From cb9ac3db586b675a96ebd2604bf099bc189c1b28 Mon Sep 17 00:00:00 2001 From: Andrew Kvochko Date: Mon, 28 Apr 2025 19:59:17 +0300 Subject: [PATCH 01/22] ltxv: add strength parameter to conditioning. (#7849) This commit adds strength parameter to the LTXVImgToVideo node. --- comfy_extras/nodes_lt.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/comfy_extras/nodes_lt.py b/comfy_extras/nodes_lt.py index ff3fe5cdc..1a667e01a 100644 --- a/comfy_extras/nodes_lt.py +++ b/comfy_extras/nodes_lt.py @@ -38,6 +38,7 @@ class LTXVImgToVideo: "height": ("INT", {"default": 512, "min": 64, "max": nodes.MAX_RESOLUTION, "step": 32}), "length": ("INT", {"default": 97, "min": 9, "max": nodes.MAX_RESOLUTION, "step": 8}), "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), + "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0}), }} RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT") @@ -46,7 +47,7 @@ class LTXVImgToVideo: CATEGORY = "conditioning/video_models" FUNCTION = "generate" - def generate(self, positive, negative, image, vae, width, height, length, batch_size): + def generate(self, positive, negative, image, vae, width, height, length, batch_size, strength): pixels = comfy.utils.common_upscale(image.movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) encode_pixels = pixels[:, :, :, :3] t = vae.encode(encode_pixels) @@ -59,7 +60,7 @@ class LTXVImgToVideo: dtype=torch.float32, device=latent.device, ) - conditioning_latent_frames_mask[:, :, :t.shape[2]] = 0 + conditioning_latent_frames_mask[:, :, :t.shape[2]] = 1.0 - strength return (positive, negative, {"samples": latent, "noise_mask": conditioning_latent_frames_mask}, ) From 30159a7fe6bfa54fc1b9ba4c80b9041837038819 Mon Sep 17 00:00:00 2001 From: Pam <42671363+pamparamm@users.noreply.github.com> Date: Mon, 28 Apr 2025 22:03:21 +0500 Subject: [PATCH 02/22] Save v pred zsnr metadata (#7840) --- comfy/model_sampling.py | 3 ++- comfy_extras/nodes_model_merging.py | 3 +++ 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/comfy/model_sampling.py b/comfy/model_sampling.py index b79af1e92..7e7291476 100644 --- a/comfy/model_sampling.py +++ b/comfy/model_sampling.py @@ -111,13 +111,14 @@ class ModelSamplingDiscrete(torch.nn.Module): self.num_timesteps = int(timesteps) self.linear_start = linear_start self.linear_end = linear_end + self.zsnr = zsnr # self.register_buffer('betas', torch.tensor(betas, dtype=torch.float32)) # self.register_buffer('alphas_cumprod', torch.tensor(alphas_cumprod, dtype=torch.float32)) # self.register_buffer('alphas_cumprod_prev', torch.tensor(alphas_cumprod_prev, dtype=torch.float32)) sigmas = ((1 - alphas_cumprod) / alphas_cumprod) ** 0.5 - if zsnr: + if self.zsnr: sigmas = rescale_zero_terminal_snr_sigmas(sigmas) self.set_sigmas(sigmas) diff --git a/comfy_extras/nodes_model_merging.py b/comfy_extras/nodes_model_merging.py index ccf601158..78d284889 100644 --- a/comfy_extras/nodes_model_merging.py +++ b/comfy_extras/nodes_model_merging.py @@ -209,6 +209,9 @@ def save_checkpoint(model, clip=None, vae=None, clip_vision=None, filename_prefi metadata["modelspec.predict_key"] = "epsilon" elif model.model.model_type == comfy.model_base.ModelType.V_PREDICTION: metadata["modelspec.predict_key"] = "v" + extra_keys["v_pred"] = torch.tensor([]) + if getattr(model_sampling, "zsnr", False): + extra_keys["ztsnr"] = torch.tensor([]) if not args.disable_metadata: metadata["prompt"] = prompt_info From 5a50c3c7e59a867ddf0b3b7cc207c877a7b422fb Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 28 Apr 2025 10:07:21 -0700 Subject: [PATCH 03/22] Fix stream priority to support older pytorch. (#7856) --- comfy/model_management.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 516b6e2f1..78317af3c 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -962,7 +962,7 @@ def get_offload_stream(device): elif is_device_cuda(device): ss = [] for k in range(NUM_STREAMS): - ss.append(torch.cuda.Stream(device=device, priority=10)) + ss.append(torch.cuda.Stream(device=device, priority=0)) STREAMS[device] = ss s = ss[stream_counter] stream_counter = (stream_counter + 1) % len(ss) From 772b4c594549fb42f70833053be8613d79932965 Mon Sep 17 00:00:00 2001 From: Andrew Kvochko Date: Mon, 28 Apr 2025 20:42:04 +0300 Subject: [PATCH 04/22] ltxv: overwrite existing mask on conditioned frame. (#7845) This commit overwrites the noise mask on the latent frame that is being conditioned with keyframe conditioning, setting it to one. --- comfy_extras/nodes_lt.py | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/comfy_extras/nodes_lt.py b/comfy_extras/nodes_lt.py index 1a667e01a..e6dc122ca 100644 --- a/comfy_extras/nodes_lt.py +++ b/comfy_extras/nodes_lt.py @@ -153,6 +153,15 @@ class LTXVAddGuide: return node_helpers.conditioning_set_values(cond, {"keyframe_idxs": keyframe_idxs}) def append_keyframe(self, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors): + _, latent_idx = self.get_latent_index( + cond=positive, + latent_length=latent_image.shape[2], + guide_length=guiding_latent.shape[2], + frame_idx=frame_idx, + scale_factors=scale_factors, + ) + noise_mask[:, :, latent_idx:latent_idx + guiding_latent.shape[2]] = 1.0 + positive = self.add_keyframe_index(positive, frame_idx, guiding_latent, scale_factors) negative = self.add_keyframe_index(negative, frame_idx, guiding_latent, scale_factors) From c15909bb621b31b0ede693f30cc927021e473c72 Mon Sep 17 00:00:00 2001 From: chaObserv <154517000+chaObserv@users.noreply.github.com> Date: Tue, 29 Apr 2025 01:51:35 +0800 Subject: [PATCH 05/22] CFG++ for gradient estimation sampler (#7809) --- comfy/k_diffusion/sampling.py | 34 +++++++++++++++++++++++++++++----- comfy/samplers.py | 2 +- 2 files changed, 30 insertions(+), 6 deletions(-) diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index 6388d3faf..77ef748e8 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -1345,28 +1345,52 @@ def sample_res_multistep_ancestral_cfg_pp(model, x, sigmas, extra_args=None, cal return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=True) @torch.no_grad() -def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.): +def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2., cfg_pp=False): """Gradient-estimation sampler. Paper: https://openreview.net/pdf?id=o2ND9v0CeK""" extra_args = {} if extra_args is None else extra_args s_in = x.new_ones([x.shape[0]]) old_d = None + uncond_denoised = None + def post_cfg_function(args): + nonlocal uncond_denoised + uncond_denoised = args["uncond_denoised"] + return args["denoised"] + + if cfg_pp: + model_options = extra_args.get("model_options", {}).copy() + extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True) + for i in trange(len(sigmas) - 1, disable=disable): denoised = model(x, sigmas[i] * s_in, **extra_args) - d = to_d(x, sigmas[i], denoised) + if cfg_pp: + d = to_d(x, sigmas[i], uncond_denoised) + else: + d = to_d(x, sigmas[i], denoised) if callback is not None: callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) dt = sigmas[i + 1] - sigmas[i] if i == 0: # Euler method - x = x + d * dt + if cfg_pp: + x = denoised + d * sigmas[i + 1] + else: + x = x + d * dt else: # Gradient estimation - d_bar = ge_gamma * d + (1 - ge_gamma) * old_d - x = x + d_bar * dt + if cfg_pp: + d_bar = (ge_gamma - 1) * (d - old_d) + x = denoised + d * sigmas[i + 1] + d_bar * dt + else: + d_bar = ge_gamma * d + (1 - ge_gamma) * old_d + x = x + d_bar * dt old_d = d return x +@torch.no_grad() +def sample_gradient_estimation_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.): + return sample_gradient_estimation(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, ge_gamma=ge_gamma, cfg_pp=True) + @torch.no_grad() def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None, noise_scaler=None, max_stage=3): """ diff --git a/comfy/samplers.py b/comfy/samplers.py index 27dfce45a..67ae09a25 100644 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -710,7 +710,7 @@ KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_c "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu", "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm", "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp", - "gradient_estimation", "er_sde", "seeds_2", "seeds_3"] + "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3"] class KSAMPLER(Sampler): def __init__(self, sampler_function, extra_options={}, inpaint_options={}): From 7d329771f9f966fec20aa72079fc39202d877cbd Mon Sep 17 00:00:00 2001 From: Yoland Yan <4950057+yoland68@users.noreply.github.com> Date: Mon, 28 Apr 2025 10:59:22 -0700 Subject: [PATCH 06/22] Add moderation level option to OpenAIGPTImage1 node and update api_call method signature (#7804) --- comfy_api_nodes/nodes_api.py | 27 +++++++++++++++++---------- 1 file changed, 17 insertions(+), 10 deletions(-) diff --git a/comfy_api_nodes/nodes_api.py b/comfy_api_nodes/nodes_api.py index 4105ba7e1..a977bb9b7 100644 --- a/comfy_api_nodes/nodes_api.py +++ b/comfy_api_nodes/nodes_api.py @@ -1,21 +1,22 @@ +import base64 import io +import math from inspect import cleandoc -from comfy.utils import common_upscale +import numpy as np +import requests +import torch +from PIL import Image + from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeDict +from comfy.utils import common_upscale from comfy_api_nodes.apis import ( - OpenAIImageGenerationRequest, OpenAIImageEditRequest, - OpenAIImageGenerationResponse + OpenAIImageGenerationRequest, + OpenAIImageGenerationResponse, ) from comfy_api_nodes.apis.client import ApiEndpoint, HttpMethod, SynchronousOperation -import numpy as np -from PIL import Image -import requests -import torch -import math -import base64 def downscale_input(image): samples = image.movedim(-1,1) @@ -331,6 +332,11 @@ class OpenAIGPTImage1(ComfyNodeABC): "default": None, "tooltip": "Optional mask for inpainting (white areas will be replaced)", }), + "moderation": (IO.COMBO, { + "options": ["low","auto"], + "default": "low", + "tooltip": "Moderation level", + }), }, "hidden": { "auth_token": "AUTH_TOKEN_COMFY_ORG" @@ -343,7 +349,7 @@ class OpenAIGPTImage1(ComfyNodeABC): DESCRIPTION = cleandoc(__doc__ or "") API_NODE = True - def api_call(self, prompt, seed=0, quality="low", background="opaque", image=None, mask=None, n=1, size="1024x1024", auth_token=None): + def api_call(self, prompt, seed=0, quality="low", background="opaque", image=None, mask=None, n=1, size="1024x1024", auth_token=None, moderation="low"): model = "gpt-image-1" path = "/proxy/openai/images/generations" request_class = OpenAIImageGenerationRequest @@ -415,6 +421,7 @@ class OpenAIGPTImage1(ComfyNodeABC): n=n, seed=seed, size=size, + moderation=moderation, ), files=files if files else None, auth_token=auth_token From 83d04717b6cd84fca7af31ee655f7e5a3585a371 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 28 Apr 2025 12:01:15 -0700 Subject: [PATCH 07/22] Support HiDream E1 model. (#7857) --- comfy/ldm/hidream/model.py | 3 +++ comfy/model_base.py | 3 +++ 2 files changed, 6 insertions(+) diff --git a/comfy/ldm/hidream/model.py b/comfy/ldm/hidream/model.py index fcb5a9c51..0305747bf 100644 --- a/comfy/ldm/hidream/model.py +++ b/comfy/ldm/hidream/model.py @@ -699,10 +699,13 @@ class HiDreamImageTransformer2DModel(nn.Module): y: Optional[torch.Tensor] = None, context: Optional[torch.Tensor] = None, encoder_hidden_states_llama3=None, + image_cond=None, control = None, transformer_options = {}, ) -> torch.Tensor: bs, c, h, w = x.shape + if image_cond is not None: + x = torch.cat([x, image_cond], dim=-1) hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size)) timesteps = t pooled_embeds = y diff --git a/comfy/model_base.py b/comfy/model_base.py index b0c6a465b..d2aa4ce7a 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1104,4 +1104,7 @@ class HiDream(BaseModel): conditioning_llama3 = kwargs.get("conditioning_llama3", None) if conditioning_llama3 is not None: out['encoder_hidden_states_llama3'] = comfy.conds.CONDRegular(conditioning_llama3) + image_cond = kwargs.get("concat_latent_image", None) + if image_cond is not None: + out['image_cond'] = comfy.conds.CONDNoiseShape(self.process_latent_in(image_cond)) return out From 68f0d3529667a2b34b27cc0ac5051bc0e8c45b49 Mon Sep 17 00:00:00 2001 From: guill Date: Tue, 29 Apr 2025 02:58:00 -0700 Subject: [PATCH 08/22] Add support for VIDEO as a built-in type (#7844) * Add basic support for videos as types This PR adds support for VIDEO as first-class types. In order to avoid unnecessary costs, VIDEO outputs must implement the `VideoInput` ABC, but their implementation details can vary. Included are two implementations of this type which can be returned by other nodes: * `VideoFromFile` - Created with either a path on disk (as a string) or a `io.BytesIO` containing the contents of a file in a supported format (like .mp4). This implementation won't actually load the video unless necessary. It will also avoid re-encoding when saving if possible. * `VideoFromComponents` - Created from an image tensor and an optional audio tensor. Currently, only h264 encoded videos in .mp4 containers are supported for saving, but the plan is to add additional encodings/containers in the near future (particularly .webm). * Add optimization to avoid parsing entire video * Improve type declarations to reduce warnings * Make sure bytesIO objects can be read many times * Fix a potential issue when saving long videos * Fix incorrect type annotation * Add a `LoadVideo` node to make testing easier * Refactor new types out of the base comfy folder I've created a new `comfy_api` top-level module. The intention is that anything within this folder would be covered by semver-style versioning that would allow custom nodes to rely on them not introducing breaking changes. * Fix linting issue --- comfy/comfy_types/node_typing.py | 9 +- comfy_api/input/__init__.py | 8 + comfy_api/input/basic_types.py | 20 +++ comfy_api/input/video_types.py | 45 ++++++ comfy_api/input_impl/__init__.py | 7 + comfy_api/input_impl/video_types.py | 224 ++++++++++++++++++++++++++++ comfy_api/util/__init__.py | 8 + comfy_api/util/video_types.py | 51 +++++++ comfy_extras/nodes_video.py | 164 +++++++++++++++++++- folder_paths.py | 4 +- 10 files changed, 532 insertions(+), 8 deletions(-) create mode 100644 comfy_api/input/__init__.py create mode 100644 comfy_api/input/basic_types.py create mode 100644 comfy_api/input/video_types.py create mode 100644 comfy_api/input_impl/__init__.py create mode 100644 comfy_api/input_impl/video_types.py create mode 100644 comfy_api/util/__init__.py create mode 100644 comfy_api/util/video_types.py diff --git a/comfy/comfy_types/node_typing.py b/comfy/comfy_types/node_typing.py index 4ceeb3468..2ffc9c021 100644 --- a/comfy/comfy_types/node_typing.py +++ b/comfy/comfy_types/node_typing.py @@ -48,6 +48,7 @@ class IO(StrEnum): FACE_ANALYSIS = "FACE_ANALYSIS" BBOX = "BBOX" SEGS = "SEGS" + VIDEO = "VIDEO" ANY = "*" """Always matches any type, but at a price. @@ -273,7 +274,7 @@ class ComfyNodeABC(ABC): Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing """ - OUTPUT_IS_LIST: tuple[bool] + OUTPUT_IS_LIST: tuple[bool, ...] """A tuple indicating which node outputs are lists, but will be connected to nodes that expect individual items. Connected nodes that do not implement `INPUT_IS_LIST` will be executed once for every item in the list. @@ -292,7 +293,7 @@ class ComfyNodeABC(ABC): Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing """ - RETURN_TYPES: tuple[IO] + RETURN_TYPES: tuple[IO, ...] """A tuple representing the outputs of this node. Usage:: @@ -301,12 +302,12 @@ class ComfyNodeABC(ABC): Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-types """ - RETURN_NAMES: tuple[str] + RETURN_NAMES: tuple[str, ...] """The output slot names for each item in `RETURN_TYPES`, e.g. ``RETURN_NAMES = ("count", "filter_string")`` Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-names """ - OUTPUT_TOOLTIPS: tuple[str] + OUTPUT_TOOLTIPS: tuple[str, ...] """A tuple of strings to use as tooltips for node outputs, one for each item in `RETURN_TYPES`.""" FUNCTION: str """The name of the function to execute as a literal string, e.g. `FUNCTION = "execute"` diff --git a/comfy_api/input/__init__.py b/comfy_api/input/__init__.py new file mode 100644 index 000000000..66667946f --- /dev/null +++ b/comfy_api/input/__init__.py @@ -0,0 +1,8 @@ +from .basic_types import ImageInput, AudioInput +from .video_types import VideoInput + +__all__ = [ + "ImageInput", + "AudioInput", + "VideoInput", +] diff --git a/comfy_api/input/basic_types.py b/comfy_api/input/basic_types.py new file mode 100644 index 000000000..033fb7e27 --- /dev/null +++ b/comfy_api/input/basic_types.py @@ -0,0 +1,20 @@ +import torch +from typing import TypedDict + +ImageInput = torch.Tensor +""" +An image in format [B, H, W, C] where B is the batch size, C is the number of channels, +""" + +class AudioInput(TypedDict): + """ + TypedDict representing audio input. + """ + + waveform: torch.Tensor + """ + Tensor in the format [B, C, T] where B is the batch size, C is the number of channels, + """ + + sample_rate: int + diff --git a/comfy_api/input/video_types.py b/comfy_api/input/video_types.py new file mode 100644 index 000000000..0676e0e66 --- /dev/null +++ b/comfy_api/input/video_types.py @@ -0,0 +1,45 @@ +from __future__ import annotations +from abc import ABC, abstractmethod +from typing import Optional +from comfy_api.util import VideoContainer, VideoCodec, VideoComponents + +class VideoInput(ABC): + """ + Abstract base class for video input types. + """ + + @abstractmethod + def get_components(self) -> VideoComponents: + """ + Abstract method to get the video components (images, audio, and frame rate). + + Returns: + VideoComponents containing images, audio, and frame rate + """ + pass + + @abstractmethod + def save_to( + self, + path: str, + format: VideoContainer = VideoContainer.AUTO, + codec: VideoCodec = VideoCodec.AUTO, + metadata: Optional[dict] = None + ): + """ + Abstract method to save the video input to a file. + """ + pass + + # Provide a default implementation, but subclasses can provide optimized versions + # if possible. + def get_dimensions(self) -> tuple[int, int]: + """ + Returns the dimensions of the video input. + + Returns: + Tuple of (width, height) + """ + components = self.get_components() + return components.images.shape[2], components.images.shape[1] + diff --git a/comfy_api/input_impl/__init__.py b/comfy_api/input_impl/__init__.py new file mode 100644 index 000000000..02901b8b9 --- /dev/null +++ b/comfy_api/input_impl/__init__.py @@ -0,0 +1,7 @@ +from .video_types import VideoFromFile, VideoFromComponents + +__all__ = [ + # Implementations + "VideoFromFile", + "VideoFromComponents", +] diff --git a/comfy_api/input_impl/video_types.py b/comfy_api/input_impl/video_types.py new file mode 100644 index 000000000..12e5783db --- /dev/null +++ b/comfy_api/input_impl/video_types.py @@ -0,0 +1,224 @@ +from __future__ import annotations +from av.container import InputContainer +from av.subtitles.stream import SubtitleStream +from fractions import Fraction +from typing import Optional +from comfy_api.input import AudioInput +import av +import io +import json +import numpy as np +import torch +from comfy_api.input import VideoInput +from comfy_api.util import VideoContainer, VideoCodec, VideoComponents + +class VideoFromFile(VideoInput): + """ + Class representing video input from a file. + """ + + def __init__(self, file: str | io.BytesIO): + """ + Initialize the VideoFromFile object based off of either a path on disk or a BytesIO object + containing the file contents. + """ + self.__file = file + + def get_dimensions(self) -> tuple[int, int]: + """ + Returns the dimensions of the video input. + + Returns: + Tuple of (width, height) + """ + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) # Reset the BytesIO object to the beginning + with av.open(self.__file, mode='r') as container: + for stream in container.streams: + if stream.type == 'video': + assert isinstance(stream, av.VideoStream) + return stream.width, stream.height + raise ValueError(f"No video stream found in file '{self.__file}'") + + def get_components_internal(self, container: InputContainer) -> VideoComponents: + # Get video frames + frames = [] + for frame in container.decode(video=0): + img = frame.to_ndarray(format='rgb24') # shape: (H, W, 3) + img = torch.from_numpy(img) / 255.0 # shape: (H, W, 3) + frames.append(img) + + images = torch.stack(frames) if len(frames) > 0 else torch.zeros(0, 3, 0, 0) + + # Get frame rate + video_stream = next(s for s in container.streams if s.type == 'video') + frame_rate = Fraction(video_stream.average_rate) if video_stream and video_stream.average_rate else Fraction(1) + + # Get audio if available + audio = None + try: + container.seek(0) # Reset the container to the beginning + for stream in container.streams: + if stream.type != 'audio': + continue + assert isinstance(stream, av.AudioStream) + audio_frames = [] + for packet in container.demux(stream): + for frame in packet.decode(): + assert isinstance(frame, av.AudioFrame) + audio_frames.append(frame.to_ndarray()) # shape: (channels, samples) + if len(audio_frames) > 0: + audio_data = np.concatenate(audio_frames, axis=1) # shape: (channels, total_samples) + audio_tensor = torch.from_numpy(audio_data).unsqueeze(0) # shape: (1, channels, total_samples) + audio = AudioInput({ + "waveform": audio_tensor, + "sample_rate": int(stream.sample_rate) if stream.sample_rate else 1, + }) + except StopIteration: + pass # No audio stream + + metadata = container.metadata + return VideoComponents(images=images, audio=audio, frame_rate=frame_rate, metadata=metadata) + + def get_components(self) -> VideoComponents: + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) # Reset the BytesIO object to the beginning + with av.open(self.__file, mode='r') as container: + return self.get_components_internal(container) + raise ValueError(f"No video stream found in file '{self.__file}'") + + def save_to( + self, + path: str, + format: VideoContainer = VideoContainer.AUTO, + codec: VideoCodec = VideoCodec.AUTO, + metadata: Optional[dict] = None + ): + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) # Reset the BytesIO object to the beginning + with av.open(self.__file, mode='r') as container: + container_format = container.format.name + video_encoding = container.streams.video[0].codec.name if len(container.streams.video) > 0 else None + reuse_streams = True + if format != VideoContainer.AUTO and format not in container_format.split(","): + reuse_streams = False + if codec != VideoCodec.AUTO and codec != video_encoding and video_encoding is not None: + reuse_streams = False + + if not reuse_streams: + components = self.get_components_internal(container) + video = VideoFromComponents(components) + return video.save_to( + path, + format=format, + codec=codec, + metadata=metadata + ) + + streams = container.streams + with av.open(path, mode='w', options={"movflags": "use_metadata_tags"}) as output_container: + # Copy over the original metadata + for key, value in container.metadata.items(): + if metadata is None or key not in metadata: + output_container.metadata[key] = value + + # Add our new metadata + if metadata is not None: + for key, value in metadata.items(): + if isinstance(value, str): + output_container.metadata[key] = value + else: + output_container.metadata[key] = json.dumps(value) + + # Add streams to the new container + stream_map = {} + for stream in streams: + if isinstance(stream, (av.VideoStream, av.AudioStream, SubtitleStream)): + out_stream = output_container.add_stream_from_template(template=stream, opaque=True) + stream_map[stream] = out_stream + + # Write packets to the new container + for packet in container.demux(): + if packet.stream in stream_map and packet.dts is not None: + packet.stream = stream_map[packet.stream] + output_container.mux(packet) + +class VideoFromComponents(VideoInput): + """ + Class representing video input from tensors. + """ + + def __init__(self, components: VideoComponents): + self.__components = components + + def get_components(self) -> VideoComponents: + return VideoComponents( + images=self.__components.images, + audio=self.__components.audio, + frame_rate=self.__components.frame_rate + ) + + def save_to( + self, + path: str, + format: VideoContainer = VideoContainer.AUTO, + codec: VideoCodec = VideoCodec.AUTO, + metadata: Optional[dict] = None + ): + if format != VideoContainer.AUTO and format != VideoContainer.MP4: + raise ValueError("Only MP4 format is supported for now") + if codec != VideoCodec.AUTO and codec != VideoCodec.H264: + raise ValueError("Only H264 codec is supported for now") + with av.open(path, mode='w', options={'movflags': 'use_metadata_tags'}) as output: + # Add metadata before writing any streams + if metadata is not None: + for key, value in metadata.items(): + output.metadata[key] = json.dumps(value) + + frame_rate = Fraction(round(self.__components.frame_rate * 1000), 1000) + # Create a video stream + video_stream = output.add_stream('h264', rate=frame_rate) + video_stream.width = self.__components.images.shape[2] + video_stream.height = self.__components.images.shape[1] + video_stream.pix_fmt = 'yuv420p' + + # Create an audio stream + audio_sample_rate = 1 + audio_stream: Optional[av.AudioStream] = None + if self.__components.audio: + audio_sample_rate = int(self.__components.audio['sample_rate']) + audio_stream = output.add_stream('aac', rate=audio_sample_rate) + audio_stream.sample_rate = audio_sample_rate + audio_stream.format = 'fltp' + + # Encode video + for i, frame in enumerate(self.__components.images): + img = (frame * 255).clamp(0, 255).byte().cpu().numpy() # shape: (H, W, 3) + frame = av.VideoFrame.from_ndarray(img, format='rgb24') + frame = frame.reformat(format='yuv420p') # Convert to YUV420P as required by h264 + packet = video_stream.encode(frame) + output.mux(packet) + + # Flush video + packet = video_stream.encode(None) + output.mux(packet) + + if audio_stream and self.__components.audio: + # Encode audio + samples_per_frame = int(audio_sample_rate / frame_rate) + num_frames = self.__components.audio['waveform'].shape[2] // samples_per_frame + for i in range(num_frames): + start = i * samples_per_frame + end = start + samples_per_frame + # TODO(Feature) - Add support for stereo audio + chunk = self.__components.audio['waveform'][0, 0, start:end].unsqueeze(0).numpy() + audio_frame = av.AudioFrame.from_ndarray(chunk, format='fltp', layout='mono') + audio_frame.sample_rate = audio_sample_rate + audio_frame.pts = i * samples_per_frame + for packet in audio_stream.encode(audio_frame): + output.mux(packet) + + # Flush audio + for packet in audio_stream.encode(None): + output.mux(packet) + diff --git a/comfy_api/util/__init__.py b/comfy_api/util/__init__.py new file mode 100644 index 000000000..9019c46db --- /dev/null +++ b/comfy_api/util/__init__.py @@ -0,0 +1,8 @@ +from .video_types import VideoContainer, VideoCodec, VideoComponents + +__all__ = [ + # Utility Types + "VideoContainer", + "VideoCodec", + "VideoComponents", +] diff --git a/comfy_api/util/video_types.py b/comfy_api/util/video_types.py new file mode 100644 index 000000000..d09663db9 --- /dev/null +++ b/comfy_api/util/video_types.py @@ -0,0 +1,51 @@ +from __future__ import annotations +from dataclasses import dataclass +from enum import Enum +from fractions import Fraction +from typing import Optional +from comfy_api.input import ImageInput, AudioInput + +class VideoCodec(str, Enum): + AUTO = "auto" + H264 = "h264" + + @classmethod + def as_input(cls) -> list[str]: + """ + Returns a list of codec names that can be used as node input. + """ + return [member.value for member in cls] + +class VideoContainer(str, Enum): + AUTO = "auto" + MP4 = "mp4" + + @classmethod + def as_input(cls) -> list[str]: + """ + Returns a list of container names that can be used as node input. + """ + return [member.value for member in cls] + + @classmethod + def get_extension(cls, value) -> str: + """ + Returns the file extension for the container. + """ + if isinstance(value, str): + value = cls(value) + if value == VideoContainer.MP4 or value == VideoContainer.AUTO: + return "mp4" + return "" + +@dataclass +class VideoComponents: + """ + Dataclass representing the components of a video. + """ + + images: ImageInput + frame_rate: Fraction + audio: Optional[AudioInput] = None + metadata: Optional[dict] = None + diff --git a/comfy_extras/nodes_video.py b/comfy_extras/nodes_video.py index a9e244ebe..61f7171b2 100644 --- a/comfy_extras/nodes_video.py +++ b/comfy_extras/nodes_video.py @@ -5,9 +5,13 @@ import av import torch import folder_paths import json +from typing import Optional, Literal from fractions import Fraction -from comfy.comfy_types import FileLocator - +from comfy.comfy_types import IO, FileLocator, ComfyNodeABC +from comfy_api.input import ImageInput, AudioInput, VideoInput +from comfy_api.util import VideoContainer, VideoCodec, VideoComponents +from comfy_api.input_impl import VideoFromFile, VideoFromComponents +from comfy.cli_args import args class SaveWEBM: def __init__(self): @@ -75,7 +79,163 @@ class SaveWEBM: return {"ui": {"images": results, "animated": (True,)}} # TODO: frontend side +class SaveVideo(ComfyNodeABC): + def __init__(self): + self.output_dir = folder_paths.get_output_directory() + self.type: Literal["output"] = "output" + self.prefix_append = "" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "video": (IO.VIDEO, {"tooltip": "The video to save."}), + "filename_prefix": ("STRING", {"default": "video/ComfyUI", "tooltip": "The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes."}), + "format": (VideoContainer.as_input(), {"default": "auto", "tooltip": "The format to save the video as."}), + "codec": (VideoCodec.as_input(), {"default": "auto", "tooltip": "The codec to use for the video."}), + }, + "hidden": { + "prompt": "PROMPT", + "extra_pnginfo": "EXTRA_PNGINFO" + }, + } + + RETURN_TYPES = () + FUNCTION = "save_video" + + OUTPUT_NODE = True + + CATEGORY = "image/video" + DESCRIPTION = "Saves the input images to your ComfyUI output directory." + + def save_video(self, video: VideoInput, filename_prefix, format, codec, prompt=None, extra_pnginfo=None): + filename_prefix += self.prefix_append + width, height = video.get_dimensions() + full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path( + filename_prefix, + self.output_dir, + width, + height + ) + results: list[FileLocator] = list() + saved_metadata = None + if not args.disable_metadata: + metadata = {} + if extra_pnginfo is not None: + metadata.update(extra_pnginfo) + if prompt is not None: + metadata["prompt"] = prompt + if len(metadata) > 0: + saved_metadata = metadata + file = f"{filename}_{counter:05}_.{VideoContainer.get_extension(format)}" + video.save_to( + os.path.join(full_output_folder, file), + format=format, + codec=codec, + metadata=saved_metadata + ) + + results.append({ + "filename": file, + "subfolder": subfolder, + "type": self.type + }) + counter += 1 + + return { "ui": { "images": results, "animated": (True,) } } + +class CreateVideo(ComfyNodeABC): + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "images": (IO.IMAGE, {"tooltip": "The images to create a video from."}), + "fps": ("FLOAT", {"default": 30.0, "min": 1.0, "max": 120.0, "step": 1.0}), + }, + "optional": { + "audio": (IO.AUDIO, {"tooltip": "The audio to add to the video."}), + } + } + + RETURN_TYPES = (IO.VIDEO,) + FUNCTION = "create_video" + + CATEGORY = "image/video" + DESCRIPTION = "Create a video from images." + + def create_video(self, images: ImageInput, fps: float, audio: Optional[AudioInput] = None): + return (VideoFromComponents( + VideoComponents( + images=images, + audio=audio, + frame_rate=Fraction(fps), + ) + ),) + +class GetVideoComponents(ComfyNodeABC): + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "video": (IO.VIDEO, {"tooltip": "The video to extract components from."}), + } + } + RETURN_TYPES = (IO.IMAGE, IO.AUDIO, IO.FLOAT) + RETURN_NAMES = ("images", "audio", "fps") + FUNCTION = "get_components" + + CATEGORY = "image/video" + DESCRIPTION = "Extracts all components from a video: frames, audio, and framerate." + + def get_components(self, video: VideoInput): + components = video.get_components() + + return (components.images, components.audio, float(components.frame_rate)) + +class LoadVideo(ComfyNodeABC): + @classmethod + def INPUT_TYPES(cls): + input_dir = folder_paths.get_input_directory() + files = [f for f in os.listdir(input_dir) if os.path.isfile(os.path.join(input_dir, f))] + files = folder_paths.filter_files_content_types(files, ["video"]) + return {"required": + {"file": (sorted(files), {"video_upload": True})}, + } + + CATEGORY = "image/video" + + RETURN_TYPES = (IO.VIDEO,) + FUNCTION = "load_video" + def load_video(self, file): + video_path = folder_paths.get_annotated_filepath(file) + return (VideoFromFile(video_path),) + + @classmethod + def IS_CHANGED(cls, file): + video_path = folder_paths.get_annotated_filepath(file) + mod_time = os.path.getmtime(video_path) + # Instead of hashing the file, we can just use the modification time to avoid + # rehashing large files. + return mod_time + + @classmethod + def VALIDATE_INPUTS(cls, file): + if not folder_paths.exists_annotated_filepath(file): + return "Invalid video file: {}".format(file) + + return True NODE_CLASS_MAPPINGS = { "SaveWEBM": SaveWEBM, + "SaveVideo": SaveVideo, + "CreateVideo": CreateVideo, + "GetVideoComponents": GetVideoComponents, + "LoadVideo": LoadVideo, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "SaveVideo": "Save Video", + "CreateVideo": "Create Video", + "GetVideoComponents": "Get Video Components", + "LoadVideo": "Load Video", } diff --git a/folder_paths.py b/folder_paths.py index 9a525e5a1..f0b3fd103 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -4,7 +4,7 @@ import os import time import mimetypes import logging -from typing import Literal +from typing import Literal, List from collections.abc import Collection from comfy.cli_args import args @@ -141,7 +141,7 @@ def get_directory_by_type(type_name: str) -> str | None: return get_input_directory() return None -def filter_files_content_types(files: list[str], content_types: Literal["image", "video", "audio", "model"]) -> list[str]: +def filter_files_content_types(files: list[str], content_types: List[Literal["image", "video", "audio", "model"]]) -> list[str]: """ Example: files = os.listdir(folder_paths.get_input_directory()) From 005a91ce2b6455d7e76f7a5cf2e77dbd020afa1f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 29 Apr 2025 03:29:38 -0700 Subject: [PATCH 09/22] Latest desktop and portable should work on blackwell. (#7861) Removed the mention about the cards from the readme. --- README.md | 2 -- 1 file changed, 2 deletions(-) diff --git a/README.md b/README.md index 62800bb4f..24a65b942 100644 --- a/README.md +++ b/README.md @@ -149,8 +149,6 @@ Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you If you have trouble extracting it, right click the file -> properties -> unblock -If you have a 50 series Blackwell card like a 5090 or 5080 see [this discussion thread](https://github.com/comfyanonymous/ComfyUI/discussions/6643) - #### How do I share models between another UI and ComfyUI? See the [Config file](extra_model_paths.yaml.example) to set the search paths for models. In the standalone windows build you can find this file in the ComfyUI directory. Rename this file to extra_model_paths.yaml and edit it with your favorite text editor. From 45503f649925fdf8e6ddce85fabee551ee3c446b Mon Sep 17 00:00:00 2001 From: Chenlei Hu Date: Tue, 29 Apr 2025 06:32:34 -0400 Subject: [PATCH 10/22] Add release process section to README (#7855) * Add release process section to README * move * Update README.md --- README.md | 18 +++++++++++++++++- 1 file changed, 17 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index 24a65b942..0f39cfce2 100644 --- a/README.md +++ b/README.md @@ -49,7 +49,6 @@ Supports all operating systems and GPU types (NVIDIA, AMD, Intel, Apple Silicon, ## [Examples](https://comfyanonymous.github.io/ComfyUI_examples/) See what ComfyUI can do with the [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/). - ## Features - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything. - Image Models @@ -99,6 +98,23 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith Workflow examples can be found on the [Examples page](https://comfyanonymous.github.io/ComfyUI_examples/) +## Release Process + +ComfyUI follows a weekly release cycle every Friday, with three interconnected repositories: + +1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)** + - Releases a new stable version (e.g., v0.7.0) + - Serves as the foundation for the desktop release + +2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)** + - Builds a new release using the latest stable core version + - Version numbers match the core release (e.g., Desktop v1.7.0 uses Core v1.7.0) + +3. **[ComfyUI Frontend](https://github.com/Comfy-Org/ComfyUI_frontend)** + - Weekly frontend updates are merged into the core repository + - Features are frozen for the upcoming core release + - Development continues for the next release cycle + ## Shortcuts | Keybind | Explanation | From 5c5457a4ef151f9f855020ff544b4362cdcf1201 Mon Sep 17 00:00:00 2001 From: Terry Jia Date: Tue, 29 Apr 2025 11:28:04 -0400 Subject: [PATCH 11/22] support more example folders (#7836) * support more example folders * add warning message --- app/custom_node_manager.py | 43 ++++++++++++++++++++++++-------------- 1 file changed, 27 insertions(+), 16 deletions(-) diff --git a/app/custom_node_manager.py b/app/custom_node_manager.py index 42b0d75ba..27d85d9ce 100644 --- a/app/custom_node_manager.py +++ b/app/custom_node_manager.py @@ -93,16 +93,20 @@ class CustomNodeManager: def add_routes(self, routes, webapp, loadedModules): + example_workflow_folder_names = ["example_workflows", "example", "examples", "workflow", "workflows"] + @routes.get("/workflow_templates") async def get_workflow_templates(request): """Returns a web response that contains the map of custom_nodes names and their associated workflow templates. The ones without templates are omitted.""" - files = [ - file - for folder in folder_paths.get_folder_paths("custom_nodes") - for file in glob.glob( - os.path.join(folder, "*/example_workflows/*.json") - ) - ] + + files = [] + + for folder in folder_paths.get_folder_paths("custom_nodes"): + for folder_name in example_workflow_folder_names: + pattern = os.path.join(folder, f"*/{folder_name}/*.json") + matched_files = glob.glob(pattern) + files.extend(matched_files) + workflow_templates_dict = ( {} ) # custom_nodes folder name -> example workflow names @@ -118,15 +122,22 @@ class CustomNodeManager: # Serve workflow templates from custom nodes. for module_name, module_dir in loadedModules: - workflows_dir = os.path.join(module_dir, "example_workflows") - if os.path.exists(workflows_dir): - webapp.add_routes( - [ - web.static( - "/api/workflow_templates/" + module_name, workflows_dir - ) - ] - ) + for folder_name in example_workflow_folder_names: + workflows_dir = os.path.join(module_dir, folder_name) + + if os.path.exists(workflows_dir): + if folder_name != "example_workflows": + logging.warning( + "WARNING: Found example workflow folder '%s' for custom node '%s', consider renaming it to 'example_workflows'", + folder_name, module_name) + + webapp.add_routes( + [ + web.static( + "/api/workflow_templates/" + module_name, workflows_dir + ) + ] + ) @routes.get("/i18n") async def get_i18n(request): From 0a66d4b0afe4a78a200809b7d1d3beec6c6a2a8f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 29 Apr 2025 17:28:52 -0700 Subject: [PATCH 12/22] Per device stream counters for async offload. (#7873) --- comfy/model_management.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 78317af3c..44aff3762 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -946,9 +946,9 @@ if args.async_offload: NUM_STREAMS = 2 logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS)) -stream_counter = 0 +stream_counters = {} def get_offload_stream(device): - global stream_counter + stream_counter = stream_counters.get(device, 0) if NUM_STREAMS <= 1: return None @@ -958,6 +958,7 @@ def get_offload_stream(device): stream_counter = (stream_counter + 1) % len(ss) if is_device_cuda(device): ss[stream_counter].wait_stream(torch.cuda.current_stream()) + stream_counters[device] = stream_counter return s elif is_device_cuda(device): ss = [] @@ -966,6 +967,7 @@ def get_offload_stream(device): STREAMS[device] = ss s = ss[stream_counter] stream_counter = (stream_counter + 1) % len(ss) + stream_counters[device] = stream_counter return s return None From 7ee96455e2ed29293aa6076db9b4866862d41142 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 29 Apr 2025 17:38:45 -0700 Subject: [PATCH 13/22] Bump minimum pyav version to 14.2.0 (#7874) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 10cc177af..f64a05947 100644 --- a/requirements.txt +++ b/requirements.txt @@ -22,5 +22,5 @@ psutil kornia>=0.7.1 spandrel soundfile -av>=14.1.0 +av>=14.2.0 pydantic~=2.0 From dbc726f80c9ac0512d2611fad63d984b8c03886f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 29 Apr 2025 17:42:00 -0700 Subject: [PATCH 14/22] Better vace memory estimation. (#7875) --- comfy/ldm/wan/model.py | 1 + comfy/supported_models.py | 4 ++++ 2 files changed, 5 insertions(+) diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index b8eec3afb..66bee7480 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -631,6 +631,7 @@ class VaceWanModel(WanModel): if ii is not None: c_skip, c = self.vace_blocks[ii](c, x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len) x += c_skip * vace_strength + del c_skip # head x = self.head(x, e) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 5e55035cf..69bcee1f7 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -993,6 +993,10 @@ class WAN21_Vace(WAN21_T2V): "model_type": "vace", } + def __init__(self, unet_config): + super().__init__(unet_config) + self.memory_usage_factor = 1.2 * self.memory_usage_factor + def get_model(self, state_dict, prefix="", device=None): out = model_base.WAN21_Vace(self, image_to_video=False, device=device) return out From b1c7291569daaeec34ccf6df25d57886eb73d98e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 30 Apr 2025 11:18:20 -0700 Subject: [PATCH 15/22] Test updater in the windows release workflow. (#7886) --- .github/workflows/windows_release_package.yml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.github/workflows/windows_release_package.yml b/.github/workflows/windows_release_package.yml index 80a45b321..3926a65f3 100644 --- a/.github/workflows/windows_release_package.yml +++ b/.github/workflows/windows_release_package.yml @@ -88,6 +88,8 @@ jobs: cd ComfyUI_windows_portable python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu + python_embeded/python.exe -s ./update/update.py ComfyUI/ + ls - name: Upload binaries to release From 39c27a37052b5e18850a6c63cd0a4b92131db3ba Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 30 Apr 2025 11:42:18 -0700 Subject: [PATCH 16/22] Add updater test to stable release workflow. (#7887) --- .github/workflows/stable-release.yml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.github/workflows/stable-release.yml b/.github/workflows/stable-release.yml index c4302cdd6..a046ff9ea 100644 --- a/.github/workflows/stable-release.yml +++ b/.github/workflows/stable-release.yml @@ -91,6 +91,8 @@ jobs: cd ComfyUI_windows_portable python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu + python_embeded/python.exe -s ./update/update.py ComfyUI/ + ls - name: Upload binaries to release From 4ca3d842774421968cc19c319ee96daa58a98fbb Mon Sep 17 00:00:00 2001 From: Silver <65376327+silveroxides@users.noreply.github.com> Date: Thu, 1 May 2025 02:57:00 +0200 Subject: [PATCH 17/22] Support for Chroma - Flux1 Schnell distilled with CFG (#7355) * Upload files for Chroma Implementation * Remove trailing whitespace * trim more trailing whitespace..oops * remove unused imports * Add supported_inference_dtypes * Set min_length to 0 and remove attention_mask=True * Set min_length to 1 * get_mdulations added from blepping and minor changes * Add lora conversion if statement in lora.py * Update supported_models.py * update model_base.py * add uptream commits * set modelType.FLOW, will cause beta scheduler to work properly * Adjust memory usage factor and remove unnecessary code * fix mistake * reduce code duplication * remove unused imports * refactor for upstream sync * sync chroma-support with upstream via syncbranch patch * Update sd.py * Add Chroma as option for the OptimalStepsScheduler node --- comfy/ldm/chroma/layers.py | 183 +++++++++++++++++++ comfy/ldm/chroma/math.py | 44 +++++ comfy/ldm/chroma/model.py | 271 +++++++++++++++++++++++++++++ comfy/lora.py | 2 +- comfy/model_base.py | 62 +++++++ comfy/model_detection.py | 26 +++ comfy/sd.py | 5 + comfy/supported_models.py | 30 +++- comfy/text_encoders/chroma.py | 43 +++++ comfy_extras/nodes_optimalsteps.py | 3 +- nodes.py | 2 +- 11 files changed, 667 insertions(+), 4 deletions(-) create mode 100644 comfy/ldm/chroma/layers.py create mode 100644 comfy/ldm/chroma/math.py create mode 100644 comfy/ldm/chroma/model.py create mode 100644 comfy/text_encoders/chroma.py diff --git a/comfy/ldm/chroma/layers.py b/comfy/ldm/chroma/layers.py new file mode 100644 index 000000000..dd0b72f70 --- /dev/null +++ b/comfy/ldm/chroma/layers.py @@ -0,0 +1,183 @@ +import torch +from torch import Tensor, nn + +from .math import attention +from comfy.ldm.flux.layers import ( + MLPEmbedder, + RMSNorm, + QKNorm, + SelfAttention, + ModulationOut, +) + + + +class ChromaModulationOut(ModulationOut): + @classmethod + def from_offset(cls, tensor: torch.Tensor, offset: int = 0) -> ModulationOut: + return cls( + shift=tensor[:, offset : offset + 1, :], + scale=tensor[:, offset + 1 : offset + 2, :], + gate=tensor[:, offset + 2 : offset + 3, :], + ) + + + + +class Approximator(nn.Module): + def __init__(self, in_dim: int, out_dim: int, hidden_dim: int, n_layers = 5, dtype=None, device=None, operations=None): + super().__init__() + self.in_proj = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device) + self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)]) + self.norms = nn.ModuleList([RMSNorm(hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)]) + self.out_proj = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device) + + @property + def device(self): + # Get the device of the module (assumes all parameters are on the same device) + return next(self.parameters()).device + + def forward(self, x: Tensor) -> Tensor: + x = self.in_proj(x) + + for layer, norms in zip(self.layers, self.norms): + x = x + layer(norms(x)) + + x = self.out_proj(x) + + return x + + +class DoubleStreamBlock(nn.Module): + def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None): + super().__init__() + + mlp_hidden_dim = int(hidden_size * mlp_ratio) + self.num_heads = num_heads + self.hidden_size = hidden_size + self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) + + self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.img_mlp = nn.Sequential( + operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), + nn.GELU(approximate="tanh"), + operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), + ) + + self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) + + self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.txt_mlp = nn.Sequential( + operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), + nn.GELU(approximate="tanh"), + operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), + ) + self.flipped_img_txt = flipped_img_txt + + def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None): + (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec + + # prepare image for attention + img_modulated = self.img_norm1(img) + img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift + img_qkv = self.img_attn.qkv(img_modulated) + img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + img_q, img_k = self.img_attn.norm(img_q, img_k, img_v) + + # prepare txt for attention + txt_modulated = self.txt_norm1(txt) + txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift + txt_qkv = self.txt_attn.qkv(txt_modulated) + txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v) + + # run actual attention + attn = attention(torch.cat((txt_q, img_q), dim=2), + torch.cat((txt_k, img_k), dim=2), + torch.cat((txt_v, img_v), dim=2), + pe=pe, mask=attn_mask) + + txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :] + + # calculate the img bloks + img = img + img_mod1.gate * self.img_attn.proj(img_attn) + img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift) + + # calculate the txt bloks + txt += txt_mod1.gate * self.txt_attn.proj(txt_attn) + txt += txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift) + + if txt.dtype == torch.float16: + txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504) + + return img, txt + + +class SingleStreamBlock(nn.Module): + """ + A DiT block with parallel linear layers as described in + https://arxiv.org/abs/2302.05442 and adapted modulation interface. + """ + + def __init__( + self, + hidden_size: int, + num_heads: int, + mlp_ratio: float = 4.0, + qk_scale: float = None, + dtype=None, + device=None, + operations=None + ): + super().__init__() + self.hidden_dim = hidden_size + self.num_heads = num_heads + head_dim = hidden_size // num_heads + self.scale = qk_scale or head_dim**-0.5 + + self.mlp_hidden_dim = int(hidden_size * mlp_ratio) + # qkv and mlp_in + self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device) + # proj and mlp_out + self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device) + + self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations) + + self.hidden_size = hidden_size + self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + + self.mlp_act = nn.GELU(approximate="tanh") + + def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None) -> Tensor: + mod = vec + x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift + qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) + + q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + q, k = self.norm(q, k, v) + + # compute attention + attn = attention(q, k, v, pe=pe, mask=attn_mask) + # compute activation in mlp stream, cat again and run second linear layer + output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2)) + x += mod.gate * output + if x.dtype == torch.float16: + x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504) + return x + + +class LastLayer(nn.Module): + def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None): + super().__init__() + self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.linear = operations.Linear(hidden_size, out_channels, bias=True, dtype=dtype, device=device) + + def forward(self, x: Tensor, vec: Tensor) -> Tensor: + shift, scale = vec + shift = shift.squeeze(1) + scale = scale.squeeze(1) + x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :] + x = self.linear(x) + return x diff --git a/comfy/ldm/chroma/math.py b/comfy/ldm/chroma/math.py new file mode 100644 index 000000000..36b67931c --- /dev/null +++ b/comfy/ldm/chroma/math.py @@ -0,0 +1,44 @@ +import torch +from einops import rearrange +from torch import Tensor + +from comfy.ldm.modules.attention import optimized_attention +import comfy.model_management + + +def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None) -> Tensor: + q_shape = q.shape + k_shape = k.shape + + q = q.float().reshape(*q.shape[:-1], -1, 1, 2) + k = k.float().reshape(*k.shape[:-1], -1, 1, 2) + q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v) + k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v) + + heads = q.shape[1] + x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask) + return x + + +def rope(pos: Tensor, dim: int, theta: int) -> Tensor: + assert dim % 2 == 0 + if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled(): + device = torch.device("cpu") + else: + device = pos.device + + scale = torch.linspace(0, (dim - 2) / dim, steps=dim//2, dtype=torch.float64, device=device) + omega = 1.0 / (theta**scale) + out = torch.einsum("...n,d->...nd", pos.to(dtype=torch.float32, device=device), omega) + out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1) + out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2) + return out.to(dtype=torch.float32, device=pos.device) + + +def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): + xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2) + xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2) + xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1] + xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1] + return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk) + diff --git a/comfy/ldm/chroma/model.py b/comfy/ldm/chroma/model.py new file mode 100644 index 000000000..636748fc5 --- /dev/null +++ b/comfy/ldm/chroma/model.py @@ -0,0 +1,271 @@ +#Original code can be found on: https://github.com/black-forest-labs/flux + +from dataclasses import dataclass + +import torch +from torch import Tensor, nn +from einops import rearrange, repeat +import comfy.ldm.common_dit + +from comfy.ldm.flux.layers import ( + EmbedND, + timestep_embedding, +) + +from .layers import ( + DoubleStreamBlock, + LastLayer, + SingleStreamBlock, + Approximator, + ChromaModulationOut, +) + + +@dataclass +class ChromaParams: + in_channels: int + out_channels: int + context_in_dim: int + hidden_size: int + mlp_ratio: float + num_heads: int + depth: int + depth_single_blocks: int + axes_dim: list + theta: int + patch_size: int + qkv_bias: bool + in_dim: int + out_dim: int + hidden_dim: int + n_layers: int + + + + +class Chroma(nn.Module): + """ + Transformer model for flow matching on sequences. + """ + + def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs): + super().__init__() + self.dtype = dtype + params = ChromaParams(**kwargs) + self.params = params + self.patch_size = params.patch_size + self.in_channels = params.in_channels + self.out_channels = params.out_channels + if params.hidden_size % params.num_heads != 0: + raise ValueError( + f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}" + ) + pe_dim = params.hidden_size // params.num_heads + if sum(params.axes_dim) != pe_dim: + raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}") + self.hidden_size = params.hidden_size + self.num_heads = params.num_heads + self.in_dim = params.in_dim + self.out_dim = params.out_dim + self.hidden_dim = params.hidden_dim + self.n_layers = params.n_layers + self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim) + self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device) + self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device) + # set as nn identity for now, will overwrite it later. + self.distilled_guidance_layer = Approximator( + in_dim=self.in_dim, + hidden_dim=self.hidden_dim, + out_dim=self.out_dim, + n_layers=self.n_layers, + dtype=dtype, device=device, operations=operations + ) + + + self.double_blocks = nn.ModuleList( + [ + DoubleStreamBlock( + self.hidden_size, + self.num_heads, + mlp_ratio=params.mlp_ratio, + qkv_bias=params.qkv_bias, + dtype=dtype, device=device, operations=operations + ) + for _ in range(params.depth) + ] + ) + + self.single_blocks = nn.ModuleList( + [ + SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations) + for _ in range(params.depth_single_blocks) + ] + ) + + if final_layer: + self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations) + + self.skip_mmdit = [] + self.skip_dit = [] + self.lite = False + + def get_modulations(self, tensor: torch.Tensor, block_type: str, *, idx: int = 0): + # This function slices up the modulations tensor which has the following layout: + # single : num_single_blocks * 3 elements + # double_img : num_double_blocks * 6 elements + # double_txt : num_double_blocks * 6 elements + # final : 2 elements + if block_type == "final": + return (tensor[:, -2:-1, :], tensor[:, -1:, :]) + single_block_count = self.params.depth_single_blocks + double_block_count = self.params.depth + offset = 3 * idx + if block_type == "single": + return ChromaModulationOut.from_offset(tensor, offset) + # Double block modulations are 6 elements so we double 3 * idx. + offset *= 2 + if block_type in {"double_img", "double_txt"}: + # Advance past the single block modulations. + offset += 3 * single_block_count + if block_type == "double_txt": + # Advance past the double block img modulations. + offset += 6 * double_block_count + return ( + ChromaModulationOut.from_offset(tensor, offset), + ChromaModulationOut.from_offset(tensor, offset + 3), + ) + raise ValueError("Bad block_type") + + + def forward_orig( + self, + img: Tensor, + img_ids: Tensor, + txt: Tensor, + txt_ids: Tensor, + timesteps: Tensor, + guidance: Tensor = None, + control = None, + transformer_options={}, + attn_mask: Tensor = None, + ) -> Tensor: + patches_replace = transformer_options.get("patches_replace", {}) + if img.ndim != 3 or txt.ndim != 3: + raise ValueError("Input img and txt tensors must have 3 dimensions.") + + # running on sequences img + img = self.img_in(img) + + # distilled vector guidance + mod_index_length = 344 + distill_timestep = timestep_embedding(timesteps.detach().clone(), 16).to(img.device, img.dtype) + # guidance = guidance * + distil_guidance = timestep_embedding(guidance.detach().clone(), 16).to(img.device, img.dtype) + + # get all modulation index + modulation_index = timestep_embedding(torch.arange(mod_index_length), 32).to(img.device, img.dtype) + # we need to broadcast the modulation index here so each batch has all of the index + modulation_index = modulation_index.unsqueeze(0).repeat(img.shape[0], 1, 1).to(img.device, img.dtype) + # and we need to broadcast timestep and guidance along too + timestep_guidance = torch.cat([distill_timestep, distil_guidance], dim=1).unsqueeze(1).repeat(1, mod_index_length, 1).to(img.dtype).to(img.device, img.dtype) + # then and only then we could concatenate it together + input_vec = torch.cat([timestep_guidance, modulation_index], dim=-1).to(img.device, img.dtype) + + mod_vectors = self.distilled_guidance_layer(input_vec) + + txt = self.txt_in(txt) + + ids = torch.cat((txt_ids, img_ids), dim=1) + pe = self.pe_embedder(ids) + + blocks_replace = patches_replace.get("dit", {}) + for i, block in enumerate(self.double_blocks): + if i not in self.skip_mmdit: + double_mod = ( + self.get_modulations(mod_vectors, "double_img", idx=i), + self.get_modulations(mod_vectors, "double_txt", idx=i), + ) + if ("double_block", i) in blocks_replace: + def block_wrap(args): + out = {} + out["img"], out["txt"] = block(img=args["img"], + txt=args["txt"], + vec=args["vec"], + pe=args["pe"], + attn_mask=args.get("attn_mask")) + return out + + out = blocks_replace[("double_block", i)]({"img": img, + "txt": txt, + "vec": double_mod, + "pe": pe, + "attn_mask": attn_mask}, + {"original_block": block_wrap}) + txt = out["txt"] + img = out["img"] + else: + img, txt = block(img=img, + txt=txt, + vec=double_mod, + pe=pe, + attn_mask=attn_mask) + + if control is not None: # Controlnet + control_i = control.get("input") + if i < len(control_i): + add = control_i[i] + if add is not None: + img += add + + img = torch.cat((txt, img), 1) + + for i, block in enumerate(self.single_blocks): + if i not in self.skip_dit: + single_mod = self.get_modulations(mod_vectors, "single", idx=i) + if ("single_block", i) in blocks_replace: + def block_wrap(args): + out = {} + out["img"] = block(args["img"], + vec=args["vec"], + pe=args["pe"], + attn_mask=args.get("attn_mask")) + return out + + out = blocks_replace[("single_block", i)]({"img": img, + "vec": single_mod, + "pe": pe, + "attn_mask": attn_mask}, + {"original_block": block_wrap}) + img = out["img"] + else: + img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask) + + if control is not None: # Controlnet + control_o = control.get("output") + if i < len(control_o): + add = control_o[i] + if add is not None: + img[:, txt.shape[1] :, ...] += add + + img = img[:, txt.shape[1] :, ...] + final_mod = self.get_modulations(mod_vectors, "final") + img = self.final_layer(img, vec=final_mod) # (N, T, patch_size ** 2 * out_channels) + return img + + def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs): + bs, c, h, w = x.shape + patch_size = 2 + x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size)) + + img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size) + + h_len = ((h + (patch_size // 2)) // patch_size) + w_len = ((w + (patch_size // 2)) // patch_size) + img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype) + img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) + img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) + img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs) + + txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype) + out = self.forward_orig(img, img_ids, context, txt_ids, timestep, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None)) + return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w] diff --git a/comfy/lora.py b/comfy/lora.py index fff524be2..dbabd3335 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -252,7 +252,7 @@ def model_lora_keys_unet(model, key_map={}): key_lora = k[len("diffusion_model."):-len(".weight")] key_map["base_model.model.{}".format(key_lora)] = k #official hunyuan lora format - if isinstance(model, comfy.model_base.Flux): #Diffusers lora Flux + if isinstance(model, comfy.model_base.Flux) or isinstance(model, comfy.model_base.Chroma): #Diffusers lora Flux or a diffusers lora Chroma diffusers_keys = comfy.utils.flux_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.") for k in diffusers_keys: if k.endswith(".weight"): diff --git a/comfy/model_base.py b/comfy/model_base.py index d2aa4ce7a..1a06bb503 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -38,6 +38,7 @@ import comfy.ldm.lumina.model import comfy.ldm.wan.model import comfy.ldm.hunyuan3d.model import comfy.ldm.hidream.model +import comfy.ldm.chroma.model import comfy.model_management import comfy.patcher_extension @@ -1108,3 +1109,64 @@ class HiDream(BaseModel): if image_cond is not None: out['image_cond'] = comfy.conds.CONDNoiseShape(self.process_latent_in(image_cond)) return out + +class Chroma(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.chroma.model.Chroma) + + def concat_cond(self, **kwargs): + try: + #Handle Flux control loras dynamically changing the img_in weight. + num_channels = self.diffusion_model.img_in.weight.shape[1] + except: + #Some cases like tensorrt might not have the weights accessible + num_channels = self.model_config.unet_config["in_channels"] + + out_channels = self.model_config.unet_config["out_channels"] + + if num_channels <= out_channels: + return None + + image = kwargs.get("concat_latent_image", None) + noise = kwargs.get("noise", None) + device = kwargs["device"] + + if image is None: + image = torch.zeros_like(noise) + + image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + image = utils.resize_to_batch_size(image, noise.shape[0]) + image = self.process_latent_in(image) + if num_channels <= out_channels * 2: + return image + + #inpaint model + mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) + if mask is None: + mask = torch.ones_like(noise)[:, :1] + + mask = torch.mean(mask, dim=1, keepdim=True) + mask = utils.common_upscale(mask.to(device), noise.shape[-1] * 8, noise.shape[-2] * 8, "bilinear", "center") + mask = mask.view(mask.shape[0], mask.shape[2] // 8, 8, mask.shape[3] // 8, 8).permute(0, 2, 4, 1, 3).reshape(mask.shape[0], -1, mask.shape[2] // 8, mask.shape[3] // 8) + mask = utils.resize_to_batch_size(mask, noise.shape[0]) + return torch.cat((image, mask), dim=1) + + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + # upscale the attention mask, since now we + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + shape = kwargs["noise"].shape + mask_ref_size = kwargs["attention_mask_img_shape"] + # the model will pad to the patch size, and then divide + # essentially dividing and rounding up + (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size)) + attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok)) + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + guidance = 0.0 + out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor((guidance,))) + return out diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 76de78a8a..daf6d04e7 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -154,6 +154,32 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["guidance_embed"] = len(guidance_keys) > 0 return dit_config + if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma + dit_config = {} + dit_config["image_model"] = "chroma" + dit_config["depth"] = 48 + dit_config["in_channels"] = 64 + patch_size = 2 + dit_config["patch_size"] = patch_size + in_key = "{}img_in.weight".format(key_prefix) + if in_key in state_dict_keys: + dit_config["in_channels"] = state_dict[in_key].shape[1] + dit_config["out_channels"] = 64 + dit_config["context_in_dim"] = 4096 + dit_config["hidden_size"] = 3072 + dit_config["mlp_ratio"] = 4.0 + dit_config["num_heads"] = 24 + dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.') + dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.') + dit_config["axes_dim"] = [16, 56, 56] + dit_config["theta"] = 10000 + dit_config["qkv_bias"] = True + dit_config["in_dim"] = 64 + dit_config["out_dim"] = 3072 + dit_config["hidden_dim"] = 5120 + dit_config["n_layers"] = 5 + return dit_config + if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and '{}img_in.weight'.format(key_prefix) in state_dict_keys: #Flux dit_config = {} dit_config["image_model"] = "flux" diff --git a/comfy/sd.py b/comfy/sd.py index 748f6c1ec..454b5929a 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -42,6 +42,7 @@ import comfy.text_encoders.cosmos import comfy.text_encoders.lumina2 import comfy.text_encoders.wan import comfy.text_encoders.hidream +import comfy.text_encoders.chroma import comfy.model_patcher import comfy.lora @@ -714,6 +715,7 @@ class CLIPType(Enum): LUMINA2 = 12 WAN = 13 HIDREAM = 14 + CHROMA = 15 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}): @@ -829,6 +831,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data), clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None, llama_scaled_fp8=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer + elif clip_type == CLIPType.CHROMA: + clip_target.clip = comfy.text_encoders.chroma.chroma_te(**t5xxl_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.chroma.ChromaT5Tokenizer else: #CLIPType.MOCHI clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 69bcee1f7..f03f2790e 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -17,6 +17,7 @@ import comfy.text_encoders.hunyuan_video import comfy.text_encoders.cosmos import comfy.text_encoders.lumina2 import comfy.text_encoders.wan +import comfy.text_encoders.chroma from . import supported_models_base from . import latent_formats @@ -1068,7 +1069,34 @@ class HiDream(supported_models_base.BASE): def clip_target(self, state_dict={}): return None # TODO +class Chroma(supported_models_base.BASE): + unet_config = { + "image_model": "chroma", + } -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream] + unet_extra_config = { + } + + sampling_settings = { + "multiplier": 1.0, + } + + latent_format = comfy.latent_formats.Flux + + memory_usage_factor = 3.2 + + supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] + + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Chroma(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.chroma.ChromaTokenizer, comfy.text_encoders.chroma.chroma_te(**t5_detect)) + +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma] models += [SVD_img2vid] diff --git a/comfy/text_encoders/chroma.py b/comfy/text_encoders/chroma.py new file mode 100644 index 000000000..aa8dffb25 --- /dev/null +++ b/comfy/text_encoders/chroma.py @@ -0,0 +1,43 @@ +from comfy import sd1_clip +import comfy.text_encoders.t5 +import os +from transformers import T5TokenizerFast + + +class T5XXLModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=False, model_options={}): + textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json") + t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None) + if t5xxl_scaled_fp8 is not None: + model_options = model_options.copy() + model_options["scaled_fp8"] = t5xxl_scaled_fp8 + + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + + +class ChromaT5XXL(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, name="t5xxl", clip_model=T5XXLModel, model_options=model_options) + + +class T5XXLTokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer") + super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_data=tokenizer_data) + + +class ChromaT5Tokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer) + + +def chroma_te(dtype_t5=None, t5xxl_scaled_fp8=None): + class ChromaTEModel_(ChromaT5XXL): + def __init__(self, device="cpu", dtype=None, model_options={}): + if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + model_options = model_options.copy() + model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + if dtype is None: + dtype = dtype_t5 + super().__init__(device=device, dtype=dtype, model_options=model_options) + return ChromaTEModel_ diff --git a/comfy_extras/nodes_optimalsteps.py b/comfy_extras/nodes_optimalsteps.py index f6928199b..102958734 100644 --- a/comfy_extras/nodes_optimalsteps.py +++ b/comfy_extras/nodes_optimalsteps.py @@ -20,13 +20,14 @@ def loglinear_interp(t_steps, num_steps): NOISE_LEVELS = {"FLUX": [0.9968, 0.9886, 0.9819, 0.975, 0.966, 0.9471, 0.9158, 0.8287, 0.5512, 0.2808, 0.001], "Wan":[1.0, 0.997, 0.995, 0.993, 0.991, 0.989, 0.987, 0.985, 0.98, 0.975, 0.973, 0.968, 0.96, 0.946, 0.927, 0.902, 0.864, 0.776, 0.539, 0.208, 0.001], +"Chroma": [0.9919999837875366, 0.9900000095367432, 0.9879999756813049, 0.9850000143051147, 0.9819999933242798, 0.9779999852180481, 0.9729999899864197, 0.9679999947547913, 0.9610000252723694, 0.953000009059906, 0.9430000185966492, 0.9309999942779541, 0.9169999957084656, 0.8999999761581421, 0.8809999823570251, 0.8579999804496765, 0.8320000171661377, 0.8019999861717224, 0.7689999938011169, 0.7310000061988831, 0.6899999976158142, 0.6460000276565552, 0.5989999771118164, 0.550000011920929, 0.5009999871253967, 0.45100000500679016, 0.4020000100135803, 0.35499998927116394, 0.3109999895095825, 0.27000001072883606, 0.23199999332427979, 0.19900000095367432, 0.16899999976158142, 0.14300000667572021, 0.11999999731779099, 0.10100000351667404, 0.08399999886751175, 0.07000000029802322, 0.057999998331069946, 0.04800000041723251, 0.0], } class OptimalStepsScheduler: @classmethod def INPUT_TYPES(s): return {"required": - {"model_type": (["FLUX", "Wan"], ), + {"model_type": (["FLUX", "Wan", "Chroma"], ), "steps": ("INT", {"default": 20, "min": 3, "max": 1000}), "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), } diff --git a/nodes.py b/nodes.py index 73a62d930..f2ced2c35 100644 --- a/nodes.py +++ b/nodes.py @@ -917,7 +917,7 @@ class CLIPLoader: @classmethod def INPUT_TYPES(s): return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream"], ), + "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), From 08ff5fa08a92e0b3f23b9abec979a830a6cffb03 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 30 Apr 2025 20:57:30 -0400 Subject: [PATCH 18/22] Cleanup chroma PR. --- comfy/ldm/chroma/layers.py | 2 +- comfy/ldm/chroma/math.py | 44 --------------------- comfy/lora.py | 2 +- comfy/model_base.py | 63 ++++-------------------------- comfy/model_detection.py | 41 ++++++------------- comfy/sd.py | 6 +-- comfy/supported_models.py | 3 +- comfy/text_encoders/chroma.py | 43 -------------------- comfy_extras/nodes_optimalsteps.py | 2 +- 9 files changed, 25 insertions(+), 181 deletions(-) delete mode 100644 comfy/ldm/chroma/math.py delete mode 100644 comfy/text_encoders/chroma.py diff --git a/comfy/ldm/chroma/layers.py b/comfy/ldm/chroma/layers.py index dd0b72f70..35da91ee2 100644 --- a/comfy/ldm/chroma/layers.py +++ b/comfy/ldm/chroma/layers.py @@ -1,7 +1,7 @@ import torch from torch import Tensor, nn -from .math import attention +from comfy.ldm.flux.math import attention from comfy.ldm.flux.layers import ( MLPEmbedder, RMSNorm, diff --git a/comfy/ldm/chroma/math.py b/comfy/ldm/chroma/math.py deleted file mode 100644 index 36b67931c..000000000 --- a/comfy/ldm/chroma/math.py +++ /dev/null @@ -1,44 +0,0 @@ -import torch -from einops import rearrange -from torch import Tensor - -from comfy.ldm.modules.attention import optimized_attention -import comfy.model_management - - -def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None) -> Tensor: - q_shape = q.shape - k_shape = k.shape - - q = q.float().reshape(*q.shape[:-1], -1, 1, 2) - k = k.float().reshape(*k.shape[:-1], -1, 1, 2) - q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v) - k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v) - - heads = q.shape[1] - x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask) - return x - - -def rope(pos: Tensor, dim: int, theta: int) -> Tensor: - assert dim % 2 == 0 - if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled(): - device = torch.device("cpu") - else: - device = pos.device - - scale = torch.linspace(0, (dim - 2) / dim, steps=dim//2, dtype=torch.float64, device=device) - omega = 1.0 / (theta**scale) - out = torch.einsum("...n,d->...nd", pos.to(dtype=torch.float32, device=device), omega) - out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1) - out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2) - return out.to(dtype=torch.float32, device=pos.device) - - -def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): - xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2) - xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2) - xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1] - xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1] - return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk) - diff --git a/comfy/lora.py b/comfy/lora.py index dbabd3335..fff524be2 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -252,7 +252,7 @@ def model_lora_keys_unet(model, key_map={}): key_lora = k[len("diffusion_model."):-len(".weight")] key_map["base_model.model.{}".format(key_lora)] = k #official hunyuan lora format - if isinstance(model, comfy.model_base.Flux) or isinstance(model, comfy.model_base.Chroma): #Diffusers lora Flux or a diffusers lora Chroma + if isinstance(model, comfy.model_base.Flux): #Diffusers lora Flux diffusers_keys = comfy.utils.flux_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.") for k in diffusers_keys: if k.endswith(".weight"): diff --git a/comfy/model_base.py b/comfy/model_base.py index 1a06bb503..3d33086d8 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -787,8 +787,8 @@ class PixArt(BaseModel): return out class Flux(BaseModel): - def __init__(self, model_config, model_type=ModelType.FLUX, device=None): - super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.flux.model.Flux) + def __init__(self, model_config, model_type=ModelType.FLUX, device=None, unet_model=comfy.ldm.flux.model.Flux): + super().__init__(model_config, model_type, device=device, unet_model=unet_model) def concat_cond(self, **kwargs): try: @@ -1110,63 +1110,14 @@ class HiDream(BaseModel): out['image_cond'] = comfy.conds.CONDNoiseShape(self.process_latent_in(image_cond)) return out -class Chroma(BaseModel): +class Chroma(Flux): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.chroma.model.Chroma) - def concat_cond(self, **kwargs): - try: - #Handle Flux control loras dynamically changing the img_in weight. - num_channels = self.diffusion_model.img_in.weight.shape[1] - except: - #Some cases like tensorrt might not have the weights accessible - num_channels = self.model_config.unet_config["in_channels"] - - out_channels = self.model_config.unet_config["out_channels"] - - if num_channels <= out_channels: - return None - - image = kwargs.get("concat_latent_image", None) - noise = kwargs.get("noise", None) - device = kwargs["device"] - - if image is None: - image = torch.zeros_like(noise) - - image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") - image = utils.resize_to_batch_size(image, noise.shape[0]) - image = self.process_latent_in(image) - if num_channels <= out_channels * 2: - return image - - #inpaint model - mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) - if mask is None: - mask = torch.ones_like(noise)[:, :1] - - mask = torch.mean(mask, dim=1, keepdim=True) - mask = utils.common_upscale(mask.to(device), noise.shape[-1] * 8, noise.shape[-2] * 8, "bilinear", "center") - mask = mask.view(mask.shape[0], mask.shape[2] // 8, 8, mask.shape[3] // 8, 8).permute(0, 2, 4, 1, 3).reshape(mask.shape[0], -1, mask.shape[2] // 8, mask.shape[3] // 8) - mask = utils.resize_to_batch_size(mask, noise.shape[0]) - return torch.cat((image, mask), dim=1) - - def extra_conds(self, **kwargs): out = super().extra_conds(**kwargs) - cross_attn = kwargs.get("cross_attn", None) - if cross_attn is not None: - out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) - # upscale the attention mask, since now we - attention_mask = kwargs.get("attention_mask", None) - if attention_mask is not None: - shape = kwargs["noise"].shape - mask_ref_size = kwargs["attention_mask_img_shape"] - # the model will pad to the patch size, and then divide - # essentially dividing and rounding up - (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size)) - attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok)) - out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) - guidance = 0.0 - out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor((guidance,))) + + guidance = kwargs.get("guidance", 0) + if guidance is not None: + out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance])) return out diff --git a/comfy/model_detection.py b/comfy/model_detection.py index daf6d04e7..9254843ea 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -154,32 +154,6 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["guidance_embed"] = len(guidance_keys) > 0 return dit_config - if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma - dit_config = {} - dit_config["image_model"] = "chroma" - dit_config["depth"] = 48 - dit_config["in_channels"] = 64 - patch_size = 2 - dit_config["patch_size"] = patch_size - in_key = "{}img_in.weight".format(key_prefix) - if in_key in state_dict_keys: - dit_config["in_channels"] = state_dict[in_key].shape[1] - dit_config["out_channels"] = 64 - dit_config["context_in_dim"] = 4096 - dit_config["hidden_size"] = 3072 - dit_config["mlp_ratio"] = 4.0 - dit_config["num_heads"] = 24 - dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.') - dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.') - dit_config["axes_dim"] = [16, 56, 56] - dit_config["theta"] = 10000 - dit_config["qkv_bias"] = True - dit_config["in_dim"] = 64 - dit_config["out_dim"] = 3072 - dit_config["hidden_dim"] = 5120 - dit_config["n_layers"] = 5 - return dit_config - if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and '{}img_in.weight'.format(key_prefix) in state_dict_keys: #Flux dit_config = {} dit_config["image_model"] = "flux" @@ -190,7 +164,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): if in_key in state_dict_keys: dit_config["in_channels"] = state_dict[in_key].shape[1] // (patch_size * patch_size) dit_config["out_channels"] = 16 - dit_config["vec_in_dim"] = 768 + vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix) + if vec_in_key in state_dict_keys: + dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1] dit_config["context_in_dim"] = 4096 dit_config["hidden_size"] = 3072 dit_config["mlp_ratio"] = 4.0 @@ -200,7 +176,16 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["axes_dim"] = [16, 56, 56] dit_config["theta"] = 10000 dit_config["qkv_bias"] = True - dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys + if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma + dit_config["image_model"] = "chroma" + dit_config["in_channels"] = 64 + dit_config["out_channels"] = 64 + dit_config["in_dim"] = 64 + dit_config["out_dim"] = 3072 + dit_config["hidden_dim"] = 5120 + dit_config["n_layers"] = 5 + else: + dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys return dit_config if '{}t5_yproj.weight'.format(key_prefix) in state_dict_keys: #Genmo mochi preview diff --git a/comfy/sd.py b/comfy/sd.py index 454b5929a..da9b36d0e 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -42,7 +42,6 @@ import comfy.text_encoders.cosmos import comfy.text_encoders.lumina2 import comfy.text_encoders.wan import comfy.text_encoders.hidream -import comfy.text_encoders.chroma import comfy.model_patcher import comfy.lora @@ -820,7 +819,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif clip_type == CLIPType.LTXV: clip_target.clip = comfy.text_encoders.lt.ltxv_te(**t5xxl_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.lt.LTXVT5Tokenizer - elif clip_type == CLIPType.PIXART: + elif clip_type == CLIPType.PIXART or clip_type == CLIPType.CHROMA: clip_target.clip = comfy.text_encoders.pixart_t5.pixart_te(**t5xxl_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.pixart_t5.PixArtTokenizer elif clip_type == CLIPType.WAN: @@ -831,9 +830,6 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data), clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None, llama_scaled_fp8=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer - elif clip_type == CLIPType.CHROMA: - clip_target.clip = comfy.text_encoders.chroma.chroma_te(**t5xxl_detect(clip_data)) - clip_target.tokenizer = comfy.text_encoders.chroma.ChromaT5Tokenizer else: #CLIPType.MOCHI clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer diff --git a/comfy/supported_models.py b/comfy/supported_models.py index f03f2790e..d5210cfac 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -17,7 +17,6 @@ import comfy.text_encoders.hunyuan_video import comfy.text_encoders.cosmos import comfy.text_encoders.lumina2 import comfy.text_encoders.wan -import comfy.text_encoders.chroma from . import supported_models_base from . import latent_formats @@ -1095,7 +1094,7 @@ class Chroma(supported_models_base.BASE): def clip_target(self, state_dict={}): pref = self.text_encoder_key_prefix[0] t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref)) - return supported_models_base.ClipTarget(comfy.text_encoders.chroma.ChromaTokenizer, comfy.text_encoders.chroma.chroma_te(**t5_detect)) + return supported_models_base.ClipTarget(comfy.text_encoders.pixart_t5.PixArtTokenizer, comfy.text_encoders.pixart_t5.pixart_te(**t5_detect)) models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma] diff --git a/comfy/text_encoders/chroma.py b/comfy/text_encoders/chroma.py deleted file mode 100644 index aa8dffb25..000000000 --- a/comfy/text_encoders/chroma.py +++ /dev/null @@ -1,43 +0,0 @@ -from comfy import sd1_clip -import comfy.text_encoders.t5 -import os -from transformers import T5TokenizerFast - - -class T5XXLModel(sd1_clip.SDClipModel): - def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=False, model_options={}): - textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json") - t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None) - if t5xxl_scaled_fp8 is not None: - model_options = model_options.copy() - model_options["scaled_fp8"] = t5xxl_scaled_fp8 - - super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) - - -class ChromaT5XXL(sd1_clip.SD1ClipModel): - def __init__(self, device="cpu", dtype=None, model_options={}): - super().__init__(device=device, dtype=dtype, name="t5xxl", clip_model=T5XXLModel, model_options=model_options) - - -class T5XXLTokenizer(sd1_clip.SDTokenizer): - def __init__(self, embedding_directory=None, tokenizer_data={}): - tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer") - super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_data=tokenizer_data) - - -class ChromaT5Tokenizer(sd1_clip.SD1Tokenizer): - def __init__(self, embedding_directory=None, tokenizer_data={}): - super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer) - - -def chroma_te(dtype_t5=None, t5xxl_scaled_fp8=None): - class ChromaTEModel_(ChromaT5XXL): - def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: - model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 - if dtype is None: - dtype = dtype_t5 - super().__init__(device=device, dtype=dtype, model_options=model_options) - return ChromaTEModel_ diff --git a/comfy_extras/nodes_optimalsteps.py b/comfy_extras/nodes_optimalsteps.py index 102958734..e7c851ca2 100644 --- a/comfy_extras/nodes_optimalsteps.py +++ b/comfy_extras/nodes_optimalsteps.py @@ -20,7 +20,7 @@ def loglinear_interp(t_steps, num_steps): NOISE_LEVELS = {"FLUX": [0.9968, 0.9886, 0.9819, 0.975, 0.966, 0.9471, 0.9158, 0.8287, 0.5512, 0.2808, 0.001], "Wan":[1.0, 0.997, 0.995, 0.993, 0.991, 0.989, 0.987, 0.985, 0.98, 0.975, 0.973, 0.968, 0.96, 0.946, 0.927, 0.902, 0.864, 0.776, 0.539, 0.208, 0.001], -"Chroma": [0.9919999837875366, 0.9900000095367432, 0.9879999756813049, 0.9850000143051147, 0.9819999933242798, 0.9779999852180481, 0.9729999899864197, 0.9679999947547913, 0.9610000252723694, 0.953000009059906, 0.9430000185966492, 0.9309999942779541, 0.9169999957084656, 0.8999999761581421, 0.8809999823570251, 0.8579999804496765, 0.8320000171661377, 0.8019999861717224, 0.7689999938011169, 0.7310000061988831, 0.6899999976158142, 0.6460000276565552, 0.5989999771118164, 0.550000011920929, 0.5009999871253967, 0.45100000500679016, 0.4020000100135803, 0.35499998927116394, 0.3109999895095825, 0.27000001072883606, 0.23199999332427979, 0.19900000095367432, 0.16899999976158142, 0.14300000667572021, 0.11999999731779099, 0.10100000351667404, 0.08399999886751175, 0.07000000029802322, 0.057999998331069946, 0.04800000041723251, 0.0], +"Chroma": [0.992, 0.99, 0.988, 0.985, 0.982, 0.978, 0.973, 0.968, 0.961, 0.953, 0.943, 0.931, 0.917, 0.9, 0.881, 0.858, 0.832, 0.802, 0.769, 0.731, 0.69, 0.646, 0.599, 0.55, 0.501, 0.451, 0.402, 0.355, 0.311, 0.27, 0.232, 0.199, 0.169, 0.143, 0.12, 0.101, 0.084, 0.07, 0.058, 0.048, 0.001], } class OptimalStepsScheduler: From c6c19e99803a2538cee17e2f94314f77b7d48e98 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Thu, 1 May 2025 00:24:32 -0700 Subject: [PATCH 19/22] fix bug (#7894) --- comfy_api_nodes/apis/client.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/comfy_api_nodes/apis/client.py b/comfy_api_nodes/apis/client.py index 384e559dc..d3cd9ad2f 100644 --- a/comfy_api_nodes/apis/client.py +++ b/comfy_api_nodes/apis/client.py @@ -297,6 +297,10 @@ class SynchronousOperation(Generic[T, R]): # Convert request model to dict, but use None for EmptyRequest request_dict = None if isinstance(self.request, EmptyRequest) else self.request.model_dump(exclude_none=True) + if request_dict: + for key, value in request_dict.items(): + if isinstance(value, Enum): + request_dict[key] = value.value # Debug log for request logging.debug(f"[DEBUG] API Request: {self.endpoint.method.value} {self.endpoint.path}") From aa9d759df36faa2e34cc5722463749a09a7f529b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 1 May 2025 03:33:42 -0700 Subject: [PATCH 20/22] Switch ltxv to use the pytorch RMSNorm. (#7897) --- comfy/ldm/lightricks/model.py | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/comfy/ldm/lightricks/model.py b/comfy/ldm/lightricks/model.py index 6e8e06181..056e101a4 100644 --- a/comfy/ldm/lightricks/model.py +++ b/comfy/ldm/lightricks/model.py @@ -1,7 +1,6 @@ import torch from torch import nn import comfy.ldm.modules.attention -from comfy.ldm.genmo.joint_model.layers import RMSNorm import comfy.ldm.common_dit from einops import rearrange import math @@ -262,8 +261,8 @@ class CrossAttention(nn.Module): self.heads = heads self.dim_head = dim_head - self.q_norm = RMSNorm(inner_dim, dtype=dtype, device=device) - self.k_norm = RMSNorm(inner_dim, dtype=dtype, device=device) + self.q_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device) + self.k_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device) self.to_q = operations.Linear(query_dim, inner_dim, bias=True, dtype=dtype, device=device) self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device) From 6d32dc049e676a1373dddc4fd912f2ba1311d4cb Mon Sep 17 00:00:00 2001 From: Chenlei Hu Date: Thu, 1 May 2025 10:57:54 -0400 Subject: [PATCH 21/22] Update frontend to v1.18 (#7898) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index f64a05947..74a4ceb02 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.17.11 +comfyui-frontend-package==1.18.5 comfyui-workflow-templates==0.1.3 torch torchsde From 8d0661d0ba35fd54a72e4b49f68d4625febafc10 Mon Sep 17 00:00:00 2001 From: catboxanon <122327233+catboxanon@users.noreply.github.com> Date: Thu, 1 May 2025 19:32:04 -0400 Subject: [PATCH 22/22] Lint instance methods (#7903) --- comfy_extras/nodes_post_processing.py | 1 + comfy_extras/nodes_webcam.py | 2 +- pyproject.toml | 1 + 3 files changed, 3 insertions(+), 1 deletion(-) diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 5b9542015..cb1a0d883 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -141,6 +141,7 @@ class Quantize: CATEGORY = "image/postprocessing" + @staticmethod def bayer(im, pal_im, order): def normalized_bayer_matrix(n): if n == 0: diff --git a/comfy_extras/nodes_webcam.py b/comfy_extras/nodes_webcam.py index 31eddb2d6..062b15cf8 100644 --- a/comfy_extras/nodes_webcam.py +++ b/comfy_extras/nodes_webcam.py @@ -20,7 +20,7 @@ class WebcamCapture(nodes.LoadImage): CATEGORY = "image" - def load_capture(s, image, **kwargs): + def load_capture(self, image, **kwargs): return super().load_image(folder_paths.get_annotated_filepath(image)) diff --git a/pyproject.toml b/pyproject.toml index eadca662e..a9c028c7e 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -12,6 +12,7 @@ documentation = "https://docs.comfy.org/" [tool.ruff] lint.select = [ + "N805", # invalid-first-argument-name-for-method "S307", # suspicious-eval-usage "S102", # exec "T", # print-usage