1.input_x.shape != c2.input_x.shape: + return False + + def objects_concatable(obj1, obj2): + if (obj1 is None) != (obj2 is None): + return False + if obj1 is not None: + if obj1 is not obj2: + return False + return True + + if not objects_concatable(c1.control, c2.control): + return False + + if not objects_concatable(c1.patches, c2.patches): + return False + + return cond_equal_size(c1.conditioning, c2.conditioning) + +def cond_cat(c_list): + c_crossattn = [] + c_concat = [] + c_adm = [] + crossattn_max_len = 0 + + temp = {} + for x in c_list: + for k in x: + cur = temp.get(k, []) + cur.append(x[k]) + temp[k] = cur + + out = {} + for k in temp: + conds = temp[k] + out[k] = conds[0].concat(conds[1:]) + + 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 + + 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 + + to_run += [(p, UNCOND)] + + while len(to_run) > 0: + first = to_run[0] + first_shape = first[0][0].shape + to_batch_temp = [] + for x in range(len(to_run)): + if can_concat_cond(to_run[x][0], first[0]): + to_batch_temp += [x] + + to_batch_temp.reverse() + to_batch = to_batch_temp[:1] + + free_memory = model_management.get_free_memory(x_in.device) + 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: + to_batch = batch_amount + break + + input_x = [] + mult = [] + c = [] + cond_or_uncond = [] + area = [] + control = None + patches = None + for x in to_batch: + o = to_run.pop(x) + p = o[0] + input_x.append(p.input_x) + mult.append(p.mult) + c.append(p.conditioning) + area.append(p.area) + cond_or_uncond.append(o[1]) + control = p.control + patches = p.patches + + batch_chunks = len(cond_or_uncond) + input_x = torch.cat(input_x) + c = cond_cat(c) + timestep_ = torch.cat([timestep] * batch_chunks) + + if control is not None: + c['control'] = control.get_control(input_x, timestep_, c, len(cond_or_uncond)) + + transformer_options = {} + if 'transformer_options' in model_options: + transformer_options = model_options['transformer_options'].copy() + + if patches is not None: + if "patches" in transformer_options: + cur_patches = transformer_options["patches"].copy() + for p in patches: + if p in cur_patches: + cur_patches[p] = cur_patches[p] + patches[p] + else: + cur_patches[p] = patches[p] + else: + transformer_options["patches"] = patches + + transformer_options["cond_or_uncond"] = cond_or_uncond[:] + transformer_options["sigmas"] = timestep + + c['transformer_options'] = transformer_options + + if 'model_function_wrapper' in 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 + + out_cond /= out_count + del out_count + out_uncond /= out_uncond_count + del out_uncond_count + return out_cond, out_uncond #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): - def get_area_and_mult(conds, x_in, timestep_in): - area = (x_in.shape[2], x_in.shape[3], 0, 0) - strength = 1.0 - - if 'timestep_start' in conds: - timestep_start = conds['timestep_start'] - if timestep_in[0] > timestep_start: - return None - if 'timestep_end' in conds: - timestep_end = conds['timestep_end'] - if timestep_in[0] < timestep_end: - return None - if 'area' in conds: - area = conds['area'] - if 'strength' in conds: - strength = conds['strength'] - - input_x = x_in[:,:,area[2]:area[0] + area[2],area[3]:area[1] + area[3]] - if 'mask' in conds: - # Scale the mask to the size of the input - # The mask should have been resized as we began the sampling process - mask_strength = 1.0 - if "mask_strength" in conds: - mask_strength = conds["mask_strength"] - 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.unsqueeze(1).repeat(input_x.shape[0] // mask.shape[0], input_x.shape[1], 1, 1) - else: - mask = torch.ones_like(input_x) - mult = mask * strength - - if 'mask' not in conds: - rr = 8 - if area[2] != 0: - for t in range(rr): - mult[:,:,t:1+t,:] *= ((1.0/rr) * (t + 1)) - if (area[0] + area[2]) < x_in.shape[2]: - for t in range(rr): - mult[:,:,area[0] - 1 - t:area[0] - t,:] *= ((1.0/rr) * (t + 1)) - if area[3] != 0: - for t in range(rr): - mult[:,:,:,t:1+t] *= ((1.0/rr) * (t + 1)) - if (area[1] + area[3]) < x_in.shape[3]: - for t in range(rr): - mult[:,:,:,area[1] - 1 - t:area[1] - t] *= ((1.0/rr) * (t + 1)) - - conditionning = {} - model_conds = conds["model_conds"] - for c in model_conds: - conditionning[c] = model_conds[c].process_cond(batch_size=x_in.shape[0], device=x_in.device, area=area) - - control = None - if 'control' in conds: - control = conds['control'] - - patches = None - if 'gligen' in conds: - gligen = conds['gligen'] - patches = {} - gligen_type = gligen[0] - gligen_model = gligen[1] - if gligen_type == "position": - gligen_patch = gligen_model.model.set_position(input_x.shape, gligen[2], input_x.device) - else: - gligen_patch = gligen_model.model.set_empty(input_x.shape, input_x.device) - - patches['middle_patch'] = [gligen_patch] - - return (input_x, mult, conditionning, area, control, patches) - - def cond_equal_size(c1, c2): - if c1 is c2: - return True - if c1.keys() != c2.keys(): - return False - for k in c1: - if not c1[k].can_concat(c2[k]): - return False - return True - - def can_concat_cond(c1, c2): - if c1[0].shape != c2[0].shape: - return False - - #control - if (c1[4] is None) != (c2[4] is None): - return False - if c1[4] is not None: - if c1[4] is not c2[4]: - return False - - #patches - if (c1[5] is None) != (c2[5] is None): - return False - if (c1[5] is not None): - if c1[5] is not c2[5]: - return False - - return cond_equal_size(c1[2], c2[2]) - - def cond_cat(c_list): - c_crossattn = [] - c_concat = [] - c_adm = [] - crossattn_max_len = 0 - - temp = {} - for x in c_list: - for k in x: - cur = temp.get(k, []) - cur.append(x[k]) - temp[k] = cur - - out = {} - for k in temp: - conds = temp[k] - out[k] = conds[0].concat(conds[1:]) - - 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 - - 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 - - to_run += [(p, UNCOND)] - - while len(to_run) > 0: - first = to_run[0] - first_shape = first[0][0].shape - to_batch_temp = [] - for x in range(len(to_run)): - if can_concat_cond(to_run[x][0], first[0]): - to_batch_temp += [x] - - to_batch_temp.reverse() - to_batch = to_batch_temp[:1] - - free_memory = model_management.get_free_memory(x_in.device) - 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: - to_batch = batch_amount - break - - input_x = [] - mult = [] - c = [] - cond_or_uncond = [] - area = [] - control = None - patches = None - for x in to_batch: - o = to_run.pop(x) - p = o[0] - input_x += [p[0]] - mult += [p[1]] - c += [p[2]] - area += [p[3]] - cond_or_uncond += [o[1]] - control = p[4] - patches = p[5] - - batch_chunks = len(cond_or_uncond) - input_x = torch.cat(input_x) - c = cond_cat(c) - timestep_ = torch.cat([timestep] * batch_chunks) - - if control is not None: - c['control'] = control.get_control(input_x, timestep_, c, len(cond_or_uncond)) - - transformer_options = {} - if 'transformer_options' in model_options: - transformer_options = model_options['transformer_options'].copy() - - if patches is not None: - if "patches" in transformer_options: - cur_patches = transformer_options["patches"].copy() - for p in patches: - if p in cur_patches: - cur_patches[p] = cur_patches[p] + patches[p] - else: - cur_patches[p] = patches[p] - else: - transformer_options["patches"] = patches - - transformer_options["cond_or_uncond"] = cond_or_uncond[:] - transformer_options["sigmas"] = timestep - - c['transformer_options'] = transformer_options - - if 'model_function_wrapper' in 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 - - out_cond /= out_count - del out_count - out_uncond /= out_uncond_count - del out_uncond_count - return out_cond, out_uncond - - if math.isclose(cond_scale, 1.0): - uncond = None - - cond, uncond = calc_cond_uncond_batch(model, cond, uncond, x, timestep, model_options) - if "sampler_cfg_function" in model_options: - args = {"cond": x - cond, "uncond": x - uncond, "cond_scale": cond_scale, "timestep": timestep, "input": x, "sigma": timestep} - return x - model_options["sampler_cfg_function"](args) + uncond_ = None else: - return uncond + (cond - uncond) * cond_scale + 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} + 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 class CFGNoisePredictor(torch.nn.Module): def __init__(self, model): @@ -276,10 +283,7 @@ class KSamplerX0Inpaint(torch.nn.Module): 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) if denoise_mask is not None: - out *= denoise_mask - - if denoise_mask is not None: - out += self.latent_image * latent_mask + out = out * denoise_mask + self.latent_image * latent_mask return out def simple_scheduler(model, steps): diff --git a/comfy/sd.py b/comfy/sd.py index f4f84d0a0..220637a05 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -151,7 +151,7 @@ class CLIP: return self.patcher.get_key_patches() class VAE: - def __init__(self, sd=None, device=None, config=None): + def __init__(self, sd=None, device=None, config=None, dtype=None): if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format sd = diffusers_convert.convert_vae_state_dict(sd) @@ -188,8 +188,11 @@ class VAE: device = model_management.vae_device() self.device = device offload_device = model_management.vae_offload_device() - self.vae_dtype = model_management.vae_dtype() + if dtype is None: + dtype = model_management.vae_dtype() + self.vae_dtype = dtype self.first_stage_model.to(self.vae_dtype) + self.output_device = model_management.intermediate_device() self.patcher = comfy.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device) @@ -201,9 +204,9 @@ class VAE: 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 = 8, pbar = pbar) + - comfy.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = 8, pbar = pbar) + - comfy.utils.tiled_scale(samples, decode_fn, tile_x, tile_y, overlap, upscale_amount = 8, pbar = pbar)) + (comfy.utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = 8, output_device=self.output_device, pbar = pbar) + + comfy.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = 8, output_device=self.output_device, pbar = pbar) + + comfy.utils.tiled_scale(samples, decode_fn, tile_x, tile_y, overlap, upscale_amount = 8, output_device=self.output_device, pbar = pbar)) / 3.0) / 2.0, min=0.0, max=1.0) return output @@ -214,9 +217,9 @@ class VAE: 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() - samples = comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar) - samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar) - samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = (1/8), out_channels=4, pbar=pbar) + samples = comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/8), out_channels=4, 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 = (1/8), out_channels=4, 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 = (1/8), out_channels=4, output_device=self.output_device, pbar=pbar) samples /= 3.0 return samples @@ -228,15 +231,15 @@ 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] * 8), round(samples_in.shape[3] * 8)), device="cpu") + pixel_samples = torch.empty((samples_in.shape[0], 3, round(samples_in.shape[2] * 8), round(samples_in.shape[3] * 8)), 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).cpu().float() + 1.0) / 2.0, min=0.0, max=1.0) + 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) except model_management.OOM_EXCEPTION as e: print("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.cpu().movedim(1,-1) + pixel_samples = pixel_samples.to(self.output_device).movedim(1,-1) return pixel_samples def decode_tiled(self, samples, tile_x=64, tile_y=64, overlap = 16): @@ -252,10 +255,10 @@ class VAE: free_memory = model_management.get_free_memory(self.device) batch_number = int(free_memory / memory_used) batch_number = max(1, batch_number) - samples = torch.empty((pixel_samples.shape[0], 4, round(pixel_samples.shape[2] // 8), round(pixel_samples.shape[3] // 8)), device="cpu") + samples = torch.empty((pixel_samples.shape[0], 4, round(pixel_samples.shape[2] // 8), round(pixel_samples.shape[3] // 8)), 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) - samples[x:x+batch_number] = self.first_stage_model.encode(pixels_in).cpu().float() + 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.") @@ -432,11 +435,15 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o 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) + if model_config is None: raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path)) @@ -469,7 +476,7 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o print("left over keys:", left_over) if output_model: - model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device(), current_device=inital_load_device) + 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") model_management.load_model_gpu(model_patcher) @@ -480,6 +487,9 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o 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 model_config is None: @@ -500,13 +510,14 @@ def load_unet_state_dict(sd): #load unet in diffusers format else: print(diffusers_keys[k], k) offload_device = model_management.unet_offload_device() + model_config.set_manual_cast(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) - return comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device) + 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) diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index 4e9f6bffe..6ffef515e 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -1,12 +1,14 @@ import os -from transformers import CLIPTokenizer, CLIPTextModel, CLIPTextConfig, modeling_utils +from transformers import CLIPTokenizer import comfy.ops import torch import traceback import zipfile from . import model_management import contextlib +import comfy.clip_model +import json def gen_empty_tokens(special_tokens, length): start_token = special_tokens.get("start", None) @@ -37,7 +39,7 @@ class ClipTokenWeightEncoder: out, pooled = self.encode(to_encode) if pooled is not None: - first_pooled = pooled[0:1].cpu() + first_pooled = pooled[0:1].to(model_management.intermediate_device()) else: first_pooled = pooled @@ -54,8 +56,8 @@ class ClipTokenWeightEncoder: output.append(z) if (len(output) == 0): - return out[-1:].cpu(), first_pooled - return torch.cat(output, dim=-2).cpu(), first_pooled + return out[-1:].to(model_management.intermediate_device()), first_pooled + return torch.cat(output, dim=-2).to(model_management.intermediate_device()), first_pooled class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): """Uses the CLIP transformer encoder for text (from huggingface)""" @@ -65,35 +67,19 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): "hidden" ] 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, textmodel_path=None, dtype=None, - special_tokens={"start": 49406, "end": 49407, "pad": 49407},layer_norm_hidden_state=True, config_class=CLIPTextConfig, - model_class=CLIPTextModel, inner_name="text_model"): # clip-vit-base-patch32 + 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 super().__init__() assert layer in self.LAYERS - self.num_layers = 12 - if textmodel_path is not None: - self.transformer = model_class.from_pretrained(textmodel_path) - else: - if textmodel_json_config is None: - textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json") - config = config_class.from_json_file(textmodel_json_config) - self.num_layers = config.num_hidden_layers - with comfy.ops.use_comfy_ops(device, dtype): - with modeling_utils.no_init_weights(): - self.transformer = model_class(config) - self.inner_name = inner_name - if dtype is not None: - inner_model = getattr(self.transformer, self.inner_name) - if hasattr(inner_model, "embeddings"): - embeddings_bak = inner_model.embeddings.to(torch.float32) - inner_model.embeddings = None - self.transformer.to(dtype) - inner_model.embeddings = embeddings_bak - else: - previous_inputs = self.transformer.get_input_embeddings().to(torch.float32, copy=True) - self.transformer.to(dtype) - self.transformer.set_input_embeddings(previous_inputs) + if textmodel_json_config is None: + textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json") + + with open(textmodel_json_config) as f: + config = json.load(f) + + self.transformer = model_class(config, dtype, device, comfy.ops.manual_cast) + self.num_layers = self.transformer.num_layers self.max_length = max_length if freeze: @@ -108,7 +94,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): self.layer_norm_hidden_state = layer_norm_hidden_state if layer == "hidden": assert layer_idx is not None - assert abs(layer_idx) <= self.num_layers + assert abs(layer_idx) < self.num_layers self.clip_layer(layer_idx) self.layer_default = (self.layer, self.layer_idx) @@ -119,7 +105,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): param.requires_grad = False def clip_layer(self, layer_idx): - if abs(layer_idx) >= self.num_layers: + if abs(layer_idx) > self.num_layers: self.layer = "last" else: self.layer = "hidden" @@ -174,41 +160,31 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): tokens = self.set_up_textual_embeddings(tokens, backup_embeds) tokens = torch.LongTensor(tokens).to(device) - if getattr(self.transformer, self.inner_name).final_layer_norm.weight.dtype != torch.float32: - precision_scope = torch.autocast + attention_mask = None + if self.enable_attention_masks: + attention_mask = torch.zeros_like(tokens) + max_token = self.transformer.get_input_embeddings().weight.shape[0] - 1 + for x in range(attention_mask.shape[0]): + for y in range(attention_mask.shape[1]): + attention_mask[x, y] = 1 + if tokens[x, y] == max_token: + break + + outputs = self.transformer(tokens, attention_mask, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state) + self.transformer.set_input_embeddings(backup_embeds) + + if self.layer == "last": + z = outputs[0] else: - precision_scope = lambda a, dtype: contextlib.nullcontext(a) + z = outputs[1] - with precision_scope(model_management.get_autocast_device(device), dtype=torch.float32): - attention_mask = None - if self.enable_attention_masks: - attention_mask = torch.zeros_like(tokens) - max_token = self.transformer.get_input_embeddings().weight.shape[0] - 1 - for x in range(attention_mask.shape[0]): - for y in range(attention_mask.shape[1]): - attention_mask[x, y] = 1 - if tokens[x, y] == max_token: - break + if outputs[2] is not None: + pooled_output = outputs[2].float() + else: + pooled_output = None - outputs = self.transformer(input_ids=tokens, attention_mask=attention_mask, output_hidden_states=self.layer=="hidden") - self.transformer.set_input_embeddings(backup_embeds) - - if self.layer == "last": - z = outputs.last_hidden_state - elif self.layer == "pooled": - z = outputs.pooler_output[:, None, :] - else: - z = outputs.hidden_states[self.layer_idx] - if self.layer_norm_hidden_state: - z = getattr(self.transformer, self.inner_name).final_layer_norm(z) - - if hasattr(outputs, "pooler_output"): - pooled_output = outputs.pooler_output.float() - else: - pooled_output = None - - 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() + 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): diff --git a/comfy/sd2_clip.py b/comfy/sd2_clip.py index 2ee0ca055..9c878d54a 100644 --- a/comfy/sd2_clip.py +++ b/comfy/sd2_clip.py @@ -3,13 +3,13 @@ import torch import os class SD2ClipHModel(sd1_clip.SDClipModel): - def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, textmodel_path=None, dtype=None): + def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None): if layer == "penultimate": layer="hidden" - layer_idx=23 + layer_idx=-2 textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd2_clip_config.json") - super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, textmodel_path=textmodel_path, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0}) + 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": 0}) class SD2ClipHTokenizer(sd1_clip.SDTokenizer): def __init__(self, tokenizer_path=None, embedding_directory=None): diff --git a/comfy/sdxl_clip.py b/comfy/sdxl_clip.py index 673399e22..b35056bb9 100644 --- a/comfy/sdxl_clip.py +++ b/comfy/sdxl_clip.py @@ -3,13 +3,13 @@ import torch import os class SDXLClipG(sd1_clip.SDClipModel): - def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, textmodel_path=None, dtype=None): + def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None): if layer == "penultimate": layer="hidden" layer_idx=-2 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, textmodel_path=textmodel_path, dtype=dtype, + 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": 0}, layer_norm_hidden_state=False) def load_sd(self, sd): @@ -37,7 +37,7 @@ class SDXLTokenizer: class SDXLClipModel(torch.nn.Module): def __init__(self, device="cpu", dtype=None): super().__init__() - self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=11, device=device, dtype=dtype, layer_norm_hidden_state=False) + 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): diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 455323b96..2f2dee871 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -217,6 +217,16 @@ class SSD1B(SDXL): "use_temporal_attention": False, } +class Segmind_Vega(SDXL): + unet_config = { + "model_channels": 320, + "use_linear_in_transformer": True, + "transformer_depth": [0, 0, 1, 1, 2, 2], + "context_dim": 2048, + "adm_in_channels": 2816, + "use_temporal_attention": False, + } + class SVD_img2vid(supported_models_base.BASE): unet_config = { "model_channels": 320, @@ -242,5 +252,5 @@ class SVD_img2vid(supported_models_base.BASE): def clip_target(self): return None -models = [SD15, SD20, SD21UnclipL, SD21UnclipH, SDXLRefiner, SDXL, SSD1B] +models = [SD15, SD20, SD21UnclipL, SD21UnclipH, SDXLRefiner, SDXL, SSD1B, Segmind_Vega] models += [SVD_img2vid] diff --git a/comfy/supported_models_base.py b/comfy/supported_models_base.py index 3412cfea0..49087d23e 100644 --- a/comfy/supported_models_base.py +++ b/comfy/supported_models_base.py @@ -22,6 +22,8 @@ class BASE: sampling_settings = {} latent_format = latent_formats.LatentFormat + manual_cast_dtype = None + @classmethod def matches(s, unet_config): for k in s.unet_config: @@ -71,3 +73,5 @@ class BASE: replace_prefix = {"": "first_stage_model."} return utils.state_dict_prefix_replace(state_dict, replace_prefix) + def set_manual_cast(self, manual_cast_dtype): + self.manual_cast_dtype = manual_cast_dtype diff --git a/comfy/utils.py b/comfy/utils.py index 505577047..f8026ddab 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -376,7 +376,7 @@ def lanczos(samples, width, height): images = [image.resize((width, height), resample=Image.Resampling.LANCZOS) for image in images] images = [torch.from_numpy(np.array(image).astype(np.float32) / 255.0).movedim(-1, 0) for image in images] result = torch.stack(images) - return result + return result.to(samples.device, samples.dtype) def common_upscale(samples, width, height, upscale_method, crop): if crop == "center": @@ -405,17 +405,17 @@ def get_tiled_scale_steps(width, height, tile_x, tile_y, overlap): return math.ceil((height / (tile_y - overlap))) * math.ceil((width / (tile_x - overlap))) @torch.inference_mode() -def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3, pbar = None): - output = torch.empty((samples.shape[0], out_channels, round(samples.shape[2] * upscale_amount), round(samples.shape[3] * upscale_amount)), device="cpu") +def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3, output_device="cpu", pbar = None): + output = torch.empty((samples.shape[0], out_channels, round(samples.shape[2] * upscale_amount), round(samples.shape[3] * upscale_amount)), device=output_device) for b in range(samples.shape[0]): s = samples[b:b+1] - out = torch.zeros((s.shape[0], out_channels, round(s.shape[2] * upscale_amount), round(s.shape[3] * upscale_amount)), device="cpu") - out_div = torch.zeros((s.shape[0], out_channels, round(s.shape[2] * upscale_amount), round(s.shape[3] * upscale_amount)), device="cpu") + out = torch.zeros((s.shape[0], out_channels, round(s.shape[2] * upscale_amount), round(s.shape[3] * upscale_amount)), device=output_device) + 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): s_in = s[:,:,y:y+tile_y,x:x+tile_x] - ps = function(s_in).cpu() + ps = function(s_in).to(output_device) mask = torch.ones_like(ps) feather = round(overlap * upscale_amount) for t in range(feather): diff --git a/comfy_extras/nodes_canny.py b/comfy_extras/nodes_canny.py index 94d453f2c..730dded5f 100644 --- a/comfy_extras/nodes_canny.py +++ b/comfy_extras/nodes_canny.py @@ -291,7 +291,7 @@ class Canny: def detect_edge(self, image, low_threshold, high_threshold): output = canny(image.to(comfy.model_management.get_torch_device()).movedim(-1, 1), low_threshold, high_threshold) - img_out = output[1].cpu().repeat(1, 3, 1, 1).movedim(1, -1) + img_out = output[1].to(comfy.model_management.intermediate_device()).repeat(1, 3, 1, 1).movedim(1, -1) return (img_out,) NODE_CLASS_MAPPINGS = { diff --git a/comfy_extras/nodes_hypertile.py b/comfy_extras/nodes_hypertile.py index 0d7d4c954..e7446b2e5 100644 --- a/comfy_extras/nodes_hypertile.py +++ b/comfy_extras/nodes_hypertile.py @@ -2,9 +2,10 @@ import math from einops import rearrange -import random +# Use torch rng for consistency across generations +from torch import randint -def random_divisor(value: int, min_value: int, /, max_options: int = 1, counter = 0) -> int: +def random_divisor(value: int, min_value: int, /, max_options: int = 1) -> int: min_value = min(min_value, value) # All big divisors of value (inclusive) @@ -12,8 +13,10 @@ def random_divisor(value: int, min_value: int, /, max_options: int = 1, counter ns = [value // i for i in divisors[:max_options]] # has at least 1 element - random.seed(counter) - idx = random.randint(0, len(ns) - 1) + if len(ns) - 1 > 0: + idx = randint(low=0, high=len(ns) - 1, size=(1,)).item() + else: + idx = 0 return ns[idx] @@ -42,7 +45,6 @@ class HyperTile: latent_tile_size = max(32, tile_size) // 8 self.temp = None - self.counter = 1 def hypertile_in(q, k, v, extra_options): if q.shape[-1] in apply_to: @@ -53,10 +55,8 @@ class HyperTile: h, w = round(math.sqrt(hw * aspect_ratio)), round(math.sqrt(hw / aspect_ratio)) factor = 2**((q.shape[-1] // model_channels) - 1) if scale_depth else 1 - nh = random_divisor(h, latent_tile_size * factor, swap_size, self.counter) - self.counter += 1 - nw = random_divisor(w, latent_tile_size * factor, swap_size, self.counter) - self.counter += 1 + nh = random_divisor(h, latent_tile_size * factor, swap_size) + nw = random_divisor(w, latent_tile_size * factor, swap_size) if nh * nw > 1: q = rearrange(q, "b (nh h nw w) c -> (b nh nw) (h w) c", h=h // nh, w=w // nw, nh=nh, nw=nw) diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 12704f545..71660f8a5 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -226,7 +226,7 @@ class Sharpen: batch_size, height, width, channels = image.shape kernel_size = sharpen_radius * 2 + 1 - kernel = gaussian_kernel(kernel_size, sigma) * -(alpha*10) + kernel = gaussian_kernel(kernel_size, sigma, device=image.device) * -(alpha*10) center = kernel_size // 2 kernel[center, center] = kernel[center, center] - kernel.sum() + 1.0 kernel = kernel.repeat(channels, 1, 1).unsqueeze(1) diff --git a/comfy_extras/nodes_sag.py b/comfy_extras/nodes_sag.py new file mode 100644 index 000000000..4c609565a --- /dev/null +++ b/comfy_extras/nodes_sag.py @@ -0,0 +1,172 @@ +import torch +from torch import einsum +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 +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): + b, _, dim_head = q.shape + dim_head //= heads + scale = dim_head ** -0.5 + + h = heads + 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(), + (q, k, v), + ) + + # force cast to fp32 to avoid overflowing + if _ATTN_PRECISION =="fp32": + with torch.autocast(enabled=False, device_type = 'cuda'): + q, k = q.float(), k.float() + sim = einsum('b i d, b j d -> b i j', q, k) * scale + else: + sim = einsum('b i d, b j d -> b i j', q, k) * scale + + del q, k + + if mask is not None: + mask = rearrange(mask, 'b ... -> b (...)') + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b j -> (b h) () j', h=h) + sim.masked_fill_(~mask, max_neg_value) + + # attention, what we cannot get enough of + sim = sim.softmax(dim=-1) + + out = einsum('b i j, b j d -> b i d', sim.to(v.dtype), v) + out = ( + out.unsqueeze(0) + .reshape(b, heads, -1, dim_head) + .permute(0, 2, 1, 3) + .reshape(b, -1, heads * dim_head) + ) + return (out, sim) + +def create_blur_map(x0, attn, sigma=3.0, threshold=1.0): + # reshape and GAP the attention map + _, hw1, hw2 = attn.shape + b, _, lh, lw = x0.shape + attn = attn.reshape(b, -1, hw1, hw2) + # Global Average Pool + mask = attn.mean(1, keepdim=False).sum(1, keepdim=False) > threshold + ratio = round(math.sqrt(lh * lw / hw1)) + mid_shape = [math.ceil(lh / ratio), math.ceil(lw / ratio)] + + # Reshape + mask = ( + mask.reshape(b, *mid_shape) + .unsqueeze(1) + .type(attn.dtype) + ) + # Upsample + mask = F.interpolate(mask, (lh, lw)) + + blurred = gaussian_blur_2d(x0, kernel_size=9, sigma=sigma) + blurred = blurred * mask + x0 * (1 - mask) + return blurred + +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 = F.pad(img, padding, mode="reflect") + img = F.conv2d(img, kernel2d, groups=img.shape[-3]) + return img + +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}), + "blur_sigma": ("FLOAT", {"default": 2.0, "min": 0.0, "max": 10.0, "step": 0.1}), + }} + RETURN_TYPES = ("MODEL",) + FUNCTION = "patch" + + CATEGORY = "_for_testing" + + def patch(self, model, scale, blur_sigma): + m = model.clone() + + attn_scores = None + mid_block_shape = None + + # TODO: make this work properly with chunked batches + # currently, we can only save the attn from one UNet call + def attn_and_record(q, k, v, extra_options): + nonlocal attn_scores + # if uncond, save the attention scores + heads = extra_options["n_heads"] + cond_or_uncond = extra_options["cond_or_uncond"] + b = q.shape[0] // len(cond_or_uncond) + 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) + # 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) + + def post_cfg_function(args): + nonlocal attn_scores + nonlocal mid_block_shape + uncond_attn = attn_scores + + sag_scale = scale + sag_sigma = blur_sigma + sag_threshold = 1.0 + model = args["model"] + uncond_pred = args["uncond_denoised"] + uncond = args["uncond"] + cfg_result = args["denoised"] + sigma = args["sigma"] + model_options = args["model_options"] + x = args["input"] + + # create the adversarially blurred image + 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) + return cfg_result + (degraded - sag) * sag_scale + + m.set_model_sampler_post_cfg_function(post_cfg_function) + + # from diffusers: + # unet.mid_block.attentions[0].transformer_blocks[0].attn1.patch + m.set_model_attn1_replace(attn_and_record, "middle", 0, 0) + + return (m, ) + +NODE_CLASS_MAPPINGS = { + "SelfAttentionGuidance": SelfAttentionGuidance, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "SelfAttentionGuidance": "Self-Attention Guidance", +} diff --git a/nodes.py b/nodes.py index 24e591fdd..3d24750cb 100644 --- a/nodes.py +++ b/nodes.py @@ -947,8 +947,8 @@ class GLIGENTextBoxApply: return (c, ) class EmptyLatentImage: - def __init__(self, device="cpu"): - self.device = device + def __init__(self): + self.device = comfy.model_management.intermediate_device() @classmethod def INPUT_TYPES(s): @@ -961,7 +961,7 @@ class EmptyLatentImage: CATEGORY = "latent" def generate(self, width, height, batch_size=1): - latent = torch.zeros([batch_size, 4, height // 8, width // 8]) + latent = torch.zeros([batch_size, 4, height // 8, width // 8], device=self.device) return ({"samples":latent}, ) @@ -1867,6 +1867,7 @@ def init_custom_nodes(): "nodes_model_downscale.py", "nodes_images.py", "nodes_video_model.py", + "nodes_sag.py", ] for node_file in extras_files: diff --git a/tests-ui/tests/extensions.test.js b/tests-ui/tests/extensions.test.js index b82e55c32..159e5113a 100644 --- a/tests-ui/tests/extensions.test.js +++ b/tests-ui/tests/extensions.test.js @@ -52,7 +52,7 @@ describe("extensions", () => { const nodeNames = Object.keys(defs); const nodeCount = nodeNames.length; expect(mockExtension.beforeRegisterNodeDef).toHaveBeenCalledTimes(nodeCount); - for (let i = 0; i < nodeCount; i++) { + for (let i = 0; i < 10; i++) { // It should be send the JS class and the original JSON definition const nodeClass = mockExtension.beforeRegisterNodeDef.mock.calls[i][0]; const nodeDef = mockExtension.beforeRegisterNodeDef.mock.calls[i][1]; @@ -133,7 +133,7 @@ describe("extensions", () => { expect(mockExtension.nodeCreated).toHaveBeenCalledTimes(graphData.nodes.length + 2); expect(mockExtension.loadedGraphNode).toHaveBeenCalledTimes(graphData.nodes.length + 1); expect(mockExtension.afterConfigureGraph).toHaveBeenCalledTimes(2); - }); + }, 15000); it("allows custom nodeDefs and widgets to be registered", async () => { const widgetMock = jest.fn((node, inputName, inputData, app) => { diff --git a/tests-ui/tests/groupNode.test.js b/tests-ui/tests/groupNode.test.js index ce54c1154..625890a09 100644 --- a/tests-ui/tests/groupNode.test.js +++ b/tests-ui/tests/groupNode.test.js @@ -1,7 +1,7 @@ // @ts-check /// -const { start, createDefaultWorkflow } = require("../utils"); +const { start, createDefaultWorkflow, getNodeDef, checkBeforeAndAfterReload } = require("../utils"); const lg = require("../utils/litegraph"); describe("group node", () => { @@ -273,7 +273,7 @@ describe("group node", () => { let reroutes = []; let prevNode = nodes.ckpt; - for(let i = 0; i < 5; i++) { + for (let i = 0; i < 5; i++) { const reroute = ez.Reroute(); prevNode.outputs[0].connectTo(reroute.inputs[0]); prevNode = reroute; @@ -283,7 +283,7 @@ describe("group node", () => { const group = await convertToGroup(app, graph, "test", [...reroutes, ...Object.values(nodes)]); expect((await graph.toPrompt()).output).toEqual(getOutput()); - + group.menu["Convert to nodes"].call(); expect((await graph.toPrompt()).output).toEqual(getOutput()); }); @@ -383,6 +383,43 @@ describe("group node", () => { getOutput([nodes.pos.id, nodes.neg.id, nodes.empty.id, nodes.sampler.id]) ); }); + test("groups can connect to each other via internal reroutes", async () => { + const { ez, graph, app } = await start(); + + const latent = ez.EmptyLatentImage(); + const vae = ez.VAELoader(); + const latentReroute = ez.Reroute(); + const vaeReroute = ez.Reroute(); + + latent.outputs[0].connectTo(latentReroute.inputs[0]); + vae.outputs[0].connectTo(vaeReroute.inputs[0]); + + const group1 = await convertToGroup(app, graph, "test", [latentReroute, vaeReroute]); + group1.menu.Clone.call(); + expect(app.graph._nodes).toHaveLength(4); + const group2 = graph.find(app.graph._nodes[3]); + expect(group2.node.type).toEqual("workflow/test"); + expect(group2.id).not.toEqual(group1.id); + + group1.outputs.VAE.connectTo(group2.inputs.VAE); + group1.outputs.LATENT.connectTo(group2.inputs.LATENT); + + const decode = ez.VAEDecode(group2.outputs.LATENT, group2.outputs.VAE); + const preview = ez.PreviewImage(decode.outputs[0]); + + const output = { + [latent.id]: { inputs: { width: 512, height: 512, batch_size: 1 }, class_type: "EmptyLatentImage" }, + [vae.id]: { inputs: { vae_name: "vae1.safetensors" }, class_type: "VAELoader" }, + [decode.id]: { inputs: { samples: [latent.id + "", 0], vae: [vae.id + "", 0] }, class_type: "VAEDecode" }, + [preview.id]: { inputs: { images: [decode.id + "", 0] }, class_type: "PreviewImage" }, + }; + expect((await graph.toPrompt()).output).toEqual(output); + + // Ensure missing connections dont cause errors + group2.inputs.VAE.disconnect(); + delete output[decode.id].inputs.vae; + expect((await graph.toPrompt()).output).toEqual(output); + }); test("displays generated image on group node", async () => { const { ez, graph, app } = await start(); const nodes = createDefaultWorkflow(ez, graph); @@ -642,6 +679,55 @@ describe("group node", () => { 2: { inputs: { text: "positive" }, class_type: "CLIPTextEncode" }, }); }); + test("correctly handles widget inputs", async () => { + const { ez, graph, app } = await start(); + const upscaleMethods = (await getNodeDef("ImageScaleBy")).input.required["upscale_method"][0]; + + const image = ez.LoadImage(); + const scale1 = ez.ImageScaleBy(image.outputs[0]); + const scale2 = ez.ImageScaleBy(image.outputs[0]); + const preview1 = ez.PreviewImage(scale1.outputs[0]); + const preview2 = ez.PreviewImage(scale2.outputs[0]); + scale1.widgets.upscale_method.value = upscaleMethods[1]; + scale1.widgets.upscale_method.convertToInput(); + + const group = await convertToGroup(app, graph, "test", [scale1, scale2]); + expect(group.inputs.length).toBe(3); + expect(group.inputs[0].input.type).toBe("IMAGE"); + expect(group.inputs[1].input.type).toBe("IMAGE"); + expect(group.inputs[2].input.type).toBe("COMBO"); + + // Ensure links are maintained + expect(group.inputs[0].connection?.originNode?.id).toBe(image.id); + expect(group.inputs[1].connection?.originNode?.id).toBe(image.id); + expect(group.inputs[2].connection).toBeFalsy(); + + // Ensure primitive gets correct type + const primitive = ez.PrimitiveNode(); + primitive.outputs[0].connectTo(group.inputs[2]); + expect(primitive.widgets.value.widget.options.values).toBe(upscaleMethods); + expect(primitive.widgets.value.value).toBe(upscaleMethods[1]); // Ensure value is copied + primitive.widgets.value.value = upscaleMethods[1]; + + await checkBeforeAndAfterReload(graph, async (r) => { + const scale1id = r ? `${group.id}:0` : scale1.id; + const scale2id = r ? `${group.id}:1` : scale2.id; + // Ensure widget value is applied to prompt + expect((await graph.toPrompt()).output).toStrictEqual({ + [image.id]: { inputs: { image: "example.png", upload: "image" }, class_type: "LoadImage" }, + [scale1id]: { + inputs: { upscale_method: upscaleMethods[1], scale_by: 1, image: [`${image.id}`, 0] }, + class_type: "ImageScaleBy", + }, + [scale2id]: { + inputs: { upscale_method: "nearest-exact", scale_by: 1, image: [`${image.id}`, 0] }, + class_type: "ImageScaleBy", + }, + [preview1.id]: { inputs: { images: [`${scale1id}`, 0] }, class_type: "PreviewImage" }, + [preview2.id]: { inputs: { images: [`${scale2id}`, 0] }, class_type: "PreviewImage" }, + }); + }); + }); test("adds widgets in node execution order", async () => { const { ez, graph, app } = await start(); const scale = ez.LatentUpscale(); @@ -815,4 +901,73 @@ describe("group node", () => { expect(p2.widgets.control_after_generate.value).toBe("randomize"); expect(p2.widgets.control_filter_list.value).toBe("/.+/"); }); + test("internal reroutes work with converted inputs and merge options", async () => { + const { ez, graph, app } = await start(); + const vae = ez.VAELoader(); + const latent = ez.EmptyLatentImage(); + const decode = ez.VAEDecode(latent.outputs.LATENT, vae.outputs.VAE); + const scale = ez.ImageScale(decode.outputs.IMAGE); + ez.PreviewImage(scale.outputs.IMAGE); + + const r1 = ez.Reroute(); + const r2 = ez.Reroute(); + + latent.widgets.width.value = 64; + latent.widgets.height.value = 128; + + latent.widgets.width.convertToInput(); + latent.widgets.height.convertToInput(); + latent.widgets.batch_size.convertToInput(); + + scale.widgets.width.convertToInput(); + scale.widgets.height.convertToInput(); + + r1.inputs[0].input.label = "hbw"; + r1.outputs[0].connectTo(latent.inputs.height); + r1.outputs[0].connectTo(latent.inputs.batch_size); + r1.outputs[0].connectTo(scale.inputs.width); + + r2.inputs[0].input.label = "wh"; + r2.outputs[0].connectTo(latent.inputs.width); + r2.outputs[0].connectTo(scale.inputs.height); + + const group = await convertToGroup(app, graph, "test", [r1, r2, latent, decode, scale]); + + expect(group.inputs[0].input.type).toBe("VAE"); + expect(group.inputs[1].input.type).toBe("INT"); + expect(group.inputs[2].input.type).toBe("INT"); + + const p1 = ez.PrimitiveNode(); + const p2 = ez.PrimitiveNode(); + p1.outputs[0].connectTo(group.inputs[1]); + p2.outputs[0].connectTo(group.inputs[2]); + + expect(p1.widgets.value.widget.options?.min).toBe(16); // width/height min + expect(p1.widgets.value.widget.options?.max).toBe(4096); // batch max + 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?.step).toBe(80); // width/height step * 10 + + expect(p1.widgets.value.value).toBe(128); + expect(p2.widgets.value.value).toBe(64); + + p1.widgets.value.value = 16; + p2.widgets.value.value = 32; + + await checkBeforeAndAfterReload(graph, async (r) => { + const id = (v) => (r ? `${group.id}:` : "") + v; + expect((await graph.toPrompt()).output).toStrictEqual({ + 1: { inputs: { vae_name: "vae1.safetensors" }, class_type: "VAELoader" }, + [id(2)]: { inputs: { width: 32, height: 16, batch_size: 16 }, class_type: "EmptyLatentImage" }, + [id(3)]: { inputs: { samples: [id(2), 0], vae: ["1", 0] }, class_type: "VAEDecode" }, + [id(4)]: { + inputs: { upscale_method: "nearest-exact", width: 16, height: 32, crop: "disabled", image: [id(3), 0] }, + class_type: "ImageScale", + }, + 5: { inputs: { images: [id(4), 0] }, class_type: "PreviewImage" }, + }); + }); + }); }); diff --git a/tests-ui/tests/widgetInputs.test.js b/tests-ui/tests/widgetInputs.test.js index 8e191adf0..67e3fa341 100644 --- a/tests-ui/tests/widgetInputs.test.js +++ b/tests-ui/tests/widgetInputs.test.js @@ -1,7 +1,13 @@ // @ts-check /// -const { start, makeNodeDef, checkBeforeAndAfterReload, assertNotNullOrUndefined } = require("../utils"); +const { + start, + makeNodeDef, + checkBeforeAndAfterReload, + assertNotNullOrUndefined, + createDefaultWorkflow, +} = require("../utils"); const lg = require("../utils/litegraph"); /** @@ -36,7 +42,7 @@ async function connectPrimitiveAndReload(ez, graph, input, widgetType, controlWi if (controlWidgetCount) { const controlWidget = primitive.widgets.control_after_generate; expect(controlWidget.widget.type).toBe("combo"); - if(widgetType === "combo") { + if (widgetType === "combo") { const filterWidget = primitive.widgets.control_filter_list; expect(filterWidget.widget.type).toBe("string"); } @@ -308,8 +314,8 @@ describe("widget inputs", () => { const { ez } = await start({ mockNodeDefs: { ...makeNodeDef("TestNode1", {}, [["A", "B"]]), - ...makeNodeDef("TestNode2", { example: [["A", "B"], { forceInput: true}] }), - ...makeNodeDef("TestNode3", { example: [["A", "B", "C"], { forceInput: true}] }), + ...makeNodeDef("TestNode2", { example: [["A", "B"], { forceInput: true }] }), + ...makeNodeDef("TestNode3", { example: [["A", "B", "C"], { forceInput: true }] }), }, }); @@ -330,7 +336,7 @@ describe("widget inputs", () => { const n1 = ez.TestNode1(); n1.widgets.example.convertToInput(); - const p = ez.PrimitiveNode() + const p = ez.PrimitiveNode(); p.outputs[0].connectTo(n1.inputs[0]); const value = p.widgets.value; @@ -380,7 +386,7 @@ describe("widget inputs", () => { // Check random control.value = "randomize"; filter.value = "/D/"; - for(let i = 0; i < 100; i++) { + for (let i = 0; i < 100; i++) { control["afterQueued"](); expect(value.value === "D" || value.value === "DD").toBeTruthy(); } @@ -392,4 +398,160 @@ describe("widget inputs", () => { control["afterQueued"](); expect(value.value).toBe("B"); }); + + describe("reroutes", () => { + async function checkOutput(graph, values) { + expect((await graph.toPrompt()).output).toStrictEqual({ + 1: { inputs: { ckpt_name: "model1.safetensors" }, class_type: "CheckpointLoaderSimple" }, + 2: { inputs: { text: "positive", clip: ["1", 1] }, class_type: "CLIPTextEncode" }, + 3: { inputs: { text: "negative", clip: ["1", 1] }, class_type: "CLIPTextEncode" }, + 4: { + inputs: { width: values.width ?? 512, height: values.height ?? 512, batch_size: values?.batch_size ?? 1 }, + class_type: "EmptyLatentImage", + }, + 5: { + inputs: { + seed: 0, + steps: 20, + cfg: 8, + sampler_name: "euler", + scheduler: values?.scheduler ?? "normal", + denoise: 1, + model: ["1", 0], + positive: ["2", 0], + negative: ["3", 0], + latent_image: ["4", 0], + }, + class_type: "KSampler", + }, + 6: { inputs: { samples: ["5", 0], vae: ["1", 2] }, class_type: "VAEDecode" }, + 7: { + inputs: { filename_prefix: values.filename_prefix ?? "ComfyUI", images: ["6", 0] }, + class_type: "SaveImage", + }, + }); + } + + async function waitForWidget(node) { + // widgets are created slightly after the graph is ready + // hard to find an exact hook to get these so just wait for them to be ready + for (let i = 0; i < 10; i++) { + await new Promise((r) => setTimeout(r, 10)); + if (node.widgets?.value) { + return; + } + } + } + + it("can connect primitive via a reroute path to a widget input", async () => { + const { ez, graph } = await start(); + const nodes = createDefaultWorkflow(ez, graph); + + nodes.empty.widgets.width.convertToInput(); + nodes.sampler.widgets.scheduler.convertToInput(); + nodes.save.widgets.filename_prefix.convertToInput(); + + let widthReroute = ez.Reroute(); + let schedulerReroute = ez.Reroute(); + let fileReroute = ez.Reroute(); + + let widthNext = widthReroute; + let schedulerNext = schedulerReroute; + let fileNext = fileReroute; + + for (let i = 0; i < 5; i++) { + let next = ez.Reroute(); + widthNext.outputs[0].connectTo(next.inputs[0]); + widthNext = next; + + next = ez.Reroute(); + schedulerNext.outputs[0].connectTo(next.inputs[0]); + schedulerNext = next; + + next = ez.Reroute(); + fileNext.outputs[0].connectTo(next.inputs[0]); + fileNext = next; + } + + widthNext.outputs[0].connectTo(nodes.empty.inputs.width); + schedulerNext.outputs[0].connectTo(nodes.sampler.inputs.scheduler); + fileNext.outputs[0].connectTo(nodes.save.inputs.filename_prefix); + + let widthPrimitive = ez.PrimitiveNode(); + let schedulerPrimitive = ez.PrimitiveNode(); + let filePrimitive = ez.PrimitiveNode(); + + widthPrimitive.outputs[0].connectTo(widthReroute.inputs[0]); + schedulerPrimitive.outputs[0].connectTo(schedulerReroute.inputs[0]); + filePrimitive.outputs[0].connectTo(fileReroute.inputs[0]); + expect(widthPrimitive.widgets.value.value).toBe(512); + widthPrimitive.widgets.value.value = 1024; + expect(schedulerPrimitive.widgets.value.value).toBe("normal"); + schedulerPrimitive.widgets.value.value = "simple"; + expect(filePrimitive.widgets.value.value).toBe("ComfyUI"); + filePrimitive.widgets.value.value = "ComfyTest"; + + await checkBeforeAndAfterReload(graph, async () => { + widthPrimitive = graph.find(widthPrimitive); + schedulerPrimitive = graph.find(schedulerPrimitive); + filePrimitive = graph.find(filePrimitive); + await waitForWidget(filePrimitive); + expect(widthPrimitive.widgets.length).toBe(2); + expect(schedulerPrimitive.widgets.length).toBe(3); + expect(filePrimitive.widgets.length).toBe(1); + + await checkOutput(graph, { + width: 1024, + scheduler: "simple", + filename_prefix: "ComfyTest", + }); + }); + }); + it("can connect primitive via a reroute path to multiple widget inputs", async () => { + const { ez, graph } = await start(); + const nodes = createDefaultWorkflow(ez, graph); + + nodes.empty.widgets.width.convertToInput(); + nodes.empty.widgets.height.convertToInput(); + nodes.empty.widgets.batch_size.convertToInput(); + + let reroute = ez.Reroute(); + let prevReroute = reroute; + for (let i = 0; i < 5; i++) { + const next = ez.Reroute(); + prevReroute.outputs[0].connectTo(next.inputs[0]); + prevReroute = next; + } + + const r1 = ez.Reroute(prevReroute.outputs[0]); + const r2 = ez.Reroute(prevReroute.outputs[0]); + const r3 = ez.Reroute(r2.outputs[0]); + const r4 = ez.Reroute(r2.outputs[0]); + + r1.outputs[0].connectTo(nodes.empty.inputs.width); + r3.outputs[0].connectTo(nodes.empty.inputs.height); + r4.outputs[0].connectTo(nodes.empty.inputs.batch_size); + + let primitive = ez.PrimitiveNode(); + primitive.outputs[0].connectTo(reroute.inputs[0]); + expect(primitive.widgets.value.value).toBe(1); + primitive.widgets.value.value = 64; + + await checkBeforeAndAfterReload(graph, async (r) => { + primitive = graph.find(primitive); + await waitForWidget(primitive); + + // Ensure widget configs are merged + expect(primitive.widgets.value.widget.options?.min).toBe(16); // width/height min + expect(primitive.widgets.value.widget.options?.max).toBe(4096); // batch max + expect(primitive.widgets.value.widget.options?.step).toBe(80); // width/height step * 10 + + await checkOutput(graph, { + width: 64, + height: 64, + batch_size: 64, + }); + }); + }); + }); }); diff --git a/tests-ui/utils/ezgraph.js b/tests-ui/utils/ezgraph.js index 898b82db0..8a55246ee 100644 --- a/tests-ui/utils/ezgraph.js +++ b/tests-ui/utils/ezgraph.js @@ -78,6 +78,14 @@ export class EzInput extends EzSlot { this.input = input; } + get connection() { + const link = this.node.node.inputs?.[this.index]?.link; + if (link == null) { + return null; + } + return new EzConnection(this.node.app, this.node.app.graph.links[link]); + } + disconnect() { this.node.node.disconnectInput(this.index); } @@ -117,7 +125,7 @@ export class EzOutput extends EzSlot { const inp = input.input; const inName = inp.name || inp.label || inp.type; throw new Error( - `Connecting from ${input.node.node.type}[${inName}#${input.index}] -> ${this.node.node.type}[${ + `Connecting from ${input.node.node.type}#${input.node.id}[${inName}#${input.index}] -> ${this.node.node.type}#${this.node.id}[${ this.output.name ?? this.output.type }#${this.index}] failed.` ); @@ -179,6 +187,7 @@ export class EzWidget { set value(v) { this.widget.value = v; + this.widget.callback?.call?.(this.widget, v) } get isConvertedToInput() { @@ -319,7 +328,7 @@ export class EzGraph { } stringify() { - return JSON.stringify(this.app.graph.serialize(), undefined, "\t"); + return JSON.stringify(this.app.graph.serialize(), undefined); } /** diff --git a/tests-ui/utils/index.js b/tests-ui/utils/index.js index 3a018f566..6a08e8594 100644 --- a/tests-ui/utils/index.js +++ b/tests-ui/utils/index.js @@ -104,3 +104,12 @@ export function createDefaultWorkflow(ez, graph) { return { ckpt, pos, neg, empty, sampler, decode, save }; } + +export async function getNodeDefs() { + const { api } = require("../../web/scripts/api"); + return api.getNodeDefs(); +} + +export async function getNodeDef(nodeId) { + return (await getNodeDefs())[nodeId]; +} \ No newline at end of file diff --git a/web/extensions/core/groupNode.js b/web/extensions/core/groupNode.js index 6766f356d..dc962ac24 100644 --- a/web/extensions/core/groupNode.js +++ b/web/extensions/core/groupNode.js @@ -174,6 +174,11 @@ export class GroupNodeConfig { node.index = i; this.processNode(node, seenInputs, seenOutputs); } + + for (const p of this.#convertedToProcess) { + p(); + } + this.#convertedToProcess = null; await app.registerNodeDef("workflow/" + this.name, this.nodeDef); } @@ -192,7 +197,10 @@ export class GroupNodeConfig { if (!this.linksFrom[sourceNodeId]) { this.linksFrom[sourceNodeId] = {}; } - this.linksFrom[sourceNodeId][sourceNodeSlot] = l; + if (!this.linksFrom[sourceNodeId][sourceNodeSlot]) { + this.linksFrom[sourceNodeId][sourceNodeSlot] = []; + } + this.linksFrom[sourceNodeId][sourceNodeSlot].push(l); if (!this.linksTo[targetNodeId]) { this.linksTo[targetNodeId] = {}; @@ -230,11 +238,11 @@ export class GroupNodeConfig { // Skip as its not linked if (!linksFrom) return; - let type = linksFrom["0"][5]; + let type = linksFrom["0"][0][5]; if (type === "COMBO") { // Use the array items const source = node.outputs[0].widget.name; - const fromTypeName = this.nodeData.nodes[linksFrom["0"][2]].type; + const fromTypeName = this.nodeData.nodes[linksFrom["0"][0][2]].type; const fromType = globalDefs[fromTypeName]; const input = fromType.input.required[source] ?? fromType.input.optional[source]; type = input[0]; @@ -258,10 +266,33 @@ export class GroupNodeConfig { return null; } + let config = {}; let rerouteType = "*"; if (linksFrom) { - const [, , id, slot] = linksFrom["0"]; - rerouteType = this.nodeData.nodes[id].inputs[slot].type; + for (const [, , id, slot] of linksFrom["0"]) { + const node = this.nodeData.nodes[id]; + const input = node.inputs[slot]; + if (rerouteType === "*") { + rerouteType = input.type; + } + if (input.widget) { + const targetDef = globalDefs[node.type]; + const targetWidget = + targetDef.input.required[input.widget.name] ?? targetDef.input.optional[input.widget.name]; + + const widget = [targetWidget[0], config]; + const res = mergeIfValid( + { + widget, + }, + targetWidget, + false, + null, + widget + ); + config = res?.customConfig ?? config; + } + } } else if (linksTo) { const [id, slot] = linksTo["0"]; rerouteType = this.nodeData.nodes[id].outputs[slot].type; @@ -282,10 +313,11 @@ export class GroupNodeConfig { } } + config.forceInput = true; return { input: { required: { - [rerouteType]: [rerouteType, {}], + [rerouteType]: [rerouteType, config], }, }, output: [rerouteType], @@ -420,10 +452,18 @@ export class GroupNodeConfig { defaultInput: true, }); this.nodeDef.input.required[name] = config; + this.newToOldWidgetMap[name] = { node, inputName }; + + if (!this.oldToNewWidgetMap[node.index]) { + this.oldToNewWidgetMap[node.index] = {}; + } + this.oldToNewWidgetMap[node.index][inputName] = name; + inputMap[slots.length + i] = this.inputCount++; } } + #convertedToProcess = []; processNodeInputs(node, seenInputs, inputs) { const inputMapping = []; @@ -434,7 +474,11 @@ export class GroupNodeConfig { const linksTo = this.linksTo[node.index] ?? {}; const inputMap = (this.oldToNewInputMap[node.index] = {}); this.processInputSlots(inputs, node, slots, linksTo, inputMap, seenInputs); - this.processConvertedWidgets(inputs, node, slots, converted, linksTo, inputMap, seenInputs); + + // Converted inputs have to be processed after all other nodes as they'll be at the end of the list + this.#convertedToProcess.push(() => + this.processConvertedWidgets(inputs, node, slots, converted, linksTo, inputMap, seenInputs) + ); return inputMapping; } @@ -597,11 +641,19 @@ export class GroupNodeHandler { const output = this.groupData.newToOldOutputMap[link.origin_slot]; let innerNode = this.innerNodes[output.node.index]; let l; - while (innerNode.type === "Reroute") { + while (innerNode?.type === "Reroute") { l = innerNode.getInputLink(0); innerNode = innerNode.getInputNode(0); } + if (!innerNode) { + return null; + } + + if (l && GroupNodeHandler.isGroupNode(innerNode)) { + return innerNode.updateLink(l); + } + link.origin_id = innerNode.id; link.origin_slot = l?.origin_slot ?? output.slot; return link; @@ -665,6 +717,8 @@ export class GroupNodeHandler { top = newNode.pos[1]; } + if (!newNode.widgets) continue; + const map = this.groupData.oldToNewWidgetMap[innerNode.index]; if (map) { const widgets = Object.keys(map); @@ -721,7 +775,7 @@ export class GroupNodeHandler { } }; - const reconnectOutputs = () => { + const reconnectOutputs = (selectedIds) => { for (let groupOutputId = 0; groupOutputId < node.outputs?.length; groupOutputId++) { const output = node.outputs[groupOutputId]; if (!output.links) continue; @@ -861,7 +915,7 @@ export class GroupNodeHandler { if (innerNode.type === "PrimitiveNode") { innerNode.primitiveValue = newValue; const primitiveLinked = this.groupData.primitiveToWidget[old.node.index]; - for (const linked of primitiveLinked) { + for (const linked of primitiveLinked ?? []) { const node = this.innerNodes[linked.nodeId]; const widget = node.widgets.find((w) => w.name === linked.inputName); @@ -870,6 +924,18 @@ export class GroupNodeHandler { } } continue; + } else if (innerNode.type === "Reroute") { + const rerouteLinks = this.groupData.linksFrom[old.node.index]; + for (const [_, , targetNodeId, targetSlot] of rerouteLinks["0"]) { + const node = this.innerNodes[targetNodeId]; + const input = node.inputs[targetSlot]; + if (input.widget) { + const widget = node.widgets?.find((w) => w.name === input.widget.name); + if (widget) { + widget.value = newValue; + } + } + } } const widget = innerNode.widgets?.find((w) => w.name === old.inputName); @@ -897,33 +963,58 @@ export class GroupNodeHandler { this.node.widgets[targetWidgetIndex + i].value = primitiveNode.widgets[i].value; } } + return true; } + populateReroute(node, nodeId, map) { + if (node.type !== "Reroute") return; + + const link = this.groupData.linksFrom[nodeId]?.[0]?.[0]; + if (!link) return; + const [, , targetNodeId, targetNodeSlot] = link; + const targetNode = this.groupData.nodeData.nodes[targetNodeId]; + const inputs = targetNode.inputs; + const targetWidget = inputs?.[targetNodeSlot].widget; + if (!targetWidget) return; + + const offset = inputs.length - (targetNode.widgets_values?.length ?? 0); + const v = targetNode.widgets_values?.[targetNodeSlot - offset]; + if (v == null) return; + + const widgetName = Object.values(map)[0]; + const widget = this.node.widgets.find(w => w.name === widgetName); + if(widget) { + widget.value = v; + } + } + + populateWidgets() { + if (!this.node.widgets) return; + for (let nodeId = 0; nodeId < this.groupData.nodeData.nodes.length; nodeId++) { const node = this.groupData.nodeData.nodes[nodeId]; - - if (!node.widgets_values?.length) continue; - - const map = this.groupData.oldToNewWidgetMap[nodeId]; + const map = this.groupData.oldToNewWidgetMap[nodeId] ?? {}; const widgets = Object.keys(map); + if (!node.widgets_values?.length) { + // special handling for populating values into reroutes + // this allows primitives connect to them to pick up the correct value + this.populateReroute(node, nodeId, map); + continue; + } + let linkedShift = 0; for (let i = 0; i < widgets.length; i++) { const oldName = widgets[i]; const newName = map[oldName]; const widgetIndex = this.node.widgets.findIndex((w) => w.name === newName); const mainWidget = this.node.widgets[widgetIndex]; - if (!newName) { - // New name will be null if its a converted widget - this.populatePrimitive(node, nodeId, oldName, i, linkedShift); - + if (this.populatePrimitive(node, nodeId, oldName, i, linkedShift)) { // Find the inner widget and shift by the number of linked widgets as they will have been removed too const innerWidget = this.innerNodes[nodeId].widgets?.find((w) => w.name === oldName); linkedShift += innerWidget.linkedWidgets?.length ?? 0; - continue; } - if (widgetIndex === -1) { continue; } diff --git a/web/extensions/core/maskeditor.js b/web/extensions/core/maskeditor.js index 8ace79562..bb2f16d42 100644 --- a/web/extensions/core/maskeditor.js +++ b/web/extensions/core/maskeditor.js @@ -33,6 +33,18 @@ function loadedImageToBlob(image) { return blob; } +function loadImage(imagePath) { + return new Promise((resolve, reject) => { + const image = new Image(); + + image.onload = function() { + resolve(image); + }; + + image.src = imagePath; + }); +} + async function uploadMask(filepath, formData) { await api.fetchApi('/upload/mask', { method: 'POST', @@ -50,25 +62,25 @@ async function uploadMask(filepath, formData) { ClipspaceDialog.invalidatePreview(); } -function prepareRGB(image, backupCanvas, backupCtx) { +function prepare_mask(image, maskCanvas, maskCtx) { // paste mask data into alpha channel - backupCtx.drawImage(image, 0, 0, backupCanvas.width, backupCanvas.height); - const backupData = backupCtx.getImageData(0, 0, backupCanvas.width, backupCanvas.height); + maskCtx.drawImage(image, 0, 0, maskCanvas.width, maskCanvas.height); + const maskData = maskCtx.getImageData(0, 0, maskCanvas.width, maskCanvas.height); - // refine mask image - for (let i = 0; i < backupData.data.length; i += 4) { - if(backupData.data[i+3] == 255) - backupData.data[i+3] = 0; + // invert mask + for (let i = 0; i < maskData.data.length; i += 4) { + if(maskData.data[i+3] == 255) + maskData.data[i+3] = 0; else - backupData.data[i+3] = 255; + maskData.data[i+3] = 255; - backupData.data[i] = 0; - backupData.data[i+1] = 0; - backupData.data[i+2] = 0; + maskData.data[i] = 0; + maskData.data[i+1] = 0; + maskData.data[i+2] = 0; } - backupCtx.globalCompositeOperation = 'source-over'; - backupCtx.putImageData(backupData, 0, 0); + maskCtx.globalCompositeOperation = 'source-over'; + maskCtx.putImageData(maskData, 0, 0); } class MaskEditorDialog extends ComfyDialog { @@ -155,10 +167,6 @@ class MaskEditorDialog extends ComfyDialog { // If it is specified as relative, using it only as a hidden placeholder for padding is recommended // to prevent anomalies where it exceeds a certain size and goes outside of the window. - var placeholder = document.createElement("div"); - placeholder.style.position = "relative"; - placeholder.style.height = "50px"; - var bottom_panel = document.createElement("div"); bottom_panel.style.position = "absolute"; bottom_panel.style.bottom = "0px"; @@ -180,18 +188,16 @@ class MaskEditorDialog extends ComfyDialog { this.brush = brush; this.element.appendChild(imgCanvas); this.element.appendChild(maskCanvas); - this.element.appendChild(placeholder); // must below z-index than bottom_panel to avoid covering button this.element.appendChild(bottom_panel); document.body.appendChild(brush); - var brush_size_slider = this.createLeftSlider(self, "Thickness", (event) => { + 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); - self.backupCtx.clearRect(0, 0, self.backupCanvas.width, self.backupCanvas.height); }); var cancelButton = this.createRightButton("Cancel", () => { document.removeEventListener("mouseup", MaskEditorDialog.handleMouseUp); @@ -207,40 +213,42 @@ class MaskEditorDialog extends ComfyDialog { this.element.appendChild(imgCanvas); this.element.appendChild(maskCanvas); - this.element.appendChild(placeholder); // must below z-index than bottom_panel to avoid covering button this.element.appendChild(bottom_panel); bottom_panel.appendChild(clearButton); bottom_panel.appendChild(this.saveButton); bottom_panel.appendChild(cancelButton); - bottom_panel.appendChild(brush_size_slider); + bottom_panel.appendChild(this.brush_size_slider); + + imgCanvas.style.position = "absolute"; + maskCanvas.style.position = "absolute"; - imgCanvas.style.position = "relative"; imgCanvas.style.top = "200"; imgCanvas.style.left = "0"; - maskCanvas.style.position = "absolute"; + maskCanvas.style.top = imgCanvas.style.top; + maskCanvas.style.left = imgCanvas.style.left; } - show() { + async show() { + this.zoom_ratio = 1.0; + this.pan_x = 0; + this.pan_y = 0; + if(!this.is_layout_created) { // layout const imgCanvas = document.createElement('canvas'); const maskCanvas = document.createElement('canvas'); - const backupCanvas = document.createElement('canvas'); imgCanvas.id = "imageCanvas"; maskCanvas.id = "maskCanvas"; - backupCanvas.id = "backupCanvas"; this.setlayout(imgCanvas, maskCanvas); // prepare content this.imgCanvas = imgCanvas; this.maskCanvas = maskCanvas; - this.backupCanvas = backupCanvas; - this.maskCtx = maskCanvas.getContext('2d'); - this.backupCtx = backupCanvas.getContext('2d'); + this.maskCtx = maskCanvas.getContext('2d', {willReadFrequently: true }); this.setEventHandler(maskCanvas); @@ -252,6 +260,8 @@ class MaskEditorDialog extends ComfyDialog { mutations.forEach(function(mutation) { if (mutation.type === 'attributes' && mutation.attributeName === 'style') { if(self.last_display_style && self.last_display_style != 'none' && self.element.style.display == 'none') { + document.removeEventListener("mouseup", MaskEditorDialog.handleMouseUp); + self.brush.style.display = "none"; ComfyApp.onClipspaceEditorClosed(); } @@ -264,7 +274,8 @@ class MaskEditorDialog extends ComfyDialog { observer.observe(this.element, config); } - this.setImages(this.imgCanvas, this.backupCanvas); + // The keydown event needs to be reconfigured when closing the dialog as it gets removed. + document.addEventListener('keydown', MaskEditorDialog.handleKeyDown); if(ComfyApp.clipspace_return_node) { this.saveButton.innerText = "Save to node"; @@ -275,97 +286,157 @@ class MaskEditorDialog extends ComfyDialog { this.saveButton.disabled = false; this.element.style.display = "block"; + this.element.style.width = "85%"; + this.element.style.margin = "0 7.5%"; + this.element.style.height = "100vh"; + this.element.style.top = "50%"; + this.element.style.left = "42%"; this.element.style.zIndex = 8888; // NOTE: alert dialog must be high priority. + + await this.setImages(this.imgCanvas); + + this.is_visible = true; } isOpened() { return this.element.style.display == "block"; } - setImages(imgCanvas, backupCanvas) { - const imgCtx = imgCanvas.getContext('2d'); - const backupCtx = backupCanvas.getContext('2d'); + invalidateCanvas(orig_image, mask_image) { + this.imgCanvas.width = orig_image.width; + this.imgCanvas.height = orig_image.height; + + this.maskCanvas.width = orig_image.width; + this.maskCanvas.height = orig_image.height; + + let imgCtx = this.imgCanvas.getContext('2d', {willReadFrequently: true }); + 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); + } + + async setImages(imgCanvas) { + let self = this; + + const imgCtx = imgCanvas.getContext('2d', {willReadFrequently: true }); const maskCtx = this.maskCtx; const maskCanvas = this.maskCanvas; - backupCtx.clearRect(0,0,this.backupCanvas.width,this.backupCanvas.height); imgCtx.clearRect(0,0,this.imgCanvas.width,this.imgCanvas.height); maskCtx.clearRect(0,0,this.maskCanvas.width,this.maskCanvas.height); // image load - const orig_image = new Image(); - window.addEventListener("resize", () => { - // repositioning - imgCanvas.width = window.innerWidth - 250; - imgCanvas.height = window.innerHeight - 200; - - // redraw image - let drawWidth = orig_image.width; - let drawHeight = orig_image.height; - if (orig_image.width > imgCanvas.width) { - drawWidth = imgCanvas.width; - drawHeight = (drawWidth / orig_image.width) * orig_image.height; - } - - if (drawHeight > imgCanvas.height) { - drawHeight = imgCanvas.height; - drawWidth = (drawHeight / orig_image.height) * orig_image.width; - } - - imgCtx.drawImage(orig_image, 0, 0, drawWidth, drawHeight); - - // update mask - maskCanvas.width = drawWidth; - maskCanvas.height = drawHeight; - maskCanvas.style.top = imgCanvas.offsetTop + "px"; - maskCanvas.style.left = imgCanvas.offsetLeft + "px"; - backupCtx.drawImage(maskCanvas, 0, 0, maskCanvas.width, maskCanvas.height, 0, 0, backupCanvas.width, backupCanvas.height); - maskCtx.drawImage(backupCanvas, 0, 0, backupCanvas.width, backupCanvas.height, 0, 0, maskCanvas.width, maskCanvas.height); - }); - const filepath = ComfyApp.clipspace.images; - const touched_image = new Image(); - - touched_image.onload = function() { - backupCanvas.width = touched_image.width; - backupCanvas.height = touched_image.height; - - prepareRGB(touched_image, backupCanvas, backupCtx); - }; - const alpha_url = new URL(ComfyApp.clipspace.imgs[ComfyApp.clipspace['selectedIndex']].src) alpha_url.searchParams.delete('channel'); alpha_url.searchParams.delete('preview'); alpha_url.searchParams.set('channel', 'a'); - touched_image.src = alpha_url; + let mask_image = await loadImage(alpha_url); // original image load - orig_image.onload = function() { - window.dispatchEvent(new Event('resize')); - }; - const rgb_url = new URL(ComfyApp.clipspace.imgs[ComfyApp.clipspace['selectedIndex']].src); rgb_url.searchParams.delete('channel'); rgb_url.searchParams.set('channel', 'rgb'); - orig_image.src = rgb_url; - this.image = orig_image; + this.image = new Image(); + this.image.onload = function() { + maskCanvas.width = self.image.width; + maskCanvas.height = self.image.height; + + self.invalidateCanvas(self.image, mask_image); + self.initializeCanvasPanZoom(); + }; + this.image.src = rgb_url; } - setEventHandler(maskCanvas) { - maskCanvas.addEventListener("contextmenu", (event) => { - event.preventDefault(); - }); + initializeCanvasPanZoom() { + // set initialize + let drawWidth = this.image.width; + let drawHeight = this.image.height; + let width = this.element.clientWidth; + let height = this.element.clientHeight; + + if (this.image.width > width) { + drawWidth = width; + drawHeight = (drawWidth / this.image.width) * this.image.height; + } + + if (drawHeight > height) { + drawHeight = height; + drawWidth = (drawHeight / this.image.height) * this.image.width; + } + + this.zoom_ratio = drawWidth/this.image.width; + + const canvasX = (width - drawWidth) / 2; + const canvasY = (height - drawHeight) / 2; + this.pan_x = canvasX; + this.pan_y = canvasY; + + this.invalidatePanZoom(); + } + + + invalidatePanZoom() { + let raw_width = this.image.width * this.zoom_ratio; + let raw_height = this.image.height * this.zoom_ratio; + + if(this.pan_x + raw_width < 10) { + this.pan_x = 10 - raw_width; + } + + if(this.pan_y + raw_height < 10) { + this.pan_y = 10 - raw_height; + } + + let width = `${raw_width}px`; + let height = `${raw_height}px`; + + let left = `${this.pan_x}px`; + let top = `${this.pan_y}px`; + + this.maskCanvas.style.width = width; + this.maskCanvas.style.height = height; + this.maskCanvas.style.left = left; + this.maskCanvas.style.top = top; + + this.imgCanvas.style.width = width; + this.imgCanvas.style.height = height; + this.imgCanvas.style.left = left; + this.imgCanvas.style.top = top; + } + + + setEventHandler(maskCanvas) { const self = this; - maskCanvas.addEventListener('wheel', (event) => this.handleWheelEvent(self,event)); - maskCanvas.addEventListener('pointerdown', (event) => this.handlePointerDown(self,event)); - document.addEventListener('pointerup', MaskEditorDialog.handlePointerUp); - maskCanvas.addEventListener('pointermove', (event) => this.draw_move(self,event)); - maskCanvas.addEventListener('touchmove', (event) => this.draw_move(self,event)); - maskCanvas.addEventListener('pointerover', (event) => { this.brush.style.display = "block"; }); - maskCanvas.addEventListener('pointerleave', (event) => { this.brush.style.display = "none"; }); - document.addEventListener('keydown', MaskEditorDialog.handleKeyDown); + + if(!this.handler_registered) { + maskCanvas.addEventListener("contextmenu", (event) => { + event.preventDefault(); + }); + + this.element.addEventListener('wheel', (event) => this.handleWheelEvent(self,event)); + this.element.addEventListener('pointermove', (event) => this.pointMoveEvent(self,event)); + this.element.addEventListener('touchmove', (event) => this.pointMoveEvent(self,event)); + + this.element.addEventListener('dragstart', (event) => { + if(event.ctrlKey) { + event.preventDefault(); + } + }); + + maskCanvas.addEventListener('pointerdown', (event) => this.handlePointerDown(self,event)); + maskCanvas.addEventListener('pointermove', (event) => this.draw_move(self,event)); + maskCanvas.addEventListener('touchmove', (event) => this.draw_move(self,event)); + maskCanvas.addEventListener('pointerover', (event) => { this.brush.style.display = "block"; }); + maskCanvas.addEventListener('pointerleave', (event) => { this.brush.style.display = "none"; }); + + document.addEventListener('pointerup', MaskEditorDialog.handlePointerUp); + + this.handler_registered = true; + } } brush_size = 10; @@ -378,8 +449,10 @@ class MaskEditorDialog extends ComfyDialog { const self = MaskEditorDialog.instance; if (event.key === ']') { self.brush_size = Math.min(self.brush_size+2, 100); + self.brush_slider_input.value = self.brush_size; } else if (event.key === '[') { self.brush_size = Math.max(self.brush_size-2, 1); + self.brush_slider_input.value = self.brush_size; } else if(event.key === 'Enter') { self.save(); } @@ -389,6 +462,10 @@ class MaskEditorDialog extends ComfyDialog { static handlePointerUp(event) { event.preventDefault(); + + this.mousedown_x = null; + this.mousedown_y = null; + MaskEditorDialog.instance.drawing_mode = false; } @@ -398,24 +475,70 @@ class MaskEditorDialog extends ComfyDialog { var centerX = self.cursorX; var centerY = self.cursorY; - brush.style.width = self.brush_size * 2 + "px"; - brush.style.height = self.brush_size * 2 + "px"; - brush.style.left = (centerX - self.brush_size) + "px"; - brush.style.top = (centerY - self.brush_size) + "px"; + brush.style.width = self.brush_size * 2 * this.zoom_ratio + "px"; + brush.style.height = self.brush_size * 2 * this.zoom_ratio + "px"; + brush.style.left = (centerX - self.brush_size * this.zoom_ratio) + "px"; + brush.style.top = (centerY - self.brush_size * this.zoom_ratio) + "px"; } handleWheelEvent(self, event) { - if(event.deltaY < 0) - self.brush_size = Math.min(self.brush_size+2, 100); - else - self.brush_size = Math.max(self.brush_size-2, 1); + event.preventDefault(); - self.brush_slider_input.value = self.brush_size; + if(event.ctrlKey) { + // zoom canvas + if(event.deltaY < 0) { + this.zoom_ratio = Math.min(10.0, this.zoom_ratio+0.2); + } + else { + this.zoom_ratio = Math.max(0.2, this.zoom_ratio-0.2); + } + + this.invalidatePanZoom(); + } + else { + // adjust brush size + if(event.deltaY < 0) + this.brush_size = Math.min(this.brush_size+2, 100); + else + this.brush_size = Math.max(this.brush_size-2, 1); + + this.brush_slider_input.value = this.brush_size; + + this.updateBrushPreview(this); + } + } + + pointMoveEvent(self, event) { + this.cursorX = event.pageX; + this.cursorY = event.pageY; self.updateBrushPreview(self); + + if(event.ctrlKey) { + event.preventDefault(); + self.pan_move(self, event); + } + } + + pan_move(self, event) { + if(event.buttons == 1) { + if(this.mousedown_x) { + let deltaX = this.mousedown_x - event.clientX; + let deltaY = this.mousedown_y - event.clientY; + + self.pan_x = this.mousedown_pan_x - deltaX; + self.pan_y = this.mousedown_pan_y - deltaY; + + self.invalidatePanZoom(); + } + } } draw_move(self, event) { + if(event.ctrlKey) { + return; + } + event.preventDefault(); this.cursorX = event.pageX; @@ -439,6 +562,9 @@ class MaskEditorDialog extends ComfyDialog { y = event.targetTouches[0].clientY - maskRect.top; } + x /= self.zoom_ratio; + y /= self.zoom_ratio; + var brush_size = this.brush_size; if(event instanceof PointerEvent && event.pointerType == 'pen') { brush_size *= event.pressure; @@ -489,8 +615,8 @@ class MaskEditorDialog extends ComfyDialog { } else if(event.buttons == 2 || event.buttons == 5 || event.buttons == 32) { const maskRect = self.maskCanvas.getBoundingClientRect(); - const x = event.offsetX || event.targetTouches[0].clientX - maskRect.left; - const y = event.offsetY || event.targetTouches[0].clientY - maskRect.top; + 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; var brush_size = this.brush_size; if(event instanceof PointerEvent && event.pointerType == 'pen') { @@ -540,6 +666,17 @@ class MaskEditorDialog extends ComfyDialog { } handlePointerDown(self, event) { + if(event.ctrlKey) { + if (event.buttons == 1) { + this.mousedown_x = event.clientX; + this.mousedown_y = event.clientY; + + this.mousedown_pan_x = this.pan_x; + this.mousedown_pan_y = this.pan_y; + } + return; + } + var brush_size = this.brush_size; if(event instanceof PointerEvent && event.pointerType == 'pen') { brush_size *= event.pressure; @@ -551,8 +688,8 @@ class MaskEditorDialog extends ComfyDialog { event.preventDefault(); const maskRect = self.maskCanvas.getBoundingClientRect(); - const x = event.offsetX || event.targetTouches[0].clientX - maskRect.left; - const y = event.offsetY || event.targetTouches[0].clientY - maskRect.top; + 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) { @@ -570,15 +707,18 @@ class MaskEditorDialog extends ComfyDialog { } async save() { - const backupCtx = this.backupCanvas.getContext('2d', {willReadFrequently:true}); + const backupCanvas = document.createElement('canvas'); + const backupCtx = backupCanvas.getContext('2d', {willReadFrequently:true}); + backupCanvas.width = this.image.width; + backupCanvas.height = this.image.height; - backupCtx.clearRect(0,0,this.backupCanvas.width,this.backupCanvas.height); + backupCtx.clearRect(0,0, backupCanvas.width, backupCanvas.height); backupCtx.drawImage(this.maskCanvas, 0, 0, this.maskCanvas.width, this.maskCanvas.height, - 0, 0, this.backupCanvas.width, this.backupCanvas.height); + 0, 0, backupCanvas.width, backupCanvas.height); // paste mask data into alpha channel - const backupData = backupCtx.getImageData(0, 0, this.backupCanvas.width, this.backupCanvas.height); + const backupData = backupCtx.getImageData(0, 0, backupCanvas.width, backupCanvas.height); // refine mask image for (let i = 0; i < backupData.data.length; i += 4) { @@ -615,7 +755,7 @@ class MaskEditorDialog extends ComfyDialog { ComfyApp.clipspace.widgets[index].value = item; } - const dataURL = this.backupCanvas.toDataURL(); + const dataURL = backupCanvas.toDataURL(); const blob = dataURLToBlob(dataURL); let original_url = new URL(this.image.src); diff --git a/web/extensions/core/rerouteNode.js b/web/extensions/core/rerouteNode.js index 499a171da..4feff91e5 100644 --- a/web/extensions/core/rerouteNode.js +++ b/web/extensions/core/rerouteNode.js @@ -1,10 +1,11 @@ import { app } from "../../scripts/app.js"; +import { mergeIfValid, getWidgetConfig, setWidgetConfig } from "./widgetInputs.js"; // Node that allows you to redirect connections for cleaner graphs app.registerExtension({ name: "Comfy.RerouteNode", - registerCustomNodes() { + registerCustomNodes(app) { class RerouteNode { constructor() { if (!this.properties) { @@ -16,6 +17,12 @@ app.registerExtension({ this.addInput("", "*"); this.addOutput(this.properties.showOutputText ? "*" : "", "*"); + this.onAfterGraphConfigured = function () { + requestAnimationFrame(() => { + this.onConnectionsChange(LiteGraph.INPUT, null, true, null); + }); + }; + this.onConnectionsChange = function (type, index, connected, link_info) { this.applyOrientation(); @@ -47,6 +54,7 @@ app.registerExtension({ const linkId = currentNode.inputs[0].link; if (linkId !== null) { const link = app.graph.links[linkId]; + if (!link) return; const node = app.graph.getNodeById(link.origin_id); const type = node.constructor.type; if (type === "Reroute") { @@ -54,8 +62,7 @@ app.registerExtension({ // We've found a circle currentNode.disconnectInput(link.target_slot); currentNode = null; - } - else { + } else { // Move the previous node currentNode = node; } @@ -94,8 +101,11 @@ app.registerExtension({ updateNodes.push(node); } else { // We've found an output - const nodeOutType = node.inputs && node.inputs[link?.target_slot] && node.inputs[link.target_slot].type ? node.inputs[link.target_slot].type : null; - if (inputType && nodeOutType !== inputType) { + const nodeOutType = + node.inputs && node.inputs[link?.target_slot] && node.inputs[link.target_slot].type + ? node.inputs[link.target_slot].type + : null; + if (inputType && inputType !== "*" && nodeOutType !== inputType) { // The output doesnt match our input so disconnect it node.disconnectInput(link.target_slot); } else { @@ -111,6 +121,9 @@ app.registerExtension({ const displayType = inputType || outputType || "*"; const color = LGraphCanvas.link_type_colors[displayType]; + let widgetConfig; + let targetWidget; + let widgetType; // Update the types of each node for (const node of updateNodes) { // If we dont have an input type we are always wildcard but we'll show the output type @@ -125,10 +138,38 @@ app.registerExtension({ const link = app.graph.links[l]; if (link) { link.color = color; + + if (app.configuringGraph) continue; + const targetNode = app.graph.getNodeById(link.target_id); + const targetInput = targetNode.inputs?.[link.target_slot]; + if (targetInput?.widget) { + const config = getWidgetConfig(targetInput); + if (!widgetConfig) { + widgetConfig = config[1] ?? {}; + widgetType = config[0]; + } + if (!targetWidget) { + targetWidget = targetNode.widgets?.find((w) => w.name === targetInput.widget.name); + } + + const merged = mergeIfValid(targetInput, [config[0], widgetConfig]); + if (merged.customConfig) { + widgetConfig = merged.customConfig; + } + } } } } + for (const node of updateNodes) { + if (widgetConfig && outputType) { + node.inputs[0].widget = { name: "value" }; + setWidgetConfig(node.inputs[0], [widgetType ?? displayType, widgetConfig], targetWidget); + } else { + setWidgetConfig(node.inputs[0], null); + } + } + if (inputNode) { const link = app.graph.links[inputNode.inputs[0].link]; if (link) { @@ -173,8 +214,8 @@ app.registerExtension({ }, { // naming is inverted with respect to LiteGraphNode.horizontal - // LiteGraphNode.horizontal == true means that - // each slot in the inputs and outputs are layed out horizontally, + // LiteGraphNode.horizontal == true means that + // each slot in the inputs and outputs are layed out horizontally, // which is the opposite of the visual orientation of the inputs and outputs as a node content: "Set " + (this.properties.horizontal ? "Horizontal" : "Vertical"), callback: () => { @@ -187,7 +228,7 @@ app.registerExtension({ applyOrientation() { this.horizontal = this.properties.horizontal; if (this.horizontal) { - // we correct the input position, because LiteGraphNode.horizontal + // we correct the input position, because LiteGraphNode.horizontal // doesn't account for title presence // which reroute nodes don't have this.inputs[0].pos = [this.size[0] / 2, 0]; diff --git a/web/extensions/core/saveImageExtraOutput.js b/web/extensions/core/saveImageExtraOutput.js index 99e2213bf..a0506b43b 100644 --- a/web/extensions/core/saveImageExtraOutput.js +++ b/web/extensions/core/saveImageExtraOutput.js @@ -1,5 +1,5 @@ import { app } from "../../scripts/app.js"; - +import { applyTextReplacements } from "../../scripts/utils.js"; // Use widget values and dates in output filenames app.registerExtension({ @@ -7,84 +7,19 @@ app.registerExtension({ async beforeRegisterNodeDef(nodeType, nodeData, app) { if (nodeData.name === "SaveImage") { const onNodeCreated = nodeType.prototype.onNodeCreated; - - // Simple date formatter - const parts = { - d: (d) => d.getDate(), - M: (d) => d.getMonth() + 1, - h: (d) => d.getHours(), - m: (d) => d.getMinutes(), - s: (d) => d.getSeconds(), - }; - const format = - Object.keys(parts) - .map((k) => k + k + "?") - .join("|") + "|yyy?y?"; - - function formatDate(text, date) { - return text.replace(new RegExp(format, "g"), function (text) { - if (text === "yy") return (date.getFullYear() + "").substring(2); - if (text === "yyyy") return date.getFullYear(); - if (text[0] in parts) { - const p = parts[text[0]](date); - return (p + "").padStart(text.length, "0"); - } - return text; - }); - } - - // When the SaveImage node is created we want to override the serialization of the output name widget to run our S&R + // When the SaveImage node is created we want to override the serialization of the output name widget to run our S&R nodeType.prototype.onNodeCreated = function () { const r = onNodeCreated ? onNodeCreated.apply(this, arguments) : undefined; const widget = this.widgets.find((w) => w.name === "filename_prefix"); widget.serializeValue = () => { - return widget.value.replace(/%([^%]+)%/g, function (match, text) { - const split = text.split("."); - if (split.length !== 2) { - // Special handling for dates - if (split[0].startsWith("date:")) { - return formatDate(split[0].substring(5), new Date()); - } - - if (text !== "width" && text !== "height") { - // Dont warn on standard replacements - console.warn("Invalid replacement pattern", text); - } - return match; - } - - // Find node with matching S&R property name - let nodes = app.graph._nodes.filter((n) => n.properties?.["Node name for S&R"] === split[0]); - // If we cant, see if there is a node with that title - if (!nodes.length) { - nodes = app.graph._nodes.filter((n) => n.title === split[0]); - } - if (!nodes.length) { - console.warn("Unable to find node", split[0]); - return match; - } - - if (nodes.length > 1) { - console.warn("Multiple nodes matched", split[0], "using first match"); - } - - const node = nodes[0]; - - const widget = node.widgets?.find((w) => w.name === split[1]); - if (!widget) { - console.warn("Unable to find widget", split[1], "on node", split[0], node); - return match; - } - - return ((widget.value ?? "") + "").replaceAll(/\/|\\/g, "_"); - }); + return applyTextReplacements(app, widget.value); }; return r; }; } else { - // When any other node is created add a property to alias the node + // When any other node is created add a property to alias the node const onNodeCreated = nodeType.prototype.onNodeCreated; nodeType.prototype.onNodeCreated = function () { const r = onNodeCreated ? onNodeCreated.apply(this, arguments) : undefined; diff --git a/web/extensions/core/undoRedo.js b/web/extensions/core/undoRedo.js index c6613b0f0..3cb137520 100644 --- a/web/extensions/core/undoRedo.js +++ b/web/extensions/core/undoRedo.js @@ -71,24 +71,21 @@ function graphEqual(a, b, root = true) { } const undoRedo = async (e) => { + const updateState = async (source, target) => { + const prevState = source.pop(); + if (prevState) { + target.push(activeState); + isOurLoad = true; + await app.loadGraphData(prevState, false); + activeState = prevState; + } + } if (e.ctrlKey || e.metaKey) { if (e.key === "y") { - const prevState = redo.pop(); - if (prevState) { - undo.push(activeState); - isOurLoad = true; - await app.loadGraphData(prevState); - activeState = prevState; - } + updateState(redo, undo); return true; } else if (e.key === "z") { - const prevState = undo.pop(); - if (prevState) { - redo.push(activeState); - isOurLoad = true; - await app.loadGraphData(prevState); - activeState = prevState; - } + updateState(undo, redo); return true; } } diff --git a/web/extensions/core/widgetInputs.js b/web/extensions/core/widgetInputs.js index b6fa411f7..3f1c1f8c1 100644 --- a/web/extensions/core/widgetInputs.js +++ b/web/extensions/core/widgetInputs.js @@ -1,10 +1,16 @@ import { ComfyWidgets, addValueControlWidgets } from "../../scripts/widgets.js"; import { app } from "../../scripts/app.js"; +import { applyTextReplacements } from "../../scripts/utils.js"; const CONVERTED_TYPE = "converted-widget"; const VALID_TYPES = ["STRING", "combo", "number", "BOOLEAN"]; const CONFIG = Symbol(); const GET_CONFIG = Symbol(); +const TARGET = Symbol(); // Used for reroutes to specify the real target widget + +export function getWidgetConfig(slot) { + return slot.widget[CONFIG] ?? slot.widget[GET_CONFIG](); +} function getConfig(widgetName) { const { nodeData } = this.constructor; @@ -100,7 +106,6 @@ function getWidgetType(config) { return { type }; } - function isValidCombo(combo, obj) { // New input isnt a combo if (!(obj instanceof Array)) { @@ -121,6 +126,31 @@ function isValidCombo(combo, obj) { return true; } +export function setWidgetConfig(slot, config, target) { + if (!slot.widget) return; + if (config) { + slot.widget[GET_CONFIG] = () => config; + slot.widget[TARGET] = target; + } else { + delete slot.widget; + } + + if (slot.link) { + const link = app.graph.links[slot.link]; + if (link) { + const originNode = app.graph.getNodeById(link.origin_id); + if (originNode.type === "PrimitiveNode") { + if (config) { + originNode.recreateWidget(); + } else if(!app.configuringGraph) { + originNode.disconnectOutput(0); + originNode.onLastDisconnect(); + } + } + } + } +} + export function mergeIfValid(output, config2, forceUpdate, recreateWidget, config1) { if (!config1) { config1 = output.widget[CONFIG] ?? output.widget[GET_CONFIG](); @@ -150,7 +180,7 @@ export function mergeIfValid(output, config2, forceUpdate, recreateWidget, confi const isNumber = config1[0] === "INT" || config1[0] === "FLOAT"; for (const k of keys.values()) { - if (k !== "default" && k !== "forceInput" && k !== "defaultInput") { + if (k !== "default" && k !== "forceInput" && k !== "defaultInput" && k !== "control_after_generate" && k !== "multiline") { let v1 = config1[1][k]; let v2 = config2[1]?.[k]; @@ -405,11 +435,16 @@ app.registerExtension({ }; }, registerCustomNodes() { + const replacePropertyName = "Run widget replace on values"; class PrimitiveNode { constructor() { this.addOutput("connect to widget input", "*"); this.serialize_widgets = true; this.isVirtualNode = true; + + if (!this.properties || !(replacePropertyName in this.properties)) { + this.addProperty(replacePropertyName, false, "boolean"); + } } applyToGraph(extraLinks = []) { @@ -430,18 +465,29 @@ app.registerExtension({ } let links = [...get_links(this).map((l) => app.graph.links[l]), ...extraLinks]; + let v = this.widgets?.[0].value; + if(v && this.properties[replacePropertyName]) { + v = applyTextReplacements(app, v); + } + // For each output link copy our value over the original widget value for (const linkInfo of links) { const node = this.graph.getNodeById(linkInfo.target_id); const input = node.inputs[linkInfo.target_slot]; - const widgetName = input.widget.name; - if (widgetName) { - const widget = node.widgets.find((w) => w.name === widgetName); - if (widget) { - widget.value = this.widgets[0].value; - if (widget.callback) { - widget.callback(widget.value, app.canvas, node, app.canvas.graph_mouse, {}); - } + let widget; + if (input.widget[TARGET]) { + widget = input.widget[TARGET]; + } else { + const widgetName = input.widget.name; + if (widgetName) { + widget = node.widgets.find((w) => w.name === widgetName); + } + } + + if (widget) { + widget.value = v; + if (widget.callback) { + widget.callback(widget.value, app.canvas, node, app.canvas.graph_mouse, {}); } } } @@ -494,14 +540,13 @@ app.registerExtension({ this.#mergeWidgetConfig(); if (!links?.length) { - this.#onLastDisconnect(); + this.onLastDisconnect(); } } } onConnectOutput(slot, type, input, target_node, target_slot) { // Fires before the link is made allowing us to reject it if it isn't valid - // No widget, we cant connect if (!input.widget) { if (!(input.type in ComfyWidgets)) return false; @@ -519,6 +564,10 @@ app.registerExtension({ #onFirstConnection(recreating) { // First connection can fire before the graph is ready on initial load so random things can be missing + if (!this.outputs[0].links) { + this.onLastDisconnect(); + return; + } const linkId = this.outputs[0].links[0]; const link = this.graph.links[linkId]; if (!link) return; @@ -546,10 +595,10 @@ app.registerExtension({ this.outputs[0].name = type; this.outputs[0].widget = widget; - this.#createWidget(widget[CONFIG] ?? config, theirNode, widget.name, recreating); + this.#createWidget(widget[CONFIG] ?? config, theirNode, widget.name, recreating, widget[TARGET]); } - #createWidget(inputData, node, widgetName, recreating) { + #createWidget(inputData, node, widgetName, recreating, targetWidget) { let type = inputData[0]; if (type instanceof Array) { @@ -563,7 +612,9 @@ app.registerExtension({ widget = this.addWidget(type, "value", null, () => {}, {}); } - if (node?.widgets && widget) { + if (targetWidget) { + widget.value = targetWidget.value; + } else if (node?.widgets && widget) { const theirWidget = node.widgets.find((w) => w.name === widgetName); if (theirWidget) { widget.value = theirWidget.value; @@ -577,11 +628,19 @@ app.registerExtension({ } addValueControlWidgets(this, widget, control_value, undefined, inputData); let filter = this.widgets_values?.[2]; - if(filter && this.widgets.length === 3) { + if (filter && this.widgets.length === 3) { this.widgets[2].value = filter; } } + // Restore any saved control values + const controlValues = this.controlValues; + if(this.lastType === this.widgets[0].type && controlValues?.length === this.widgets.length - 1) { + for(let i = 0; i < controlValues.length; i++) { + this.widgets[i + 1].value = controlValues[i]; + } + } + // When our value changes, update other widgets to reflect our changes // e.g. so LoadImage shows correct image const callback = widget.callback; @@ -610,12 +669,14 @@ app.registerExtension({ } } - #recreateWidget() { - const values = this.widgets.map((w) => w.value); + recreateWidget() { + const values = this.widgets?.map((w) => w.value); this.#removeWidgets(); this.#onFirstConnection(true); - for (let i = 0; i < this.widgets?.length; i++) this.widgets[i].value = values[i]; - return this.widgets[0]; + if (values?.length) { + for (let i = 0; i < this.widgets?.length; i++) this.widgets[i].value = values[i]; + } + return this.widgets?.[0]; } #mergeWidgetConfig() { @@ -631,7 +692,7 @@ app.registerExtension({ if (links?.length < 2 && hasConfig) { // Copy the widget options from the source if (links.length) { - this.#recreateWidget(); + this.recreateWidget(); } return; @@ -657,7 +718,7 @@ app.registerExtension({ // Only allow connections where the configs match const output = this.outputs[0]; const config2 = input.widget[GET_CONFIG](); - return !!mergeIfValid.call(this, output, config2, forceUpdate, this.#recreateWidget); + return !!mergeIfValid.call(this, output, config2, forceUpdate, this.recreateWidget); } #removeWidgets() { @@ -668,11 +729,20 @@ app.registerExtension({ w.onRemove(); } } + + // Temporarily store the current values in case the node is being recreated + // e.g. by group node conversion + this.controlValues = []; + this.lastType = this.widgets[0]?.type; + for(let i = 1; i < this.widgets.length; i++) { + this.controlValues.push(this.widgets[i].value); + } + setTimeout(() => { delete this.lastType; delete this.controlValues }, 15); this.widgets.length = 0; } } - #onLastDisconnect() { + onLastDisconnect() { // We cant remove + re-add the output here as if you drag a link over the same link // it removes, then re-adds, causing it to break this.outputs[0].type = "*"; diff --git a/web/scripts/app.js b/web/scripts/app.js index 5faf41fb3..62169abfb 100644 --- a/web/scripts/app.js +++ b/web/scripts/app.js @@ -1559,9 +1559,12 @@ export class ComfyApp { /** * Populates the graph with the specified workflow data * @param {*} graphData A serialized graph object + * @param { boolean } clean If the graph state, e.g. images, should be cleared */ - async loadGraphData(graphData) { - this.clean(); + async loadGraphData(graphData, clean = true) { + if (clean !== false) { + this.clean(); + } let reset_invalid_values = false; if (!graphData) { @@ -1771,7 +1774,9 @@ export class ComfyApp { if (parent?.updateLink) { link = parent.updateLink(link); } - inputs[node.inputs[i].name] = [String(link.origin_id), parseInt(link.origin_slot)]; + if (link) { + inputs[node.inputs[i].name] = [String(link.origin_id), parseInt(link.origin_slot)]; + } } } } diff --git a/web/scripts/utils.js b/web/scripts/utils.js new file mode 100644 index 000000000..401aca9e4 --- /dev/null +++ b/web/scripts/utils.js @@ -0,0 +1,67 @@ +// Simple date formatter +const parts = { + d: (d) => d.getDate(), + M: (d) => d.getMonth() + 1, + h: (d) => d.getHours(), + m: (d) => d.getMinutes(), + s: (d) => d.getSeconds(), +}; +const format = + Object.keys(parts) + .map((k) => k + k + "?") + .join("|") + "|yyy?y?"; + +function formatDate(text, date) { + return text.replace(new RegExp(format, "g"), function (text) { + if (text === "yy") return (date.getFullYear() + "").substring(2); + if (text === "yyyy") return date.getFullYear(); + if (text[0] in parts) { + const p = parts[text[0]](date); + return (p + "").padStart(text.length, "0"); + } + return text; + }); +} + +export function applyTextReplacements(app, value) { + return value.replace(/%([^%]+)%/g, function (match, text) { + const split = text.split("."); + if (split.length !== 2) { + // Special handling for dates + if (split[0].startsWith("date:")) { + return formatDate(split[0].substring(5), new Date()); + } + + if (text !== "width" && text !== "height") { + // Dont warn on standard replacements + console.warn("Invalid replacement pattern", text); + } + return match; + } + + // Find node with matching S&R property name + let nodes = app.graph._nodes.filter((n) => n.properties?.["Node name for S&R"] === split[0]); + // If we cant, see if there is a node with that title + if (!nodes.length) { + nodes = app.graph._nodes.filter((n) => n.title === split[0]); + } + if (!nodes.length) { + console.warn("Unable to find node", split[0]); + return match; + } + + if (nodes.length > 1) { + console.warn("Multiple nodes matched", split[0], "using first match"); + } + + const node = nodes[0]; + + const widget = node.widgets?.find((w) => w.name === split[1]); + if (!widget) { + console.warn("Unable to find widget", split[1], "on node", split[0], node); + return match; + } + + return ((widget.value ?? "") + "").replaceAll(/\/|\\/g, "_"); + }); +}