Merge pull request #391 from huixingyun/main

Update minimal diffusers version

.gitignore

@@ -7,4 +7,5 @@ master_ip
 logs/
 *.DS_Store
 .idea
-*.pt
+*.pt
+tools/

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

cogvideox_fun/pipeline_cogvideox_control.py

@@ -1,866 +0,0 @@
-# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
-# All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import inspect
-import math
-from dataclasses import dataclass
-from typing import Callable, Dict, List, Optional, Tuple, Union
-
-import torch
-import torch.nn.functional as F
-from einops import rearrange
-
-from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
-from diffusers.models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
-from diffusers.models.embeddings import get_3d_rotary_pos_embed
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline
-from diffusers.schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
-from diffusers.utils import BaseOutput, logging, replace_example_docstring
-from diffusers.utils.torch_utils import randn_tensor
-from diffusers.video_processor import VideoProcessor
-from diffusers.image_processor import VaeImageProcessor
-from einops import rearrange
-
-
-logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
-
-
-EXAMPLE_DOC_STRING = """
-    Examples:
-        ```python
-        >>> import torch
-        >>> from diffusers import CogVideoX_Fun_Pipeline
-        >>> from diffusers.utils import export_to_video
-
-        >>> # Models: "THUDM/CogVideoX-2b" or "THUDM/CogVideoX-5b"
-        >>> pipe = CogVideoX_Fun_Pipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16).to("cuda")
-        >>> prompt = (
-        ...     "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
-        ...     "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other "
-        ...     "pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, "
-        ...     "casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. "
-        ...     "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical "
-        ...     "atmosphere of this unique musical performance."
-        ... )
-        >>> video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
-        >>> export_to_video(video, "output.mp4", fps=8)
-        ```
-"""
-
-
-# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid
-def get_resize_crop_region_for_grid(src, tgt_width, tgt_height):
-    tw = tgt_width
-    th = tgt_height
-    h, w = src
-    r = h / w
-    if r > (th / tw):
-        resize_height = th
-        resize_width = int(round(th / h * w))
-    else:
-        resize_width = tw
-        resize_height = int(round(tw / w * h))
-
-    crop_top = int(round((th - resize_height) / 2.0))
-    crop_left = int(round((tw - resize_width) / 2.0))
-
-    return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
-
-
-# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
-def retrieve_timesteps(
-    scheduler,
-    num_inference_steps: Optional[int] = None,
-    device: Optional[Union[str, torch.device]] = None,
-    timesteps: Optional[List[int]] = None,
-    sigmas: Optional[List[float]] = None,
-    **kwargs,
-):
-    """
-    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
-    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
-
-    Args:
-        scheduler (`SchedulerMixin`):
-            The scheduler to get timesteps from.
-        num_inference_steps (`int`):
-            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
-            must be `None`.
-        device (`str` or `torch.device`, *optional*):
-            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
-        timesteps (`List[int]`, *optional*):
-            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
-            `num_inference_steps` and `sigmas` must be `None`.
-        sigmas (`List[float]`, *optional*):
-            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
-            `num_inference_steps` and `timesteps` must be `None`.
-
-    Returns:
-        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
-        second element is the number of inference steps.
-    """
-    if timesteps is not None and sigmas is not None:
-        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
-    if timesteps is not None:
-        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
-        if not accepts_timesteps:
-            raise ValueError(
-                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
-                f" timestep schedules. Please check whether you are using the correct scheduler."
-            )
-        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
-        timesteps = scheduler.timesteps
-        num_inference_steps = len(timesteps)
-    elif sigmas is not None:
-        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
-        if not accept_sigmas:
-            raise ValueError(
-                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
-                f" sigmas schedules. Please check whether you are using the correct scheduler."
-            )
-        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
-        timesteps = scheduler.timesteps
-        num_inference_steps = len(timesteps)
-    else:
-        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
-        timesteps = scheduler.timesteps
-    return timesteps, num_inference_steps
-
-
-@dataclass
-class CogVideoX_Fun_PipelineOutput(BaseOutput):
-    r"""
-    Output class for CogVideo pipelines.
-
-    Args:
-        video (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
-            List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
-            denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
-            `(batch_size, num_frames, channels, height, width)`.
-    """
-
-    videos: torch.Tensor
-
-
-class CogVideoX_Fun_Pipeline_Control(DiffusionPipeline):
-    r"""
-    Pipeline for text-to-video generation using CogVideoX.
-
-    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
-    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
-
-    Args:
-        vae ([`AutoencoderKL`]):
-            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
-        transformer ([`CogVideoXTransformer3DModel`]):
-            A text conditioned `CogVideoXTransformer3DModel` to denoise the encoded video latents.
-        scheduler ([`SchedulerMixin`]):
-            A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
-    """
-
-    _optional_components = []
-    model_cpu_offload_seq = "vae->transformer->vae"
-
-    _callback_tensor_inputs = [
-        "latents",
-        "prompt_embeds",
-        "negative_prompt_embeds",
-    ]
-
-    def __init__(
-        self,
-        vae: AutoencoderKLCogVideoX,
-        transformer: CogVideoXTransformer3DModel,
-        scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
-    ):
-        super().__init__()
-
-        self.register_modules(
-            vae=vae, transformer=transformer, scheduler=scheduler
-        )
-        self.vae_scale_factor_spatial = (
-            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
-        )
-        self.vae_scale_factor_temporal = (
-            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
-        )
-
-        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
-
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
-        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
-        self.mask_processor = VaeImageProcessor(
-            vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
-        )
-
-    def prepare_latents(
-        self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, timesteps, denoise_strength, num_inference_steps,
-         latents=None, freenoise=True, context_size=None, context_overlap=None
-    ):
-        shape = (
-            batch_size,
-            (num_frames - 1) // self.vae_scale_factor_temporal + 1,
-            num_channels_latents,
-            height // self.vae_scale_factor_spatial,
-            width // self.vae_scale_factor_spatial,
-        )
-        if isinstance(generator, list) and len(generator) != batch_size:
-            raise ValueError(
-                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
-                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
-            )
-        noise = randn_tensor(shape, generator=generator, device=torch.device("cpu"), dtype=self.vae.dtype)
-        if freenoise:
-            print("Applying FreeNoise")
-            # code and comments from AnimateDiff-Evolved by Kosinkadink (https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved)
-            video_length = num_frames // 4
-            delta = context_size - context_overlap
-            for start_idx in range(0, video_length-context_size, delta):
-                # start_idx corresponds to the beginning of a context window
-                # goal: place shuffled in the delta region right after the end of the context window
-                #       if space after context window is not enough to place the noise, adjust and finish
-                place_idx = start_idx + context_size
-                # if place_idx is outside the valid indexes, we are already finished
-                if place_idx >= video_length:
-                    break
-                end_idx = place_idx - 1
-                #print("video_length:", video_length, "start_idx:", start_idx, "end_idx:", end_idx, "place_idx:", place_idx, "delta:", delta)
-
-                # if there is not enough room to copy delta amount of indexes, copy limited amount and finish
-                if end_idx + delta >= video_length:
-                    final_delta = video_length - place_idx
-                    # generate list of indexes in final delta region
-                    list_idx = torch.tensor(list(range(start_idx,start_idx+final_delta)), device=torch.device("cpu"), dtype=torch.long)
-                    # shuffle list
-                    list_idx = list_idx[torch.randperm(final_delta, generator=generator)]
-                    # apply shuffled indexes
-                    noise[:, place_idx:place_idx + final_delta, :, :, :] = noise[:, list_idx, :, :, :]
-                    break
-                # otherwise, do normal behavior
-                # generate list of indexes in delta region
-                list_idx = torch.tensor(list(range(start_idx,start_idx+delta)), device=torch.device("cpu"), dtype=torch.long)
-                # shuffle list
-                list_idx = list_idx[torch.randperm(delta, generator=generator)]
-                # apply shuffled indexes
-                #print("place_idx:", place_idx, "delta:", delta, "list_idx:", list_idx)
-                noise[:, place_idx:place_idx + delta, :, :, :] = noise[:, list_idx, :, :, :]
-        if latents is None:
-            latents = noise.to(device)
-        else:
-            latents = latents.to(device)
-            timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, denoise_strength, device)
-            latent_timestep = timesteps[:1]
-            
-            noise = randn_tensor(shape, generator=generator, device=device, dtype=self.vae.dtype)
-            frames_needed = noise.shape[1]
-            current_frames = latents.shape[1]
-            
-            if frames_needed > current_frames:
-                repeat_factor = frames_needed // current_frames
-                additional_frame = torch.randn((latents.size(0), repeat_factor, latents.size(2), latents.size(3), latents.size(4)), dtype=latents.dtype, device=latents.device)
-                latents = torch.cat((latents, additional_frame), dim=1)
-            elif frames_needed < current_frames:
-                latents = latents[:, :frames_needed, :, :, :]
-
-            latents = self.scheduler.add_noise(latents, noise, latent_timestep)
-        latents = latents * self.scheduler.init_noise_sigma # scale the initial noise by the standard deviation required by the scheduler
-        return latents, timesteps, noise
-
-    def prepare_control_latents(
-        self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
-    ):
-        # resize the mask to latents shape as we concatenate the mask to the latents
-        # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
-        # and half precision
-
-        if mask is not None:
-            mask = mask.to(device=device, dtype=self.vae.dtype)
-            bs = 1
-            new_mask = []
-            for i in range(0, mask.shape[0], bs):
-                mask_bs = mask[i : i + bs]
-                mask_bs = self.vae.encode(mask_bs)[0]
-                mask_bs = mask_bs.mode()
-                new_mask.append(mask_bs)
-            mask = torch.cat(new_mask, dim = 0)
-            mask = mask * self.vae.config.scaling_factor
-
-        if masked_image is not None:
-            masked_image = masked_image.to(device=device, dtype=self.vae.dtype)
-            bs = 1
-            new_mask_pixel_values = []
-            for i in range(0, masked_image.shape[0], bs):
-                mask_pixel_values_bs = masked_image[i : i + bs]
-                mask_pixel_values_bs = self.vae.encode(mask_pixel_values_bs)[0]
-                mask_pixel_values_bs = mask_pixel_values_bs.mode()
-                new_mask_pixel_values.append(mask_pixel_values_bs)
-            masked_image_latents = torch.cat(new_mask_pixel_values, dim = 0)
-            masked_image_latents = masked_image_latents * self.vae.config.scaling_factor
-        else:
-            masked_image_latents = None
-
-        return mask, masked_image_latents
-
-    def decode_latents(self, latents: torch.Tensor) -> torch.Tensor:
-        latents = latents.permute(0, 2, 1, 3, 4)  # [batch_size, num_channels, num_frames, height, width]
-        latents = 1 / self.vae.config.scaling_factor * latents
-
-        frames = self.vae.decode(latents).sample
-        frames = (frames / 2 + 0.5).clamp(0, 1)
-        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
-        frames = frames.cpu().float().numpy()
-        return frames
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
-    def prepare_extra_step_kwargs(self, generator, eta):
-        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
-        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
-        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
-        # and should be between [0, 1]
-
-        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
-        extra_step_kwargs = {}
-        if accepts_eta:
-            extra_step_kwargs["eta"] = eta
-
-        # check if the scheduler accepts generator
-        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
-        if accepts_generator:
-            extra_step_kwargs["generator"] = generator
-        return extra_step_kwargs
-
-    # Copied from diffusers.pipelines.latte.pipeline_latte.LattePipeline.check_inputs
-    def check_inputs(
-        self,
-        prompt,
-        height,
-        width,
-        negative_prompt,
-        callback_on_step_end_tensor_inputs,
-        prompt_embeds=None,
-        negative_prompt_embeds=None,
-    ):
-        if height % 8 != 0 or width % 8 != 0:
-            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
-
-        if callback_on_step_end_tensor_inputs is not None and not all(
-            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
-        ):
-            raise ValueError(
-                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
-            )
-        if prompt is not None and prompt_embeds is not None:
-            raise ValueError(
-                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
-                " only forward one of the two."
-            )
-        elif prompt is None and prompt_embeds is None:
-            raise ValueError(
-                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
-            )
-        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
-            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
-
-        if prompt is not None and negative_prompt_embeds is not None:
-            raise ValueError(
-                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
-                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
-            )
-
-        if negative_prompt is not None and negative_prompt_embeds is not None:
-            raise ValueError(
-                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
-                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
-            )
-
-        if prompt_embeds is not None and negative_prompt_embeds is not None:
-            if prompt_embeds.shape != negative_prompt_embeds.shape:
-                raise ValueError(
-                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
-                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
-                    f" {negative_prompt_embeds.shape}."
-                )
-
-    def fuse_qkv_projections(self) -> None:
-        r"""Enables fused QKV projections."""
-        self.fusing_transformer = True
-        self.transformer.fuse_qkv_projections()
-
-    def unfuse_qkv_projections(self) -> None:
-        r"""Disable QKV projection fusion if enabled."""
-        if not self.fusing_transformer:
-            logger.warning("The Transformer was not initially fused for QKV projections. Doing nothing.")
-        else:
-            self.transformer.unfuse_qkv_projections()
-            self.fusing_transformer = False
-
-    def _prepare_rotary_positional_embeddings(
-        self,
-        height: int,
-        width: int,
-        num_frames: int,
-        device: torch.device,
-        start_frame: Optional[int] = None,
-        end_frame: Optional[int] = None,
-        context_frames: Optional[int] = None,
-        ) -> Tuple[torch.Tensor, torch.Tensor]:
-        grid_height = height // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
-        grid_width = width // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
-        base_size_width = 720 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
-        base_size_height = 480 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
-
-        grid_crops_coords = get_resize_crop_region_for_grid(
-            (grid_height, grid_width), base_size_width, base_size_height
-        )
-        freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
-            embed_dim=self.transformer.config.attention_head_dim,
-            crops_coords=grid_crops_coords,
-            grid_size=(grid_height, grid_width),
-            temporal_size=num_frames,
-            use_real=True,
-        )
-        
-        if start_frame is not None or context_frames is not None:
-            freqs_cos = freqs_cos.view(num_frames, grid_height * grid_width, -1)
-            freqs_sin = freqs_sin.view(num_frames, grid_height * grid_width, -1)
-            if context_frames is not None:
-                freqs_cos = freqs_cos[context_frames]
-                freqs_sin = freqs_sin[context_frames]
-            else:
-                freqs_cos = freqs_cos[start_frame:end_frame]
-                freqs_sin = freqs_sin[start_frame:end_frame]
-
-            freqs_cos = freqs_cos.view(-1, freqs_cos.shape[-1])
-            freqs_sin = freqs_sin.view(-1, freqs_sin.shape[-1])
-
-        freqs_cos = freqs_cos.to(device=device)
-        freqs_sin = freqs_sin.to(device=device)
-        return freqs_cos, freqs_sin
-
-    @property
-    def guidance_scale(self):
-        return self._guidance_scale
-
-    @property
-    def num_timesteps(self):
-        return self._num_timesteps
-
-    @property
-    def interrupt(self):
-        return self._interrupt
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
-    def get_timesteps(self, num_inference_steps, strength, device):
-        # get the original timestep using init_timestep
-        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
-
-        t_start = max(num_inference_steps - init_timestep, 0)
-        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
-
-        return timesteps, num_inference_steps - t_start
-
-    @torch.no_grad()
-    @replace_example_docstring(EXAMPLE_DOC_STRING)
-    def __call__(
-        self,
-        prompt: Optional[Union[str, List[str]]] = None,
-        negative_prompt: Optional[Union[str, List[str]]] = None,
-        height: int = 480,
-        width: int = 720,
-        video: Union[torch.FloatTensor] = None,
-        control_video: Union[torch.FloatTensor] = None,
-        num_frames: int = 49,
-        num_inference_steps: int = 50,
-        timesteps: Optional[List[int]] = None,
-        guidance_scale: float = 6,
-        use_dynamic_cfg: bool = False,
-        denoise_strength: float = 1.0,
-        num_videos_per_prompt: int = 1,
-        eta: float = 0.0,
-        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
-        latents: Optional[torch.FloatTensor] = None,
-        prompt_embeds: Optional[torch.FloatTensor] = None,
-        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
-        output_type: str = "numpy",
-        return_dict: bool = False,
-        callback_on_step_end: Optional[
-            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
-        ] = None,
-        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
-        max_sequence_length: int = 226,
-        comfyui_progressbar: bool = False,
-        control_strength: float = 1.0,
-        control_start_percent: float = 0.0,
-        control_end_percent: float = 1.0,
-        scheduler_name: str = "DPM",
-        context_schedule: Optional[str] = None,
-        context_frames: Optional[int] = None,
-        context_stride: Optional[int] = None,
-        context_overlap: Optional[int] = None,
-        freenoise: Optional[bool] = True,
-        tora: Optional[dict] = None,
-    ) -> Union[CogVideoX_Fun_PipelineOutput, Tuple]:
-        """
-        Function invoked when calling the pipeline for generation.
-
-        Args:
-            prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
-                instead.
-            negative_prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts not to guide the image generation. If not defined, one has to pass
-                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
-                less than `1`).
-            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
-                The height in pixels of the generated image. This is set to 1024 by default for the best results.
-            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
-                The width in pixels of the generated image. This is set to 1024 by default for the best results.
-            num_frames (`int`, defaults to `48`):
-                Number of frames to generate. Must be divisible by self.vae_scale_factor_temporal. Generated video will
-                contain 1 extra frame because CogVideoX_Fun is conditioned with (num_seconds * fps + 1) frames where
-                num_seconds is 6 and fps is 4. However, since videos can be saved at any fps, the only condition that
-                needs to be satisfied is that of divisibility mentioned above.
-            num_inference_steps (`int`, *optional*, defaults to 50):
-                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
-                expense of slower inference.
-            timesteps (`List[int]`, *optional*):
-                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
-                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
-                passed will be used. Must be in descending order.
-            guidance_scale (`float`, *optional*, defaults to 7.0):
-                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
-                `guidance_scale` is defined as `w` of equation 2. of [Imagen
-                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
-                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
-                usually at the expense of lower image quality.
-            num_videos_per_prompt (`int`, *optional*, defaults to 1):
-                The number of videos to generate per prompt.
-            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
-                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
-                to make generation deterministic.
-            latents (`torch.FloatTensor`, *optional*):
-                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
-                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
-                tensor will ge generated by sampling using the supplied random `generator`.
-            prompt_embeds (`torch.FloatTensor`, *optional*):
-                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
-                provided, text embeddings will be generated from `prompt` input argument.
-            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
-                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
-                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
-                argument.
-            output_type (`str`, *optional*, defaults to `"pil"`):
-                The output format of the generate image. Choose between
-                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
-            return_dict (`bool`, *optional*, defaults to `True`):
-                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
-                of a plain tuple.
-            callback_on_step_end (`Callable`, *optional*):
-                A function that calls at the end of each denoising steps during the inference. The function is called
-                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
-                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
-                `callback_on_step_end_tensor_inputs`.
-            callback_on_step_end_tensor_inputs (`List`, *optional*):
-                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
-                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
-                `._callback_tensor_inputs` attribute of your pipeline class.
-            max_sequence_length (`int`, defaults to `226`):
-                Maximum sequence length in encoded prompt. Must be consistent with
-                `self.transformer.config.max_text_seq_length` otherwise may lead to poor results.
-
-        Examples:
-
-        Returns:
-            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoX_Fun_PipelineOutput`] or `tuple`:
-            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoX_Fun_PipelineOutput`] if `return_dict` is True, otherwise a
-            `tuple`. When returning a tuple, the first element is a list with the generated images.
-        """
-
-        # if num_frames > 49:
-        #     raise ValueError(
-        #         "The number of frames must be less than 49 for now due to static positional embeddings. This will be updated in the future to remove this limitation."
-        #     )
-
-        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
-            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
-
-        height = height or self.transformer.config.sample_size * self.vae_scale_factor_spatial
-        width = width or self.transformer.config.sample_size * self.vae_scale_factor_spatial
-        num_videos_per_prompt = 1
-
-        # 1. Check inputs. Raise error if not correct
-        self.check_inputs(
-            prompt,
-            height,
-            width,
-            negative_prompt,
-            callback_on_step_end_tensor_inputs,
-            prompt_embeds,
-            negative_prompt_embeds,
-        )
-        self._guidance_scale = guidance_scale
-        self._interrupt = False
-
-        # 2. Default call parameters
-        if prompt is not None and isinstance(prompt, str):
-            batch_size = 1
-        elif prompt is not None and isinstance(prompt, list):
-            batch_size = len(prompt)
-        else:
-            batch_size = prompt_embeds.shape[0]
-
-        device = self._execution_device
-
-        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
-        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
-        # corresponds to doing no classifier free guidance.
-        do_classifier_free_guidance = guidance_scale > 1.0
-
-        if do_classifier_free_guidance:
-            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
-
-        # 4. Prepare timesteps
-        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
-        self._num_timesteps = len(timesteps)
-        if comfyui_progressbar:
-            from comfy.utils import ProgressBar
-            pbar = ProgressBar(num_inference_steps + 2)
-
-        # 5. Prepare latents.
-        latent_channels = self.vae.config.latent_channels
-        latents, timesteps, noise = self.prepare_latents(
-            batch_size * num_videos_per_prompt,
-            latent_channels,
-            num_frames,
-            height,
-            width,
-            self.vae.dtype,
-            device,
-            generator,
-            timesteps,
-            denoise_strength,
-            num_inference_steps,
-            latents,
-            context_size=context_frames,
-            context_overlap=context_overlap,
-            freenoise=freenoise,
-        )
-        if comfyui_progressbar:
-            pbar.update(1)
-
-
-        control_video_latents_input = (
-            torch.cat([control_video] * 2) if do_classifier_free_guidance else control_video
-        )
-        control_latents = rearrange(control_video_latents_input, "b c f h w -> b f c h w")
-
-        control_latents = control_latents * control_strength
-
-        if comfyui_progressbar:
-            pbar.update(1)
-
-        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
-        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
-
-        
-
-        # 8. Denoising loop
-        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
-
-        if context_schedule is not None:
-            print(f"Context schedule enabled: {context_frames} frames, {context_stride} stride, {context_overlap} overlap")
-            use_context_schedule = True
-            from .context import get_context_scheduler
-            context = get_context_scheduler(context_schedule)
-
-        else:
-            use_context_schedule = False
-            print(" context schedule disabled")
-            # 7. Create rotary embeds if required
-            image_rotary_emb = (
-                self._prepare_rotary_positional_embeddings(height, width, latents.size(1), device)
-                if self.transformer.config.use_rotary_positional_embeddings
-                else None
-            )
-            if tora is not None and do_classifier_free_guidance:
-                video_flow_features = tora["video_flow_features"].repeat(1, 2, 1, 1, 1).contiguous()
-
-        if tora is not None:
-            for module in self.transformer.fuser_list:
-                for param in module.parameters():
-                    param.data = param.data.to(device)
-
-        with self.progress_bar(total=num_inference_steps) as progress_bar:
-            # for DPM-solver++
-            old_pred_original_sample = None
-            for i, t in enumerate(timesteps):
-                if self.interrupt:
-                    continue
-                if use_context_schedule:
-                    
-                    latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
-                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
-
-                    # Calculate the current step percentage
-                    current_step_percentage = i / num_inference_steps
-
-                    # Determine if control_latents should be applied
-                    apply_control = control_start_percent <= current_step_percentage <= control_end_percent
-                    current_control_latents = control_latents if apply_control else torch.zeros_like(control_latents)
-
-                    # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
-                    timestep = t.expand(latent_model_input.shape[0])
-
-                    context_queue = list(context(
-                        i, num_inference_steps, latents.shape[1], context_frames, context_stride, context_overlap,
-                    ))
-                    counter = torch.zeros_like(latent_model_input)
-                    noise_pred = torch.zeros_like(latent_model_input)
-
-                    image_rotary_emb = (
-                            self._prepare_rotary_positional_embeddings(height, width, context_frames, device)
-                            if self.transformer.config.use_rotary_positional_embeddings
-                            else None
-                        )
-
-                    for c in context_queue:
-                        partial_latent_model_input = latent_model_input[:, c, :, :, :]
-                        partial_control_latents = current_control_latents[:, c, :, :, :]
-
-                        # predict noise model_output
-                        noise_pred[:, c, :, :, :] += self.transformer(
-                            hidden_states=partial_latent_model_input,
-                            encoder_hidden_states=prompt_embeds,
-                            timestep=timestep,
-                            image_rotary_emb=image_rotary_emb,
-                            return_dict=False,
-                            control_latents=partial_control_latents,
-                        )[0]
-
-                        counter[:, c, :, :, :] += 1
-                        noise_pred = noise_pred.float()
-                        
-                    noise_pred /= counter
-                    if do_classifier_free_guidance:
-                        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-                        noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
-                       
-                    # compute the previous noisy sample x_t -> x_t-1
-                    if not isinstance(self.scheduler, CogVideoXDPMScheduler):
-                        latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
-                    else:
-                        latents, old_pred_original_sample = self.scheduler.step(
-                            noise_pred,
-                            old_pred_original_sample,
-                            t,
-                            timesteps[i - 1] if i > 0 else None,
-                            latents,
-                            **extra_step_kwargs,
-                            return_dict=False,
-                        )
-                    latents = latents.to(prompt_embeds.dtype)
-
-                    # call the callback, if provided
-                    if callback_on_step_end is not None:
-                        callback_kwargs = {}
-                        for k in callback_on_step_end_tensor_inputs:
-                            callback_kwargs[k] = locals()[k]
-                        callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
-
-                        latents = callback_outputs.pop("latents", latents)
-                        prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
-                        negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
-
-                    if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
-                        progress_bar.update()
-                    if comfyui_progressbar:
-                        pbar.update(1)
-                else:
-                    latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
-                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
-
-                    # Calculate the current step percentage
-                    current_step_percentage = i / num_inference_steps
-
-                    # Determine if control_latents should be applied
-                    apply_control = control_start_percent <= current_step_percentage <= control_end_percent
-                    current_control_latents = control_latents if apply_control else torch.zeros_like(control_latents)
-
-                    # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
-                    timestep = t.expand(latent_model_input.shape[0])
-
-                    # predict noise model_output
-                    noise_pred = self.transformer(
-                        hidden_states=latent_model_input,
-                        encoder_hidden_states=prompt_embeds,
-                        timestep=timestep,
-                        image_rotary_emb=image_rotary_emb,
-                        return_dict=False,
-                        control_latents=current_control_latents,
-                        video_flow_features=video_flow_features if (tora is not None and tora["start_percent"] <= current_step_percentage <= tora["end_percent"]) else None,
-
-                    )[0]
-                    noise_pred = noise_pred.float()
-
-                    if do_classifier_free_guidance:
-                        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-                        noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
-
-                    # compute the previous noisy sample x_t -> x_t-1
-                    if not isinstance(self.scheduler, CogVideoXDPMScheduler):
-                        latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
-                    else:
-                        latents, old_pred_original_sample = self.scheduler.step(
-                            noise_pred,
-                            old_pred_original_sample,
-                            t,
-                            timesteps[i - 1] if i > 0 else None,
-                            latents,
-                            **extra_step_kwargs,
-                            return_dict=False,
-                        )
-                    latents = latents.to(prompt_embeds.dtype)
-
-                    # call the callback, if provided
-                    if callback_on_step_end is not None:
-                        callback_kwargs = {}
-                        for k in callback_on_step_end_tensor_inputs:
-                            callback_kwargs[k] = locals()[k]
-                        callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
-
-                        latents = callback_outputs.pop("latents", latents)
-                        prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
-                        negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
-
-                    if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
-                        progress_bar.update()
-                    if comfyui_progressbar:
-                        pbar.update(1)
-
-        # if output_type == "numpy":
-        #     video = self.decode_latents(latents)
-        # elif not output_type == "latent":
-        #     video = self.decode_latents(latents)
-        #     video = self.video_processor.postprocess_video(video=video, output_type=output_type)
-        # else:
-        #     video = latents
-
-        # Offload all models
-        self.maybe_free_model_hooks()
-
-        # if not return_dict:
-        #     video = torch.from_numpy(video)
-
-        return latents

cogvideox_fun/transformer_3d.py

@@ -1,823 +0,0 @@
-# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
-# All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from typing import Any, Dict, Optional, Tuple, Union
-
-import os
-import json
-import torch
-import glob
-import torch.nn.functional as F
-from torch import nn
-
-from diffusers.configuration_utils import ConfigMixin, register_to_config
-from diffusers.utils import is_torch_version, logging
-from diffusers.utils.torch_utils import maybe_allow_in_graph
-from diffusers.models.attention import Attention, FeedForward
-from diffusers.models.attention_processor import AttentionProcessor#, CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
-from diffusers.models.embeddings import TimestepEmbedding, Timesteps, get_3d_sincos_pos_embed
-from diffusers.models.modeling_outputs import Transformer2DModelOutput
-from diffusers.models.modeling_utils import ModelMixin
-from diffusers.models.normalization import AdaLayerNorm, CogVideoXLayerNormZero
-
-logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
-
-from einops import rearrange
-try:
-    from sageattention import sageattn
-    SAGEATTN_IS_AVAILABLE = True
-except:
-    SAGEATTN_IS_AVAILABLE = False
-
-def fft(tensor):
-    tensor_fft = torch.fft.fft2(tensor)
-    tensor_fft_shifted = torch.fft.fftshift(tensor_fft)
-    B, C, H, W = tensor.size()
-    radius = min(H, W) // 5
-            
-    Y, X = torch.meshgrid(torch.arange(H), torch.arange(W))
-    center_x, center_y = W // 2, H // 2
-    mask = (X - center_x) ** 2 + (Y - center_y) ** 2 <= radius ** 2
-    low_freq_mask = mask.unsqueeze(0).unsqueeze(0).to(tensor.device)
-    high_freq_mask = ~low_freq_mask
-            
-    low_freq_fft = tensor_fft_shifted * low_freq_mask
-    high_freq_fft = tensor_fft_shifted * high_freq_mask
-
-    return low_freq_fft, high_freq_fft
-
-class CogVideoXAttnProcessor2_0:
-    r"""
-    Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
-    query and key vectors, but does not include spatial normalization.
-    """
-
-    def __init__(self):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        image_rotary_emb: Optional[torch.Tensor] = None,
-        attention_mode: Optional[str] = None,
-    ) -> torch.Tensor:
-        text_seq_length = encoder_hidden_states.size(1)
-
-        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
-
-        batch_size, sequence_length, _ = (
-            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-        )
-
-        if attention_mask is not None:
-            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
-            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
-
-        query = attn.to_q(hidden_states)
-        key = attn.to_k(hidden_states)
-        value = attn.to_v(hidden_states)
-
-        inner_dim = key.shape[-1]
-        head_dim = inner_dim // attn.heads
-
-        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-
-        if attn.norm_q is not None:
-            query = attn.norm_q(query)
-        if attn.norm_k is not None:
-            key = attn.norm_k(key)
-
-        # Apply RoPE if needed
-        if image_rotary_emb is not None:
-            from diffusers.models.embeddings import apply_rotary_emb
-
-            query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
-            if not attn.is_cross_attention:
-                key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
-                 
-        if attention_mode == "sageattn":
-            if SAGEATTN_IS_AVAILABLE:
-               hidden_states = sageattn(query, key, value, attn_mask=attention_mask, dropout_p=0.0,is_causal=False)
-            else:
-                raise ImportError("sageattn not found")
-        else:
-            hidden_states = F.scaled_dot_product_attention(
-            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
-            )
-
-        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
-
-        # linear proj
-        hidden_states = attn.to_out[0](hidden_states)
-        # dropout
-        hidden_states = attn.to_out[1](hidden_states)
-
-        encoder_hidden_states, hidden_states = hidden_states.split(
-            [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
-        )
-        return hidden_states, encoder_hidden_states
-    
-class CogVideoXPatchEmbed(nn.Module):
-    def __init__(
-        self,
-        patch_size: int = 2,
-        in_channels: int = 16,
-        embed_dim: int = 1920,
-        text_embed_dim: int = 4096,
-        bias: bool = True,
-    ) -> None:
-        super().__init__()
-        self.patch_size = patch_size
-
-        self.proj = nn.Conv2d(
-            in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
-        )
-        self.text_proj = nn.Linear(text_embed_dim, embed_dim)
-
-    def forward(self, text_embeds: torch.Tensor, image_embeds: torch.Tensor):
-        r"""
-        Args:
-            text_embeds (`torch.Tensor`):
-                Input text embeddings. Expected shape: (batch_size, seq_length, embedding_dim).
-            image_embeds (`torch.Tensor`):
-                Input image embeddings. Expected shape: (batch_size, num_frames, channels, height, width).
-        """
-        text_embeds = self.text_proj(text_embeds)
-
-        batch, num_frames, channels, height, width = image_embeds.shape
-        image_embeds = image_embeds.reshape(-1, channels, height, width)
-        image_embeds = self.proj(image_embeds)
-        image_embeds = image_embeds.view(batch, num_frames, *image_embeds.shape[1:])
-        image_embeds = image_embeds.flatten(3).transpose(2, 3)  # [batch, num_frames, height x width, channels]
-        image_embeds = image_embeds.flatten(1, 2)  # [batch, num_frames x height x width, channels]
-
-        embeds = torch.cat(
-            [text_embeds, image_embeds], dim=1
-        ).contiguous()  # [batch, seq_length + num_frames x height x width, channels]
-        return embeds
-
-@maybe_allow_in_graph
-class CogVideoXBlock(nn.Module):
-    r"""
-    Transformer block used in [CogVideoX](https://github.com/THUDM/CogVideo) model.
-
-    Parameters:
-        dim (`int`):
-            The number of channels in the input and output.
-        num_attention_heads (`int`):
-            The number of heads to use for multi-head attention.
-        attention_head_dim (`int`):
-            The number of channels in each head.
-        time_embed_dim (`int`):
-            The number of channels in timestep embedding.
-        dropout (`float`, defaults to `0.0`):
-            The dropout probability to use.
-        activation_fn (`str`, defaults to `"gelu-approximate"`):
-            Activation function to be used in feed-forward.
-        attention_bias (`bool`, defaults to `False`):
-            Whether or not to use bias in attention projection layers.
-        qk_norm (`bool`, defaults to `True`):
-            Whether or not to use normalization after query and key projections in Attention.
-        norm_elementwise_affine (`bool`, defaults to `True`):
-            Whether to use learnable elementwise affine parameters for normalization.
-        norm_eps (`float`, defaults to `1e-5`):
-            Epsilon value for normalization layers.
-        final_dropout (`bool` defaults to `False`):
-            Whether to apply a final dropout after the last feed-forward layer.
-        ff_inner_dim (`int`, *optional*, defaults to `None`):
-            Custom hidden dimension of Feed-forward layer. If not provided, `4 * dim` is used.
-        ff_bias (`bool`, defaults to `True`):
-            Whether or not to use bias in Feed-forward layer.
-        attention_out_bias (`bool`, defaults to `True`):
-            Whether or not to use bias in Attention output projection layer.
-    """
-
-    def __init__(
-        self,
-        dim: int,
-        num_attention_heads: int,
-        attention_head_dim: int,
-        time_embed_dim: int,
-        dropout: float = 0.0,
-        activation_fn: str = "gelu-approximate",
-        attention_bias: bool = False,
-        qk_norm: bool = True,
-        norm_elementwise_affine: bool = True,
-        norm_eps: float = 1e-5,
-        final_dropout: bool = True,
-        ff_inner_dim: Optional[int] = None,
-        ff_bias: bool = True,
-        attention_out_bias: bool = True,
-        attention_mode: Optional[str] = None,
-    ):
-        super().__init__()
-
-        # 1. Self Attention
-        self.norm1 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)
-
-        self.attn1 = Attention(
-            query_dim=dim,
-            dim_head=attention_head_dim,
-            heads=num_attention_heads,
-            qk_norm="layer_norm" if qk_norm else None,
-            eps=1e-6,
-            bias=attention_bias,
-            out_bias=attention_out_bias,
-            processor=CogVideoXAttnProcessor2_0(),
-        )
-
-        # 2. Feed Forward
-        self.norm2 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)
-
-        self.ff = FeedForward(
-            dim,
-            dropout=dropout,
-            activation_fn=activation_fn,
-            final_dropout=final_dropout,
-            inner_dim=ff_inner_dim,
-            bias=ff_bias,
-        )
-        self.cached_hidden_states = []
-        self.cached_encoder_hidden_states = []
-        self.attention_mode = attention_mode
-        
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: torch.Tensor,
-        temb: torch.Tensor,
-        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
-        video_flow_feature: Optional[torch.Tensor] = None,
-        fuser=None,
-        block_use_fastercache=False,
-        fastercache_counter=0,
-        fastercache_start_step=15,
-        fastercache_device="cuda:0",
-    ) -> torch.Tensor:
-        text_seq_length = encoder_hidden_states.size(1)
-
-        # norm & modulate
-        norm_hidden_states, norm_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(
-            hidden_states, encoder_hidden_states, temb
-        )
-        # Tora Motion-guidance Fuser
-        if video_flow_feature is not None:
-            H, W = video_flow_feature.shape[-2:]
-            T = norm_hidden_states.shape[1] // H // W
-            h = rearrange(norm_hidden_states, "B (T H W) C -> (B T) C H W", H=H, W=W)
-            h = fuser(h, video_flow_feature.to(h), T=T)
-            norm_hidden_states = rearrange(h, "(B T) C H W ->  B (T H W) C", T=T)
-            del h, fuser        
-        
-        #region fastercache
-        if block_use_fastercache:
-            B = norm_hidden_states.shape[0]
-            if fastercache_counter >= fastercache_start_step + 3 and fastercache_counter%3!=0 and self.cached_hidden_states[-1].shape[0] >= B:
-                attn_hidden_states = (
-                    self.cached_hidden_states[1][:B] + 
-                    (self.cached_hidden_states[1][:B] - self.cached_hidden_states[0][:B]) 
-                    * 0.3
-                    ).to(norm_hidden_states.device, non_blocking=True)
-                attn_encoder_hidden_states = (
-                    self.cached_encoder_hidden_states[1][:B] + 
-                    (self.cached_encoder_hidden_states[1][:B] - self.cached_encoder_hidden_states[0][:B])
-                    * 0.3
-                    ).to(norm_hidden_states.device, non_blocking=True)
-            else:
-                attn_hidden_states, attn_encoder_hidden_states = self.attn1(
-                    hidden_states=norm_hidden_states,
-                    encoder_hidden_states=norm_encoder_hidden_states,
-                    image_rotary_emb=image_rotary_emb,
-                    attention_mode=self.attention_mode,
-                )
-                if fastercache_counter == fastercache_start_step:
-                    self.cached_hidden_states = [attn_hidden_states.to(fastercache_device), attn_hidden_states.to(fastercache_device)]
-                    self.cached_encoder_hidden_states = [attn_encoder_hidden_states.to(fastercache_device), attn_encoder_hidden_states.to(fastercache_device)]
-                elif fastercache_counter > fastercache_start_step:
-                    self.cached_hidden_states[-1].copy_(attn_hidden_states.to(fastercache_device))
-                    self.cached_encoder_hidden_states[-1].copy_(attn_encoder_hidden_states.to(fastercache_device))
-        else:
-            attn_hidden_states, attn_encoder_hidden_states = self.attn1(
-                hidden_states=norm_hidden_states,
-                encoder_hidden_states=norm_encoder_hidden_states,
-                image_rotary_emb=image_rotary_emb,
-                attention_mode=self.attention_mode,
-            )
-        
-        hidden_states = hidden_states + gate_msa * attn_hidden_states
-        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder_hidden_states
-
-        # norm & modulate
-        norm_hidden_states, norm_encoder_hidden_states, gate_ff, enc_gate_ff = self.norm2(
-            hidden_states, encoder_hidden_states, temb
-        )
-
-        # feed-forward
-        norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
-        ff_output = self.ff(norm_hidden_states)
-
-        hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
-        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
-
-        return hidden_states, encoder_hidden_states
-
-
-class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
-    """
-    A Transformer model for video-like data in [CogVideoX](https://github.com/THUDM/CogVideo).
-
-    Parameters:
-        num_attention_heads (`int`, defaults to `30`):
-            The number of heads to use for multi-head attention.
-        attention_head_dim (`int`, defaults to `64`):
-            The number of channels in each head.
-        in_channels (`int`, defaults to `16`):
-            The number of channels in the input.
-        out_channels (`int`, *optional*, defaults to `16`):
-            The number of channels in the output.
-        flip_sin_to_cos (`bool`, defaults to `True`):
-            Whether to flip the sin to cos in the time embedding.
-        time_embed_dim (`int`, defaults to `512`):
-            Output dimension of timestep embeddings.
-        text_embed_dim (`int`, defaults to `4096`):
-            Input dimension of text embeddings from the text encoder.
-        num_layers (`int`, defaults to `30`):
-            The number of layers of Transformer blocks to use.
-        dropout (`float`, defaults to `0.0`):
-            The dropout probability to use.
-        attention_bias (`bool`, defaults to `True`):
-            Whether or not to use bias in the attention projection layers.
-        sample_width (`int`, defaults to `90`):
-            The width of the input latents.
-        sample_height (`int`, defaults to `60`):
-            The height of the input latents.
-        sample_frames (`int`, defaults to `49`):
-            The number of frames in the input latents. Note that this parameter was incorrectly initialized to 49
-            instead of 13 because CogVideoX processed 13 latent frames at once in its default and recommended settings,
-            but cannot be changed to the correct value to ensure backwards compatibility. To create a transformer with
-            K latent frames, the correct value to pass here would be: ((K - 1) * temporal_compression_ratio + 1).
-        patch_size (`int`, defaults to `2`):
-            The size of the patches to use in the patch embedding layer.
-        temporal_compression_ratio (`int`, defaults to `4`):
-            The compression ratio across the temporal dimension. See documentation for `sample_frames`.
-        max_text_seq_length (`int`, defaults to `226`):
-            The maximum sequence length of the input text embeddings.
-        activation_fn (`str`, defaults to `"gelu-approximate"`):
-            Activation function to use in feed-forward.
-        timestep_activation_fn (`str`, defaults to `"silu"`):
-            Activation function to use when generating the timestep embeddings.
-        norm_elementwise_affine (`bool`, defaults to `True`):
-            Whether or not to use elementwise affine in normalization layers.
-        norm_eps (`float`, defaults to `1e-5`):
-            The epsilon value to use in normalization layers.
-        spatial_interpolation_scale (`float`, defaults to `1.875`):
-            Scaling factor to apply in 3D positional embeddings across spatial dimensions.
-        temporal_interpolation_scale (`float`, defaults to `1.0`):
-            Scaling factor to apply in 3D positional embeddings across temporal dimensions.
-    """
-
-    _supports_gradient_checkpointing = True
-
-    @register_to_config
-    def __init__(
-        self,
-        num_attention_heads: int = 30,
-        attention_head_dim: int = 64,
-        in_channels: int = 16,
-        out_channels: Optional[int] = 16,
-        flip_sin_to_cos: bool = True,
-        freq_shift: int = 0,
-        time_embed_dim: int = 512,
-        text_embed_dim: int = 4096,
-        num_layers: int = 30,
-        dropout: float = 0.0,
-        attention_bias: bool = True,
-        sample_width: int = 90,
-        sample_height: int = 60,
-        sample_frames: int = 49,
-        patch_size: int = 2,
-        temporal_compression_ratio: int = 4,
-        max_text_seq_length: int = 226,
-        activation_fn: str = "gelu-approximate",
-        timestep_activation_fn: str = "silu",
-        norm_elementwise_affine: bool = True,
-        norm_eps: float = 1e-5,
-        spatial_interpolation_scale: float = 1.875,
-        temporal_interpolation_scale: float = 1.0,
-        use_rotary_positional_embeddings: bool = False,
-        add_noise_in_inpaint_model: bool = False,
-        attention_mode: Optional[str] = None,
-    ):
-        super().__init__()
-        inner_dim = num_attention_heads * attention_head_dim
-
-        post_patch_height = sample_height // patch_size
-        post_patch_width = sample_width // patch_size
-        post_time_compression_frames = (sample_frames - 1) // temporal_compression_ratio + 1
-        self.num_patches = post_patch_height * post_patch_width * post_time_compression_frames
-        self.post_patch_height = post_patch_height
-        self.post_patch_width = post_patch_width
-        self.post_time_compression_frames = post_time_compression_frames
-        self.patch_size = patch_size
-
-        # 1. Patch embedding
-        self.patch_embed = CogVideoXPatchEmbed(patch_size, in_channels, inner_dim, text_embed_dim, bias=True)
-        self.embedding_dropout = nn.Dropout(dropout)
-
-        # 2. 3D positional embeddings
-        spatial_pos_embedding = get_3d_sincos_pos_embed(
-            inner_dim,
-            (post_patch_width, post_patch_height),
-            post_time_compression_frames,
-            spatial_interpolation_scale,
-            temporal_interpolation_scale,
-        )
-        spatial_pos_embedding = torch.from_numpy(spatial_pos_embedding).flatten(0, 1)
-        pos_embedding = torch.zeros(1, max_text_seq_length + self.num_patches, inner_dim, requires_grad=False)
-        pos_embedding.data[:, max_text_seq_length:].copy_(spatial_pos_embedding)
-        self.register_buffer("pos_embedding", pos_embedding, persistent=False)
-
-        # 3. Time embeddings
-        self.time_proj = Timesteps(inner_dim, flip_sin_to_cos, freq_shift)
-        self.time_embedding = TimestepEmbedding(inner_dim, time_embed_dim, timestep_activation_fn)
-
-        # 4. Define spatio-temporal transformers blocks
-        self.transformer_blocks = nn.ModuleList(
-            [
-                CogVideoXBlock(
-                    dim=inner_dim,
-                    num_attention_heads=num_attention_heads,
-                    attention_head_dim=attention_head_dim,
-                    time_embed_dim=time_embed_dim,
-                    dropout=dropout,
-                    activation_fn=activation_fn,
-                    attention_bias=attention_bias,
-                    norm_elementwise_affine=norm_elementwise_affine,
-                    norm_eps=norm_eps,
-                )
-                for _ in range(num_layers)
-            ]
-        )
-        self.norm_final = nn.LayerNorm(inner_dim, norm_eps, norm_elementwise_affine)
-
-        # 5. Output blocks
-        self.norm_out = AdaLayerNorm(
-            embedding_dim=time_embed_dim,
-            output_dim=2 * inner_dim,
-            norm_elementwise_affine=norm_elementwise_affine,
-            norm_eps=norm_eps,
-            chunk_dim=1,
-        )
-        self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels)
-
-        self.gradient_checkpointing = False
-
-        self.fuser_list = None
-
-        self.use_fastercache = False
-        self.fastercache_counter = 0
-        self.fastercache_start_step = 15
-        self.fastercache_lf_step = 40
-        self.fastercache_hf_step = 30
-        self.fastercache_device = "cuda"
-        self.fastercache_num_blocks_to_cache = len(self.transformer_blocks)
-        self.attention_mode = attention_mode
-
-    def _set_gradient_checkpointing(self, module, value=False):
-        self.gradient_checkpointing = value
-
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
-            if hasattr(module, "set_processor"):
-                if not isinstance(processor, dict):
-                    module.set_processor(processor)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedCogVideoXAttnProcessor2_0
-    def fuse_qkv_projections(self):
-        """
-        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
-        are fused. For cross-attention modules, key and value projection matrices are fused.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-        """
-        self.original_attn_processors = None
-
-        for _, attn_processor in self.attn_processors.items():
-            if "Added" in str(attn_processor.__class__.__name__):
-                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
-
-        self.original_attn_processors = self.attn_processors
-
-        for module in self.modules():
-            if isinstance(module, Attention):
-                module.fuse_projections(fuse=True)
-
-        self.set_attn_processor(FusedCogVideoXAttnProcessor2_0())
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
-    def unfuse_qkv_projections(self):
-        """Disables the fused QKV projection if enabled.
-
-        <Tip warning={true}>
-
-        This API is 🧪 experimental.
-
-        </Tip>
-
-        """
-        if self.original_attn_processors is not None:
-            self.set_attn_processor(self.original_attn_processors)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: torch.Tensor,
-        timestep: Union[int, float, torch.LongTensor],
-        timestep_cond: Optional[torch.Tensor] = None,
-        inpaint_latents: Optional[torch.Tensor] = None,
-        control_latents: Optional[torch.Tensor] = None,
-        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
-        video_flow_features: Optional[torch.Tensor] = None,
-        return_dict: bool = True,
-    ):
-        batch_size, num_frames, channels, height, width = hidden_states.shape
-
-        # 1. Time embedding
-        timesteps = timestep
-        t_emb = self.time_proj(timesteps)
-
-        # timesteps does not contain any weights and will always return f32 tensors
-        # but time_embedding might actually be running in fp16. so we need to cast here.
-        # there might be better ways to encapsulate this.
-        t_emb = t_emb.to(dtype=hidden_states.dtype)
-        emb = self.time_embedding(t_emb, timestep_cond)
-
-        # 2. Patch embedding
-        if inpaint_latents is not None:
-            hidden_states = torch.concat([hidden_states, inpaint_latents], 2)
-        if control_latents is not None:
-            hidden_states = torch.concat([hidden_states, control_latents], 2)
-        hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
-
-        # 3. Position embedding
-        text_seq_length = encoder_hidden_states.shape[1]
-        if not self.config.use_rotary_positional_embeddings:
-            seq_length = height * width * num_frames // (self.config.patch_size**2)
-            # pos_embeds = self.pos_embedding[:, : text_seq_length + seq_length]
-            pos_embeds = self.pos_embedding
-            emb_size = hidden_states.size()[-1]
-            pos_embeds_without_text = pos_embeds[:, text_seq_length: ].view(1, self.post_time_compression_frames, self.post_patch_height, self.post_patch_width, emb_size)
-            pos_embeds_without_text = pos_embeds_without_text.permute([0, 4, 1, 2, 3])
-            pos_embeds_without_text = F.interpolate(pos_embeds_without_text,size=[self.post_time_compression_frames, height // self.config.patch_size, width // self.config.patch_size],mode='trilinear',align_corners=False)
-            pos_embeds_without_text = pos_embeds_without_text.permute([0, 2, 3, 4, 1]).view(1, -1, emb_size)
-            pos_embeds = torch.cat([pos_embeds[:, :text_seq_length], pos_embeds_without_text], dim = 1)
-            pos_embeds = pos_embeds[:, : text_seq_length + seq_length]
-            hidden_states = hidden_states + pos_embeds
-            hidden_states = self.embedding_dropout(hidden_states)
-
-        encoder_hidden_states = hidden_states[:, :text_seq_length]
-        hidden_states = hidden_states[:, text_seq_length:]
-
-        if self.use_fastercache:
-            self.fastercache_counter+=1
-        if self.fastercache_counter >= self.fastercache_start_step + 3 and self.fastercache_counter % 5 !=0:
-            # 4. Transformer blocks
-            for i, block in enumerate(self.transformer_blocks):
-                hidden_states, encoder_hidden_states = block(
-                    hidden_states=hidden_states[:1],
-                    encoder_hidden_states=encoder_hidden_states[:1],
-                    temb=emb[:1],
-                    image_rotary_emb=image_rotary_emb,
-                    video_flow_feature=video_flow_features[i][:1] if video_flow_features is not None else None,
-                    fuser = self.fuser_list[i] if self.fuser_list is not None else None,
-                    block_use_fastercache = i <= self.fastercache_num_blocks_to_cache,
-                    fastercache_start_step = self.fastercache_start_step,
-                    fastercache_counter = self.fastercache_counter,
-                    fastercache_device = self.fastercache_device
-                )
-
-            if not self.config.use_rotary_positional_embeddings:
-                # CogVideoX-2B
-                hidden_states = self.norm_final(hidden_states)
-            else:
-                # CogVideoX-5B
-                hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
-                hidden_states = self.norm_final(hidden_states)
-                hidden_states = hidden_states[:, text_seq_length:]
-
-            # 5. Final block
-            hidden_states = self.norm_out(hidden_states, temb=emb[:1])
-            hidden_states = self.proj_out(hidden_states)
-
-            # 6. Unpatchify
-            p = self.config.patch_size
-            output = hidden_states.reshape(1, num_frames, height // p, width // p, channels, p, p)
-            output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
-
-            (bb, tt, cc, hh, ww) = output.shape
-            cond = rearrange(output, "B T C H W -> (B T) C H W", B=bb, C=cc, T=tt, H=hh, W=ww)
-            lf_c, hf_c = fft(cond.float())
-            #lf_step = 40
-            #hf_step = 30
-            if self.fastercache_counter <= self.fastercache_lf_step:
-                self.delta_lf = self.delta_lf * 1.1
-            if self.fastercache_counter >= self.fastercache_hf_step:
-                self.delta_hf = self.delta_hf * 1.1
-   
-            new_hf_uc = self.delta_hf + hf_c
-            new_lf_uc = self.delta_lf + lf_c
-
-            combine_uc = new_lf_uc + new_hf_uc
-            combined_fft = torch.fft.ifftshift(combine_uc)
-            recovered_uncond = torch.fft.ifft2(combined_fft).real
-            recovered_uncond = rearrange(recovered_uncond.to(output.dtype), "(B T) C H W -> B T C H W", B=bb, C=cc, T=tt, H=hh, W=ww)
-            output = torch.cat([output, recovered_uncond])
-        else:
-            # 4. Transformer blocks
-            for i, block in enumerate(self.transformer_blocks):
-                hidden_states, encoder_hidden_states = block(
-                    hidden_states=hidden_states,
-                    encoder_hidden_states=encoder_hidden_states,
-                    temb=emb,
-                    image_rotary_emb=image_rotary_emb,
-                    video_flow_feature=video_flow_features[i] if video_flow_features is not None else None,
-                    fuser = self.fuser_list[i] if self.fuser_list is not None else None,
-                    block_use_fastercache = i <= self.fastercache_num_blocks_to_cache,
-                    fastercache_counter = self.fastercache_counter,
-                    fastercache_start_step = self.fastercache_start_step,
-                    fastercache_device = self.fastercache_device
-                )
-
-            if not self.config.use_rotary_positional_embeddings:
-                # CogVideoX-2B
-                hidden_states = self.norm_final(hidden_states)
-            else:
-                # CogVideoX-5B
-                hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
-                hidden_states = self.norm_final(hidden_states)
-                hidden_states = hidden_states[:, text_seq_length:]
-
-            # 5. Final block
-            hidden_states = self.norm_out(hidden_states, temb=emb)
-            hidden_states = self.proj_out(hidden_states)
-
-            # 6. Unpatchify
-            p = self.config.patch_size
-            output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, channels, p, p)
-            output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
-
-            if self.fastercache_counter >= self.fastercache_start_step + 1: 
-                (bb, tt, cc, hh, ww) = output.shape
-                cond = rearrange(output[0:1].float(), "B T C H W -> (B T) C H W", B=bb//2, C=cc, T=tt, H=hh, W=ww)
-                uncond = rearrange(output[1:2].float(), "B T C H W -> (B T) C H W", B=bb//2, C=cc, T=tt, H=hh, W=ww)
-
-                lf_c, hf_c = fft(cond)
-                lf_uc, hf_uc = fft(uncond)
-
-                self.delta_lf = lf_uc - lf_c
-                self.delta_hf = hf_uc - hf_c
-            
-
-        if not return_dict:
-            return (output,)
-        return Transformer2DModelOutput(sample=output)
-
-    @classmethod
-    def from_pretrained_2d(cls, pretrained_model_path, subfolder=None, transformer_additional_kwargs={}):
-        if subfolder is not None:
-            pretrained_model_path = os.path.join(pretrained_model_path, subfolder)
-        print(f"loaded 3D transformer's pretrained weights from {pretrained_model_path} ...")
-
-        config_file = os.path.join(pretrained_model_path, 'config.json')
-        if not os.path.isfile(config_file):
-            raise RuntimeError(f"{config_file} does not exist")
-        with open(config_file, "r") as f:
-            config = json.load(f)
-
-        from diffusers.utils import WEIGHTS_NAME
-        model = cls.from_config(config, **transformer_additional_kwargs)
-        model_file = os.path.join(pretrained_model_path, WEIGHTS_NAME)
-        model_file_safetensors = model_file.replace(".bin", ".safetensors")
-        if os.path.exists(model_file):
-            state_dict = torch.load(model_file, map_location="cpu")
-        elif os.path.exists(model_file_safetensors):
-            from safetensors.torch import load_file, safe_open
-            state_dict = load_file(model_file_safetensors)
-        else:
-            from safetensors.torch import load_file, safe_open
-            model_files_safetensors = glob.glob(os.path.join(pretrained_model_path, "*.safetensors"))
-            state_dict = {}
-            for model_file_safetensors in model_files_safetensors:
-                _state_dict = load_file(model_file_safetensors)
-                for key in _state_dict:
-                    state_dict[key] = _state_dict[key]
-        
-        if model.state_dict()['patch_embed.proj.weight'].size() != state_dict['patch_embed.proj.weight'].size():
-            new_shape   = model.state_dict()['patch_embed.proj.weight'].size()
-            if len(new_shape) == 5:
-                state_dict['patch_embed.proj.weight'] = state_dict['patch_embed.proj.weight'].unsqueeze(2).expand(new_shape).clone()
-                state_dict['patch_embed.proj.weight'][:, :, :-1] = 0
-            else:
-                if model.state_dict()['patch_embed.proj.weight'].size()[1] > state_dict['patch_embed.proj.weight'].size()[1]:
-                    model.state_dict()['patch_embed.proj.weight'][:, :state_dict['patch_embed.proj.weight'].size()[1], :, :] = state_dict['patch_embed.proj.weight']
-                    model.state_dict()['patch_embed.proj.weight'][:, state_dict['patch_embed.proj.weight'].size()[1]:, :, :] = 0
-                    state_dict['patch_embed.proj.weight'] = model.state_dict()['patch_embed.proj.weight']
-                else:
-                    model.state_dict()['patch_embed.proj.weight'][:, :, :, :] = state_dict['patch_embed.proj.weight'][:, :model.state_dict()['patch_embed.proj.weight'].size()[1], :, :]
-                    state_dict['patch_embed.proj.weight'] = model.state_dict()['patch_embed.proj.weight']
-
-        tmp_state_dict = {} 
-        for key in state_dict:
-            if key in model.state_dict().keys() and model.state_dict()[key].size() == state_dict[key].size():
-                tmp_state_dict[key] = state_dict[key]
-            else:
-                print(key, "Size don't match, skip")
-        state_dict = tmp_state_dict
-
-        m, u = model.load_state_dict(state_dict, strict=False)
-        print(f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
-        print(m)
-        
-        params = [p.numel() if "mamba" in n else 0 for n, p in model.named_parameters()]
-        print(f"### Mamba Parameters: {sum(params) / 1e6} M")
-
-        params = [p.numel() if "attn1." in n else 0 for n, p in model.named_parameters()]
-        print(f"### attn1 Parameters: {sum(params) / 1e6} M")
-        
-        return model

cogvideox_fun/utils.py

@@ -1,26 +1,6 @@
-import os
-import gc
 import numpy as np
-import torch
 from PIL import Image
 
-# Copyright (c) OpenMMLab. All rights reserved.
-
-def tensor2pil(image):
-    return Image.fromarray(np.clip(255. * image.cpu().numpy(), 0, 255).astype(np.uint8))
-
-def numpy2pil(image):
-    return Image.fromarray(np.clip(255. * image, 0, 255).astype(np.uint8))
-
-def to_pil(image):
-    if isinstance(image, Image.Image):
-        return image
-    if isinstance(image, torch.Tensor):
-        return tensor2pil(image)
-    if isinstance(image, np.ndarray):
-        return numpy2pil(image)
-    raise ValueError(f"Cannot convert {type(image)} to PIL.Image")
-
 ASPECT_RATIO_512 = {
     '0.25': [256.0, 1024.0], '0.26': [256.0, 992.0], '0.27': [256.0, 960.0], '0.28': [256.0, 928.0],
     '0.32': [288.0, 896.0], '0.33': [288.0, 864.0], '0.35': [288.0, 832.0], '0.4': [320.0, 800.0],
@@ -54,126 +34,10 @@ def get_closest_ratio(height: float, width: float, ratios: dict = ASPECT_RATIO_5
     closest_ratio = min(ratios.keys(), key=lambda ratio: abs(float(ratio) - aspect_ratio))
     return ratios[closest_ratio], float(closest_ratio)
 
-
 def get_width_and_height_from_image_and_base_resolution(image, base_resolution):
     target_pixels = int(base_resolution) * int(base_resolution)
     original_width, original_height = Image.open(image).size
     ratio = (target_pixels / (original_width * original_height)) ** 0.5
     width_slider = round(original_width * ratio)
     height_slider = round(original_height * ratio)
-    return height_slider, width_slider
-
-def get_image_to_video_latent(validation_image_start, validation_image_end, video_length, sample_size):
-    if validation_image_start is not None and validation_image_end is not None:
-        if type(validation_image_start) is str and os.path.isfile(validation_image_start):
-            image_start = clip_image = Image.open(validation_image_start).convert("RGB")
-            image_start = image_start.resize([sample_size[1], sample_size[0]])
-            clip_image = clip_image.resize([sample_size[1], sample_size[0]])
-        else:
-            image_start = clip_image = validation_image_start
-            image_start = [_image_start.resize([sample_size[1], sample_size[0]]) for _image_start in image_start]
-            clip_image = [_clip_image.resize([sample_size[1], sample_size[0]]) for _clip_image in clip_image]
-
-        if type(validation_image_end) is str and os.path.isfile(validation_image_end):
-            image_end = Image.open(validation_image_end).convert("RGB")
-            image_end = image_end.resize([sample_size[1], sample_size[0]])
-        else:
-            image_end = validation_image_end
-            image_end = [_image_end.resize([sample_size[1], sample_size[0]]) for _image_end in image_end]
-
-        if type(image_start) is list:
-            clip_image = clip_image[0]
-            start_video = torch.cat(
-                [torch.from_numpy(np.array(_image_start)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0) for _image_start in image_start], 
-                dim=2
-            )
-            input_video = torch.tile(start_video[:, :, :1], [1, 1, video_length, 1, 1])
-            input_video[:, :, :len(image_start)] = start_video
-            
-            input_video_mask = torch.zeros_like(input_video[:, :1])
-            input_video_mask[:, :, len(image_start):] = 255
-        else:
-            input_video = torch.tile(
-                torch.from_numpy(np.array(image_start)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0), 
-                [1, 1, video_length, 1, 1]
-            )
-            input_video_mask = torch.zeros_like(input_video[:, :1])
-            input_video_mask[:, :, 1:] = 255
-
-        if type(image_end) is list:
-            image_end = [_image_end.resize(image_start[0].size if type(image_start) is list else image_start.size) for _image_end in image_end]
-            end_video = torch.cat(
-                [torch.from_numpy(np.array(_image_end)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0) for _image_end in image_end], 
-                dim=2
-            )
-            input_video[:, :, -len(end_video):] = end_video
-            
-            input_video_mask[:, :, -len(image_end):] = 0
-        else:
-            image_end = image_end.resize(image_start[0].size if type(image_start) is list else image_start.size)
-            input_video[:, :, -1:] = torch.from_numpy(np.array(image_end)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0)
-            input_video_mask[:, :, -1:] = 0
-
-        input_video = input_video / 255
-
-    elif validation_image_start is not None:
-        if type(validation_image_start) is str and os.path.isfile(validation_image_start):
-            image_start = clip_image = Image.open(validation_image_start).convert("RGB")
-            image_start = image_start.resize([sample_size[1], sample_size[0]])
-            clip_image = clip_image.resize([sample_size[1], sample_size[0]])
-        else:
-            image_start = clip_image = validation_image_start
-            image_start = [_image_start.resize([sample_size[1], sample_size[0]]) for _image_start in image_start]
-            clip_image = [_clip_image.resize([sample_size[1], sample_size[0]]) for _clip_image in clip_image]
-        image_end = None
-        
-        if type(image_start) is list:
-            clip_image = clip_image[0]
-            start_video = torch.cat(
-                [torch.from_numpy(np.array(_image_start)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0) for _image_start in image_start], 
-                dim=2
-            )
-            input_video = torch.tile(start_video[:, :, :1], [1, 1, video_length, 1, 1])
-            input_video[:, :, :len(image_start)] = start_video
-            input_video = input_video / 255
-            
-            input_video_mask = torch.zeros_like(input_video[:, :1])
-            input_video_mask[:, :, len(image_start):] = 255
-        else:
-            input_video = torch.tile(
-                torch.from_numpy(np.array(image_start)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0), 
-                [1, 1, video_length, 1, 1]
-            ) / 255
-            input_video_mask = torch.zeros_like(input_video[:, :1])
-            input_video_mask[:, :, 1:, ] = 255
-    else:
-        image_start = None
-        image_end = None
-        input_video = torch.zeros([1, 3, video_length, sample_size[0], sample_size[1]])
-        input_video_mask = torch.ones([1, 1, video_length, sample_size[0], sample_size[1]]) * 255
-        clip_image = None
-
-    del image_start
-    del image_end
-    gc.collect()
-
-    return  input_video, input_video_mask, clip_image
-
-def get_video_to_video_latent(input_video_path, video_length, sample_size, validation_video_mask=None):
-    input_video = input_video_path
-
-    input_video = torch.from_numpy(np.array(input_video))[:video_length]
-    input_video = input_video.permute([3, 0, 1, 2]).unsqueeze(0) / 255
-
-    if validation_video_mask is not None:
-        validation_video_mask = Image.open(validation_video_mask).convert('L').resize((sample_size[1], sample_size[0]))
-        input_video_mask = np.where(np.array(validation_video_mask) < 240, 0, 255)
-        
-        input_video_mask = torch.from_numpy(np.array(input_video_mask)).unsqueeze(0).unsqueeze(-1).permute([3, 0, 1, 2]).unsqueeze(0)
-        input_video_mask = torch.tile(input_video_mask, [1, 1, input_video.size()[2], 1, 1])
-        input_video_mask = input_video_mask.to(input_video.device, input_video.dtype)
-    else:
-        input_video_mask = torch.zeros_like(input_video[:, :1])
-        input_video_mask[:, :, :] = 255
-
-    return  input_video, input_video_mask, None
+    return height_slider, width_slider

convert_weight_sat2hf.py

@@ -1,303 +0,0 @@
-"""
-
-The script demonstrates how to convert the weights of the CogVideoX model from SAT to Hugging Face format.
-This script supports the conversion of the following models:
-- CogVideoX-2B
-- CogVideoX-5B, CogVideoX-5B-I2V
-- CogVideoX1.1-5B, CogVideoX1.1-5B-I2V
-
-Original Script:
-https://github.com/huggingface/diffusers/blob/main/scripts/convert_cogvideox_to_diffusers.py
-
-"""
-import argparse
-from typing import Any, Dict
-
-import torch
-from transformers import T5EncoderModel, T5Tokenizer
-
-from diffusers import (
-    AutoencoderKLCogVideoX,
-    CogVideoXDDIMScheduler,
-    CogVideoXImageToVideoPipeline,
-    CogVideoXPipeline,
-    #CogVideoXTransformer3DModel,
-)
-from custom_cogvideox_transformer_3d import CogVideoXTransformer3DModel
-
-
-def reassign_query_key_value_inplace(key: str, state_dict: Dict[str, Any]):
-    to_q_key = key.replace("query_key_value", "to_q")
-    to_k_key = key.replace("query_key_value", "to_k")
-    to_v_key = key.replace("query_key_value", "to_v")
-    to_q, to_k, to_v = torch.chunk(state_dict[key], chunks=3, dim=0)
-    state_dict[to_q_key] = to_q
-    state_dict[to_k_key] = to_k
-    state_dict[to_v_key] = to_v
-    state_dict.pop(key)
-
-
-def reassign_query_key_layernorm_inplace(key: str, state_dict: Dict[str, Any]):
-    layer_id, weight_or_bias = key.split(".")[-2:]
-
-    if "query" in key:
-        new_key = f"transformer_blocks.{layer_id}.attn1.norm_q.{weight_or_bias}"
-    elif "key" in key:
-        new_key = f"transformer_blocks.{layer_id}.attn1.norm_k.{weight_or_bias}"
-
-    state_dict[new_key] = state_dict.pop(key)
-
-
-def reassign_adaln_norm_inplace(key: str, state_dict: Dict[str, Any]):
-    layer_id, _, weight_or_bias = key.split(".")[-3:]
-
-    weights_or_biases = state_dict[key].chunk(12, dim=0)
-    norm1_weights_or_biases = torch.cat(weights_or_biases[0:3] + weights_or_biases[6:9])
-    norm2_weights_or_biases = torch.cat(weights_or_biases[3:6] + weights_or_biases[9:12])
-
-    norm1_key = f"transformer_blocks.{layer_id}.norm1.linear.{weight_or_bias}"
-    state_dict[norm1_key] = norm1_weights_or_biases
-
-    norm2_key = f"transformer_blocks.{layer_id}.norm2.linear.{weight_or_bias}"
-    state_dict[norm2_key] = norm2_weights_or_biases
-
-    state_dict.pop(key)
-
-
-def remove_keys_inplace(key: str, state_dict: Dict[str, Any]):
-    state_dict.pop(key)
-
-
-def replace_up_keys_inplace(key: str, state_dict: Dict[str, Any]):
-    key_split = key.split(".")
-    layer_index = int(key_split[2])
-    replace_layer_index = 4 - 1 - layer_index
-
-    key_split[1] = "up_blocks"
-    key_split[2] = str(replace_layer_index)
-    new_key = ".".join(key_split)
-
-    state_dict[new_key] = state_dict.pop(key)
-
-
-TRANSFORMER_KEYS_RENAME_DICT = {
-    "transformer.final_layernorm": "norm_final",
-    "transformer": "transformer_blocks",
-    "attention": "attn1",
-    "mlp": "ff.net",
-    "dense_h_to_4h": "0.proj",
-    "dense_4h_to_h": "2",
-    ".layers": "",
-    "dense": "to_out.0",
-    "input_layernorm": "norm1.norm",
-    "post_attn1_layernorm": "norm2.norm",
-    "time_embed.0": "time_embedding.linear_1",
-    "time_embed.2": "time_embedding.linear_2",
-    "mixins.patch_embed": "patch_embed",
-    "mixins.final_layer.norm_final": "norm_out.norm",
-    "mixins.final_layer.linear": "proj_out",
-    "mixins.final_layer.adaLN_modulation.1": "norm_out.linear",
-    "mixins.pos_embed.pos_embedding": "patch_embed.pos_embedding",  # Specific to CogVideoX-5b-I2V
-}
-
-TRANSFORMER_SPECIAL_KEYS_REMAP = {
-    "query_key_value": reassign_query_key_value_inplace,
-    "query_layernorm_list": reassign_query_key_layernorm_inplace,
-    "key_layernorm_list": reassign_query_key_layernorm_inplace,
-    "adaln_layer.adaLN_modulations": reassign_adaln_norm_inplace,
-    "embed_tokens": remove_keys_inplace,
-    "freqs_sin": remove_keys_inplace,
-    "freqs_cos": remove_keys_inplace,
-    "position_embedding": remove_keys_inplace,
-}
-
-VAE_KEYS_RENAME_DICT = {
-    "block.": "resnets.",
-    "down.": "down_blocks.",
-    "downsample": "downsamplers.0",
-    "upsample": "upsamplers.0",
-    "nin_shortcut": "conv_shortcut",
-    "encoder.mid.block_1": "encoder.mid_block.resnets.0",
-    "encoder.mid.block_2": "encoder.mid_block.resnets.1",
-    "decoder.mid.block_1": "decoder.mid_block.resnets.0",
-    "decoder.mid.block_2": "decoder.mid_block.resnets.1",
-}
-
-VAE_SPECIAL_KEYS_REMAP = {
-    "loss": remove_keys_inplace,
-    "up.": replace_up_keys_inplace,
-}
-
-TOKENIZER_MAX_LENGTH = 226
-
-
-def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
-    state_dict = saved_dict
-    if "model" in saved_dict.keys():
-        state_dict = state_dict["model"]
-    if "module" in saved_dict.keys():
-        state_dict = state_dict["module"]
-    if "state_dict" in saved_dict.keys():
-        state_dict = state_dict["state_dict"]
-    return state_dict
-
-
-def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
-    state_dict[new_key] = state_dict.pop(old_key)
-
-
-def convert_transformer(
-        ckpt_path: str,
-        num_layers: int,
-        num_attention_heads: int,
-        use_rotary_positional_embeddings: bool,
-        i2v: bool,
-        dtype: torch.dtype,
-):
-    PREFIX_KEY = "model.diffusion_model."
-
-    original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
-    transformer = CogVideoXTransformer3DModel(
-        in_channels=32 if i2v else 16,
-        num_layers=num_layers,
-        num_attention_heads=num_attention_heads,
-        use_rotary_positional_embeddings=use_rotary_positional_embeddings,
-        use_learned_positional_embeddings=i2v,
-    ).to(dtype=dtype)
-
-    for key in list(original_state_dict.keys()):
-        new_key = key[len(PREFIX_KEY):]
-        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
-            new_key = new_key.replace(replace_key, rename_key)
-        update_state_dict_inplace(original_state_dict, key, new_key)
-
-    for key in list(original_state_dict.keys()):
-        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
-            if special_key not in key:
-                continue
-            handler_fn_inplace(key, original_state_dict)
-    transformer.load_state_dict(original_state_dict, strict=True)
-    return transformer
-
-
-def convert_vae(ckpt_path: str, scaling_factor: float, dtype: torch.dtype):
-    original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
-    vae = AutoencoderKLCogVideoX(scaling_factor=scaling_factor).to(dtype=dtype)
-
-    for key in list(original_state_dict.keys()):
-        new_key = key[:]
-        for replace_key, rename_key in VAE_KEYS_RENAME_DICT.items():
-            new_key = new_key.replace(replace_key, rename_key)
-        update_state_dict_inplace(original_state_dict, key, new_key)
-
-    for key in list(original_state_dict.keys()):
-        for special_key, handler_fn_inplace in VAE_SPECIAL_KEYS_REMAP.items():
-            if special_key not in key:
-                continue
-            handler_fn_inplace(key, original_state_dict)
-
-    vae.load_state_dict(original_state_dict, strict=True)
-    return vae
-
-
-def get_args():
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--transformer_ckpt_path", type=str, default=None, help="Path to original transformer checkpoint"
-    )
-    parser.add_argument("--vae_ckpt_path", type=str, default=None, help="Path to original vae checkpoint")
-    parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
-    parser.add_argument("--fp16", action="store_true", default=False, help="Whether to save the model weights in fp16")
-    parser.add_argument("--bf16", action="store_true", default=False, help="Whether to save the model weights in bf16")
-    parser.add_argument(
-        "--push_to_hub", action="store_true", default=False, help="Whether to push to HF Hub after saving"
-    )
-    parser.add_argument(
-        "--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
-    )
-    # For CogVideoX-2B, num_layers is 30. For 5B, it is 42
-    parser.add_argument("--num_layers", type=int, default=30, help="Number of transformer blocks")
-    # For CogVideoX-2B, num_attention_heads is 30. For 5B, it is 48
-    parser.add_argument("--num_attention_heads", type=int, default=30, help="Number of attention heads")
-    # For CogVideoX-2B, use_rotary_positional_embeddings is False. For 5B, it is True
-    parser.add_argument(
-        "--use_rotary_positional_embeddings", action="store_true", default=False, help="Whether to use RoPE or not"
-    )
-    # For CogVideoX-2B, scaling_factor is 1.15258426. For 5B, it is 0.7
-    parser.add_argument("--scaling_factor", type=float, default=1.15258426, help="Scaling factor in the VAE")
-    # For CogVideoX-2B, snr_shift_scale is 3.0. For 5B, it is 1.0
-    parser.add_argument("--snr_shift_scale", type=float, default=3.0, help="Scaling factor in the VAE")
-    parser.add_argument("--i2v", action="store_true", default=False, help="Whether to save the model weights in fp16")
-    return parser.parse_args()
-
-
-if __name__ == "__main__":
-    args = get_args()
-
-    transformer = None
-    vae = None
-
-    if args.fp16 and args.bf16:
-        raise ValueError("You cannot pass both --fp16 and --bf16 at the same time.")
-
-    dtype = torch.float16 if args.fp16 else torch.bfloat16 if args.bf16 else torch.float32
-
-    if args.transformer_ckpt_path is not None:
-        transformer = convert_transformer(
-            args.transformer_ckpt_path,
-            args.num_layers,
-            args.num_attention_heads,
-            args.use_rotary_positional_embeddings,
-            args.i2v,
-            dtype,
-        )
-    if args.vae_ckpt_path is not None:
-        vae = convert_vae(args.vae_ckpt_path, args.scaling_factor, dtype)
-
-    #text_encoder_id = "/share/official_pretrains/hf_home/t5-v1_1-xxl"
-    #tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
-    #text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
-
-    # Apparently, the conversion does not work anymore without this :shrug:
-    #for param in text_encoder.parameters():
-    #    param.data = param.data.contiguous()
-
-    scheduler = CogVideoXDDIMScheduler.from_config(
-        {
-            "snr_shift_scale": args.snr_shift_scale,
-            "beta_end": 0.012,
-            "beta_schedule": "scaled_linear",
-            "beta_start": 0.00085,
-            "clip_sample": False,
-            "num_train_timesteps": 1000,
-            "prediction_type": "v_prediction",
-            "rescale_betas_zero_snr": True,
-            "set_alpha_to_one": True,
-            "timestep_spacing": "trailing",
-        }
-    )
-    if args.i2v:
-        pipeline_cls = CogVideoXImageToVideoPipeline
-    else:
-        pipeline_cls = CogVideoXPipeline
-
-    pipe = pipeline_cls(
-        tokenizer=None,
-        text_encoder=None,
-        vae=vae,
-        transformer=transformer,
-        scheduler=scheduler,
-    )
-
-    if args.fp16:
-        pipe = pipe.to(dtype=torch.float16)
-    if args.bf16:
-        pipe = pipe.to(dtype=torch.bfloat16)
-
-    # We don't use variant here because the model must be run in fp16 (2B) or bf16 (5B). It would be weird
-    # for users to specify variant when the default is not fp32 and they want to run with the correct default (which
-    # is either fp16/bf16 here).
-
-    # This is necessary This is necessary for users with insufficient memory,
-    # such as those using Colab and notebooks, as it can save some memory used for model loading.
-    pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB", push_to_hub=args.push_to_hub)

custom_cogvideox_transformer_3d.py

@@ -35,20 +35,56 @@ from diffusers.loaders import PeftAdapterMixin
 from diffusers.models.embeddings import apply_rotary_emb
 from .embeddings import CogVideoXPatchEmbed
 
+from .enhance_a_video.enhance import get_feta_scores
+from .enhance_a_video.globals import is_enhance_enabled, set_num_frames
+
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 try:
     from sageattention import sageattn
-    
     SAGEATTN_IS_AVAILABLE = True
 except:
     SAGEATTN_IS_AVAILABLE = False
 
-@torch.compiler.disable()
-def sageattn_func(query, key, value, attn_mask=None, dropout_p=0.0,is_causal=False):
-    return sageattn(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p,is_causal=is_causal)
+from comfy.ldm.modules.attention import optimized_attention
 
+
+def set_attention_func(attention_mode, heads):
+    if attention_mode == "sdpa" or attention_mode == "fused_sdpa":
+        def func(q, k, v, is_causal=False, attn_mask=None):
+            return F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask, dropout_p=0.0, is_causal=is_causal)
+        return func
+    elif attention_mode == "comfy":
+        def func(q, k, v, is_causal=False, attn_mask=None):
+            return optimized_attention(q, k, v, mask=attn_mask, heads=heads, skip_reshape=True)
+        return func
+
+    elif attention_mode == "sageattn" or attention_mode == "fused_sageattn":
+        @torch.compiler.disable()
+        def func(q, k, v, is_causal=False, attn_mask=None):
+            return sageattn(q.to(v), k.to(v), v, is_causal=is_causal, attn_mask=attn_mask)
+        return func
+    elif attention_mode == "sageattn_qk_int8_pv_fp16_cuda":
+        from sageattention import sageattn_qk_int8_pv_fp16_cuda
+        @torch.compiler.disable()
+        def func(q, k, v, is_causal=False, attn_mask=None):
+            return sageattn_qk_int8_pv_fp16_cuda(q.to(v), k.to(v), v, is_causal=is_causal, attn_mask=attn_mask, pv_accum_dtype="fp32")
+        return func
+    elif attention_mode == "sageattn_qk_int8_pv_fp16_triton":
+        from sageattention import sageattn_qk_int8_pv_fp16_triton
+        @torch.compiler.disable()
+        def func(q, k, v, is_causal=False, attn_mask=None):
+            return sageattn_qk_int8_pv_fp16_triton(q.to(v), k.to(v), v, is_causal=is_causal, attn_mask=attn_mask)
+        return func
+    elif attention_mode == "sageattn_qk_int8_pv_fp8_cuda":
+        from sageattention import sageattn_qk_int8_pv_fp8_cuda
+        @torch.compiler.disable()
+        def func(q, k, v, is_causal=False, attn_mask=None):
+            return sageattn_qk_int8_pv_fp8_cuda(q.to(v), k.to(v), v, is_causal=is_causal, attn_mask=attn_mask, pv_accum_dtype="fp32+fp32")
+        return func
+
+#for fastercache
 def fft(tensor):
     tensor_fft = torch.fft.fft2(tensor)
     tensor_fft_shifted = torch.fft.fftshift(tensor_fft)
@@ -66,17 +102,25 @@ def fft(tensor):
 
     return low_freq_fft, high_freq_fft
 
+#for teacache
+def poly1d(coefficients, x):
+    result = torch.zeros_like(x)
+    for i, coeff in enumerate(coefficients):
+        result += coeff * (x ** (len(coefficients) - 1 - i))
+    return result.abs()
+
+#region Attention
 class CogVideoXAttnProcessor2_0:
     r"""
     Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
     query and key vectors, but does not include spatial normalization.
     """
 
-    def __init__(self):
+    def __init__(self, attn_func, attention_mode: Optional[str] = None):
         if not hasattr(F, "scaled_dot_product_attention"):
             raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
-        
-    #@torch.compiler.disable()
+        self.attention_mode = attention_mode
+        self.attn_func = attn_func
     def __call__(
         self,
         attn: Attention,
@@ -84,7 +128,6 @@ class CogVideoXAttnProcessor2_0:
         encoder_hidden_states: torch.Tensor,
         attention_mask: Optional[torch.Tensor] = None,
         image_rotary_emb: Optional[torch.Tensor] = None,
-        attention_mode: Optional[str] = None,
     ) -> torch.Tensor:
         text_seq_length = encoder_hidden_states.size(1)
 
@@ -98,7 +141,10 @@ class CogVideoXAttnProcessor2_0:
             attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
             attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
 
-        if attention_mode != "fused_sdpa" or attention_mode != "fused_sageattn":
+        if attn.to_q.weight.dtype == torch.float16 or attn.to_q.weight.dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(attn.to_q.weight.dtype)
+
+        if not "fused" in self.attention_mode:
             query = attn.to_q(hidden_states)
             key = attn.to_k(hidden_states)
             value = attn.to_v(hidden_states)
@@ -124,17 +170,15 @@ class CogVideoXAttnProcessor2_0:
             query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
             if not attn.is_cross_attention:
                 key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
-                 
-        if attention_mode == "sageattn" or attention_mode == "fused_sageattn":
-            hidden_states = sageattn_func(query, key, value, attn_mask=attention_mask, dropout_p=0.0,is_causal=False)
-        else:
-            hidden_states = F.scaled_dot_product_attention(
-                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
-                )
-        #if torch.isinf(hidden_states).any():
-        #    raise ValueError(f"hidden_states after dot product has inf")
 
-        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        #feta
+        if is_enhance_enabled():
+            feta_scores = get_feta_scores(attn, query, key, head_dim, text_seq_length)
+                
+        hidden_states = self.attn_func(query, key, value, attn_mask=attention_mask, is_causal=False)
+       
+        if self.attention_mode != "comfy":
+            hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
 
         # linear proj
         hidden_states = attn.to_out[0](hidden_states)
@@ -144,6 +188,10 @@ class CogVideoXAttnProcessor2_0:
         encoder_hidden_states, hidden_states = hidden_states.split(
             [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
         )
+
+        if is_enhance_enabled():
+            hidden_states *= feta_scores
+
         return hidden_states, encoder_hidden_states
 
 #region Blocks
@@ -200,13 +248,15 @@ class CogVideoXBlock(nn.Module):
         ff_inner_dim: Optional[int] = None,
         ff_bias: bool = True,
         attention_out_bias: bool = True,
+        attention_mode: Optional[str] = "sdpa",
     ):
         super().__init__()
 
         # 1. Self Attention
         self.norm1 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)
-       
 
+        attn_func = set_attention_func(attention_mode, num_attention_heads)
+       
         self.attn1 = Attention(
             query_dim=dim,
             dim_head=attention_head_dim,
@@ -215,7 +265,7 @@ class CogVideoXBlock(nn.Module):
             eps=1e-6,
             bias=attention_bias,
             out_bias=attention_out_bias,
-            processor=CogVideoXAttnProcessor2_0(),
+            processor=CogVideoXAttnProcessor2_0(attn_func, attention_mode=attention_mode),
         )
 
         # 2. Feed Forward
@@ -244,7 +294,6 @@ class CogVideoXBlock(nn.Module):
         fastercache_counter=0,
         fastercache_start_step=15,
         fastercache_device="cuda:0",
-        attention_mode="sdpa",
     ) -> torch.Tensor:
         #print("hidden_states in block: ", hidden_states.shape) #1.5: torch.Size([2, 3200, 3072]) 10.: torch.Size([2, 6400, 3072])
         text_seq_length = encoder_hidden_states.size(1)
@@ -283,7 +332,6 @@ class CogVideoXBlock(nn.Module):
                     hidden_states=norm_hidden_states,
                     encoder_hidden_states=norm_encoder_hidden_states,
                     image_rotary_emb=image_rotary_emb,
-                    attention_mode=attention_mode,
                 )
                 if fastercache_counter == fastercache_start_step:
                     self.cached_hidden_states = [attn_hidden_states.to(fastercache_device), attn_hidden_states.to(fastercache_device)]
@@ -295,8 +343,7 @@ class CogVideoXBlock(nn.Module):
             attn_hidden_states, attn_encoder_hidden_states = self.attn1(
                 hidden_states=norm_hidden_states,
                 encoder_hidden_states=norm_encoder_hidden_states,
-                image_rotary_emb=image_rotary_emb,
-                attention_mode=attention_mode,
+                image_rotary_emb=image_rotary_emb
             )
         
         hidden_states = hidden_states + gate_msa * attn_hidden_states
@@ -405,6 +452,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         use_rotary_positional_embeddings: bool = False,
         use_learned_positional_embeddings: bool = False,
         patch_bias: bool = True,
+        attention_mode: Optional[str] = "sdpa",
     ):
         super().__init__()
         inner_dim = num_attention_heads * attention_head_dim
@@ -458,6 +506,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
                     dropout=dropout,
                     activation_fn=activation_fn,
                     attention_bias=attention_bias,
+                    attention_mode=attention_mode,
                     norm_elementwise_affine=norm_elementwise_affine,
                     norm_eps=norm_eps,
                 )
@@ -485,7 +534,12 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
 
         self.gradient_checkpointing = False
 
+        self.attention_mode = attention_mode
+
+        #tora
         self.fuser_list = None
+
+        #fastercache
         self.use_fastercache = False
         self.fastercache_counter = 0
         self.fastercache_start_step = 15
@@ -493,73 +547,21 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         self.fastercache_hf_step = 30
         self.fastercache_device = "cuda"
         self.fastercache_num_blocks_to_cache = len(self.transformer_blocks)
-        self.attention_mode = "sdpa"
+        
+        #teacache
+        self.use_teacache = False
+        self.teacache_rel_l1_thresh = 0.0
+        if not self.config.use_rotary_positional_embeddings:
+            #CogVideoX-2B
+            self.teacache_coefficients = [-3.10658903e+01, 2.54732368e+01, -5.92380459e+00, 1.75769064e+00, -3.61568434e-03]
+        else:
+            #CogVideoX-5B
+            self.teacache_coefficients = [-1.53880483e+03, 8.43202495e+02, -1.34363087e+02, 7.97131516e+00, -5.23162339e-02]
         
 
     def _set_gradient_checkpointing(self, module, value=False):
         self.gradient_checkpointing = value
-
-    @property
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
-    def attn_processors(self) -> Dict[str, AttentionProcessor]:
-        r"""
-        Returns:
-            `dict` of attention processors: A dictionary containing all attention processors used in the model with
-            indexed by its weight name.
-        """
-        # set recursively
-        processors = {}
-
-        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
-            if hasattr(module, "get_processor"):
-                processors[f"{name}.processor"] = module.get_processor()
-
-            for sub_name, child in module.named_children():
-                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
-
-            return processors
-
-        for name, module in self.named_children():
-            fn_recursive_add_processors(name, module, processors)
-
-        return processors
-
-    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
-    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
-        r"""
-        Sets the attention processor to use to compute attention.
-
-        Parameters:
-            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
-                The instantiated processor class or a dictionary of processor classes that will be set as the processor
-                for **all** `Attention` layers.
-
-                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
-                processor. This is strongly recommended when setting trainable attention processors.
-
-        """
-        count = len(self.attn_processors.keys())
-
-        if isinstance(processor, dict) and len(processor) != count:
-            raise ValueError(
-                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
-                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
-            )
-
-        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
-            if hasattr(module, "set_processor"):
-                if not isinstance(processor, dict):
-                    module.set_processor(processor)
-                else:
-                    module.set_processor(processor.pop(f"{name}.processor"))
-
-            for sub_name, child in module.named_children():
-                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
-
-        for name, module in self.named_children():
-            fn_recursive_attn_processor(name, module, processor)
-
-
+    #region forward
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -574,6 +576,8 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         return_dict: bool = True,
     ):
         batch_size, num_frames, channels, height, width = hidden_states.shape
+
+        set_num_frames(num_frames) #enhance a video global
    
         # 1. Time embedding
         timesteps = timestep
@@ -611,29 +615,28 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         if self.fastercache_counter >= self.fastercache_start_step + 3 and self.fastercache_counter % 5 !=0:
             # 3. Transformer blocks
             for i, block in enumerate(self.transformer_blocks):
-                    hidden_states, encoder_hidden_states = block(
-                        hidden_states=hidden_states[:1],
-                        encoder_hidden_states=encoder_hidden_states[:1],
-                        temb=emb[:1],
-                        image_rotary_emb=image_rotary_emb,
-                        video_flow_feature=video_flow_features[i][:1] if video_flow_features is not None else None,
-                        fuser = self.fuser_list[i] if self.fuser_list is not None else None,
-                        block_use_fastercache = i <= self.fastercache_num_blocks_to_cache,
-                        fastercache_counter = self.fastercache_counter,
-                        fastercache_start_step = self.fastercache_start_step,
-                        fastercache_device = self.fastercache_device,
-                        attention_mode = self.attention_mode
-                    )
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states=hidden_states[:1],
+                    encoder_hidden_states=encoder_hidden_states[:1],
+                    temb=emb[:1],
+                    image_rotary_emb=image_rotary_emb,
+                    video_flow_feature=video_flow_features[i][:1] if video_flow_features is not None else None,
+                    fuser = self.fuser_list[i] if self.fuser_list is not None else None,
+                    block_use_fastercache = i <= self.fastercache_num_blocks_to_cache,
+                    fastercache_counter = self.fastercache_counter,
+                    fastercache_start_step = self.fastercache_start_step,
+                    fastercache_device = self.fastercache_device
+                )
 
-                    if (controlnet_states is not None) and (i < len(controlnet_states)):
-                        controlnet_states_block = controlnet_states[i]
-                        controlnet_block_weight = 1.0
-                        if isinstance(controlnet_weights, (list, np.ndarray)) or torch.is_tensor(controlnet_weights):
-                            controlnet_block_weight = controlnet_weights[i]
-                        elif isinstance(controlnet_weights, (float, int)):
-                            controlnet_block_weight = controlnet_weights
-                        
-                        hidden_states = hidden_states + controlnet_states_block * controlnet_block_weight
+                if (controlnet_states is not None) and (i < len(controlnet_states)):
+                    controlnet_states_block = controlnet_states[i]
+                    controlnet_block_weight = 1.0
+                    if isinstance(controlnet_weights, (list, np.ndarray)) or torch.is_tensor(controlnet_weights):
+                        controlnet_block_weight = controlnet_weights[i]
+                    elif isinstance(controlnet_weights, (float, int)):
+                        controlnet_block_weight = controlnet_weights
+                    
+                    hidden_states = hidden_states + controlnet_states_block * controlnet_block_weight
                     
             if not self.config.use_rotary_positional_embeddings:
                 # CogVideoX-2B
@@ -681,33 +684,56 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
             recovered_uncond = rearrange(recovered_uncond.to(output.dtype), "(B T) C H W -> B T C H W", B=bb, C=cc, T=tt, H=hh, W=ww)
             output = torch.cat([output, recovered_uncond])
         else:
-            for i, block in enumerate(self.transformer_blocks):
-                hidden_states, encoder_hidden_states = block(
-                    hidden_states=hidden_states,
-                    encoder_hidden_states=encoder_hidden_states,
-                    temb=emb,
-                    image_rotary_emb=image_rotary_emb,
-                    video_flow_feature=video_flow_features[i] if video_flow_features is not None else None,
-                    fuser = self.fuser_list[i] if self.fuser_list is not None else None,
-                    block_use_fastercache = i <= self.fastercache_num_blocks_to_cache,
-                    fastercache_counter = self.fastercache_counter,
-                    fastercache_start_step = self.fastercache_start_step,
-                    fastercache_device = self.fastercache_device,
-                    attention_mode = self.attention_mode
-                )
-                #has_nan = torch.isnan(hidden_states).any()
-                #if has_nan:
-                #    raise ValueError(f"block output hidden_states has nan: {has_nan}")
+            if self.use_teacache:
+                if not hasattr(self, 'accumulated_rel_l1_distance'):
+                    should_calc = True
+                    self.accumulated_rel_l1_distance = 0
+                else:                    
+                    self.accumulated_rel_l1_distance += poly1d(self.teacache_coefficients, ((emb-self.previous_modulated_input).abs().mean() / self.previous_modulated_input.abs().mean()))
+                    if self.accumulated_rel_l1_distance < self.teacache_rel_l1_thresh:
+                        should_calc = False
+                        self.teacache_counter += 1
+                    else:
+                        should_calc = True
+                        self.accumulated_rel_l1_distance = 0
+                #print("self.accumulated_rel_l1_distance ", self.accumulated_rel_l1_distance)
+                self.previous_modulated_input = emb
+                if not should_calc:
+                    hidden_states += self.previous_residual
+                    encoder_hidden_states += self.previous_residual_encoder
+                    
+            if not self.use_teacache or (self.use_teacache and should_calc):
+                if self.use_teacache:
+                    ori_hidden_states = hidden_states.clone()
+                    ori_encoder_hidden_states = encoder_hidden_states.clone()
+                for i, block in enumerate(self.transformer_blocks):
+                    hidden_states, encoder_hidden_states = block(
+                        hidden_states=hidden_states,
+                        encoder_hidden_states=encoder_hidden_states,
+                        temb=emb,
+                        image_rotary_emb=image_rotary_emb,
+                        video_flow_feature=video_flow_features[i] if video_flow_features is not None else None,
+                        fuser = self.fuser_list[i] if self.fuser_list is not None else None,
+                        block_use_fastercache = i <= self.fastercache_num_blocks_to_cache,
+                        fastercache_counter = self.fastercache_counter,
+                        fastercache_start_step = self.fastercache_start_step,
+                        fastercache_device = self.fastercache_device
+                    )
 
-            if (controlnet_states is not None) and (i < len(controlnet_states)):
-                controlnet_states_block = controlnet_states[i]
-                controlnet_block_weight = 1.0
-                if isinstance(controlnet_weights, (list, np.ndarray)) or torch.is_tensor(controlnet_weights):
-                    controlnet_block_weight = controlnet_weights[i]
-                elif isinstance(controlnet_weights, (float, int)):
-                    controlnet_block_weight = controlnet_weights
-                
-                hidden_states = hidden_states + controlnet_states_block * controlnet_block_weight
+                    #controlnet
+                    if (controlnet_states is not None) and (i < len(controlnet_states)):
+                        controlnet_states_block = controlnet_states[i]
+                        controlnet_block_weight = 1.0
+                        if isinstance(controlnet_weights, (list, np.ndarray)) or torch.is_tensor(controlnet_weights):
+                            controlnet_block_weight = controlnet_weights[i]
+                            print(controlnet_block_weight)
+                        elif isinstance(controlnet_weights, (float, int)):
+                            controlnet_block_weight = controlnet_weights                    
+                        hidden_states = hidden_states + controlnet_states_block * controlnet_block_weight
+
+                if self.use_teacache:
+                    self.previous_residual = hidden_states - ori_hidden_states
+                    self.previous_residual_encoder = encoder_hidden_states - ori_encoder_hidden_states
                     
             if not self.config.use_rotary_positional_embeddings:
                 # CogVideoX-2B

embeddings.py

@@ -67,8 +67,9 @@ class CogVideoXPatchEmbed(nn.Module):
             post_time_compression_frames,
             self.spatial_interpolation_scale,
             self.temporal_interpolation_scale,
+            output_type="pt",
         )
-        pos_embedding = torch.from_numpy(pos_embedding).flatten(0, 1)
+        pos_embedding = pos_embedding.flatten(0, 1)
         joint_pos_embedding = torch.zeros(
             1, self.max_text_seq_length + num_patches, self.embed_dim, requires_grad=False
         )
@@ -173,6 +174,8 @@ def get_3d_rotary_pos_embed(
         grid_t = np.arange(temporal_size, dtype=np.float32)
         grid_t = np.linspace(0, temporal_size, temporal_size, endpoint=False, dtype=np.float32)
     elif grid_type == "slice":
+        if max_size is None:
+            raise ValueError("`max_size` must be provided when `grid_type` is 'slice'")
         max_h, max_w = max_size
         grid_size_h, grid_size_w = grid_size
         grid_h = np.arange(max_h, dtype=np.float32)

enhance_a_video/enhance.py

@@ -0,0 +1,82 @@
+import torch
+from einops import rearrange
+from diffusers.models.attention import Attention
+from .globals import get_enhance_weight, get_num_frames
+
+# def get_feta_scores(query, key):
+#     img_q, img_k = query, key
+   
+#     num_frames = get_num_frames()
+    
+#     B, S, N, C = img_q.shape
+
+#     # Calculate spatial dimension
+#     spatial_dim = S // num_frames
+    
+#     # Add time dimension between spatial and head dims
+#     query_image = img_q.reshape(B, spatial_dim, num_frames, N, C)
+#     key_image = img_k.reshape(B, spatial_dim, num_frames, N, C)
+    
+#     # Expand time dimension
+#     query_image = query_image.expand(-1, -1, num_frames, -1, -1)  # [B, S, T, N, C]
+#     key_image = key_image.expand(-1, -1, num_frames, -1, -1)      # [B, S, T, N, C]
+    
+#     # Reshape to match feta_score input format: [(B S) N T C]
+#     query_image = rearrange(query_image, "b s t n c -> (b s) n t c")  #torch.Size([3200, 24, 5, 128])
+#     key_image = rearrange(key_image, "b s t n c -> (b s) n t c")
+    
+#     return feta_score(query_image, key_image, C, num_frames)
+ 
+def get_feta_scores(
+        attn: Attention,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        head_dim: int,
+        text_seq_length: int,
+    ) -> torch.Tensor:
+        num_frames = get_num_frames()
+        spatial_dim = int((query.shape[2] - text_seq_length) / num_frames)
+
+        query_image = rearrange(
+            query[:, :, text_seq_length:],
+            "B N (T S) C -> (B S) N T C",
+            N=attn.heads,
+            T=num_frames,
+            S=spatial_dim,
+            C=head_dim,
+        )
+        key_image = rearrange(
+            key[:, :, text_seq_length:],
+            "B N (T S) C -> (B S) N T C",
+            N=attn.heads,
+            T=num_frames,
+            S=spatial_dim,
+            C=head_dim,
+        )
+        return feta_score(query_image, key_image, head_dim, num_frames)
+
+def feta_score(query_image, key_image, head_dim, num_frames):
+    scale = head_dim**-0.5
+    query_image = query_image * scale
+    attn_temp = query_image @ key_image.transpose(-2, -1)  # translate attn to float32
+    attn_temp = attn_temp.to(torch.float32)
+    attn_temp = attn_temp.softmax(dim=-1)
+
+    # Reshape to [batch_size * num_tokens, num_frames, num_frames]
+    attn_temp = attn_temp.reshape(-1, num_frames, num_frames)
+
+    # Create a mask for diagonal elements
+    diag_mask = torch.eye(num_frames, device=attn_temp.device).bool()
+    diag_mask = diag_mask.unsqueeze(0).expand(attn_temp.shape[0], -1, -1)
+
+    # Zero out diagonal elements
+    attn_wo_diag = attn_temp.masked_fill(diag_mask, 0)
+
+    # Calculate mean for each token's attention matrix
+    # Number of off-diagonal elements per matrix is n*n - n
+    num_off_diag = num_frames * num_frames - num_frames
+    mean_scores = attn_wo_diag.sum(dim=(1, 2)) / num_off_diag
+
+    enhance_scores = mean_scores.mean() * (num_frames + get_enhance_weight())
+    enhance_scores = enhance_scores.clamp(min=1)
+    return enhance_scores

enhance_a_video/globals.py

@@ -0,0 +1,31 @@
+NUM_FRAMES = None
+FETA_WEIGHT = None
+ENABLE_FETA = False
+
+def set_num_frames(num_frames: int):
+    global NUM_FRAMES
+    NUM_FRAMES = num_frames
+
+
+def get_num_frames() -> int:
+    return NUM_FRAMES
+
+
+def enable_enhance():
+    global ENABLE_FETA
+    ENABLE_FETA = True
+
+def disable_enhance():
+    global ENABLE_FETA
+    ENABLE_FETA = False
+
+def is_enhance_enabled() -> bool:
+    return ENABLE_FETA
+
+def set_enhance_weight(feta_weight: float):
+    global FETA_WEIGHT
+    FETA_WEIGHT = feta_weight
+
+
+def get_enhance_weight() -> float:
+    return FETA_WEIGHT

example_workflows/cogvideox_1_0_5b_I2V_02.json

@@ -1,42 +1,7 @@
 {
-  "last_node_id": 58,
-  "last_link_id": 129,
+  "last_node_id": 63,
+  "last_link_id": 149,
   "nodes": [
-    {
-      "id": 20,
-      "type": "CLIPLoader",
-      "pos": {
-        "0": -26,
-        "1": 400
-      },
-      "size": {
-        "0": 451.30548095703125,
-        "1": 82
-      },
-      "flags": {},
-      "order": 0,
-      "mode": 0,
-      "inputs": [],
-      "outputs": [
-        {
-          "name": "CLIP",
-          "type": "CLIP",
-          "links": [
-            54,
-            56
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CLIPLoader"
-      },
-      "widgets_values": [
-        "t5\\google_t5-v1_1-xxl_encoderonly-fp8_e4m3fn.safetensors",
-        "sd3"
-      ]
-    },
     {
       "id": 31,
       "type": "CogVideoTextEncode",
@@ -46,16 +11,16 @@
       },
       "size": {
         "0": 463.01251220703125,
-        "1": 124
+        "1": 144
       },
       "flags": {},
-      "order": 4,
+      "order": 6,
       "mode": 0,
       "inputs": [
         {
           "name": "clip",
           "type": "CLIP",
-          "link": 56
+          "link": 149
         }
       ],
       "outputs": [
@@ -63,10 +28,15 @@
           "name": "conditioning",
           "type": "CONDITIONING",
           "links": [
-            123
+            146
           ],
           "slot_index": 0,
           "shape": 3
+        },
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "links": null
         }
       ],
       "properties": {
@@ -78,6 +48,208 @@
         true
       ]
     },
+    {
+      "id": 63,
+      "type": "CogVideoSampler",
+      "pos": {
+        "0": 1142,
+        "1": 74
+      },
+      "size": [
+        330,
+        574
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "COGVIDEOMODEL",
+          "link": 144
+        },
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 145
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 146
+        },
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "image_cond_latents",
+          "type": "LATENT",
+          "link": 147,
+          "shape": 7
+        },
+        {
+          "name": "context_options",
+          "type": "COGCONTEXT",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "controlnet",
+          "type": "COGVIDECONTROLNET",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "tora_trajectory",
+          "type": "TORAFEATURES",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "fastercache",
+          "type": "FASTERCACHEARGS",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "links": [
+            148
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "CogVideoSampler"
+      },
+      "widgets_values": [
+        49,
+        25,
+        6,
+        0,
+        "fixed",
+        "CogVideoXDDIM",
+        1
+      ]
+    },
+    {
+      "id": 62,
+      "type": "CogVideoImageEncode",
+      "pos": {
+        "0": 1149,
+        "1": 711
+      },
+      "size": {
+        "0": 315,
+        "1": 122
+      },
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 141
+        },
+        {
+          "name": "start_image",
+          "type": "IMAGE",
+          "link": 142
+        },
+        {
+          "name": "end_image",
+          "type": "IMAGE",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "links": [
+            147
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "CogVideoImageEncode"
+      },
+      "widgets_values": [
+        false,
+        0
+      ]
+    },
+    {
+      "id": 59,
+      "type": "DownloadAndLoadCogVideoModel",
+      "pos": {
+        "0": 622,
+        "1": -25
+      },
+      "size": {
+        "0": 315,
+        "1": 218
+      },
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "block_edit",
+          "type": "TRANSFORMERBLOCKS",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "lora",
+          "type": "COGLORA",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "compile_args",
+          "type": "COMPILEARGS",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "model",
+          "type": "COGVIDEOMODEL",
+          "links": [
+            144
+          ]
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "links": [
+            132,
+            141
+          ],
+          "slot_index": 1
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DownloadAndLoadCogVideoModel"
+      },
+      "widgets_values": [
+        "THUDM/CogVideoX-5b-I2V",
+        "bf16",
+        "disabled",
+        false,
+        "sdpa",
+        "main_device"
+      ]
+    },
     {
       "id": 30,
       "type": "CogVideoTextEncode",
@@ -90,7 +262,7 @@
         "1": 168.08047485351562
       },
       "flags": {},
-      "order": 3,
+      "order": 4,
       "mode": 0,
       "inputs": [
         {
@@ -104,10 +276,18 @@
           "name": "conditioning",
           "type": "CONDITIONING",
           "links": [
-            122
+            145
           ],
           "slot_index": 0,
           "shape": 3
+        },
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "links": [
+            149
+          ],
+          "slot_index": 1
         }
       ],
       "properties": {
@@ -116,22 +296,22 @@
       "widgets_values": [
         "a majestic stag is grazing in an enhanced forest, basking in the setting sun filtered by the trees",
         1,
-        true
+        false
       ]
     },
     {
       "id": 37,
       "type": "ImageResizeKJ",
       "pos": {
-        "0": 809,
-        "1": 684
+        "0": 784,
+        "1": 731
       },
       "size": {
         "0": 315,
         "1": 266
       },
       "flags": {},
-      "order": 5,
+      "order": 3,
       "mode": 0,
       "inputs": [
         {
@@ -142,7 +322,8 @@
         {
           "name": "get_image_size",
           "type": "IMAGE",
-          "link": null
+          "link": null,
+          "shape": 7
         },
         {
           "name": "width_input",
@@ -166,7 +347,7 @@
           "name": "IMAGE",
           "type": "IMAGE",
           "links": [
-            125
+            142
           ],
           "slot_index": 0,
           "shape": 3
@@ -198,182 +379,12 @@
         "disabled"
       ]
     },
-    {
-      "id": 58,
-      "type": "CogVideoImageEncode",
-      "pos": {
-        "0": 1156,
-        "1": 650
-      },
-      "size": {
-        "0": 315,
-        "1": 122
-      },
-      "flags": {},
-      "order": 6,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "pipeline",
-          "type": "COGVIDEOPIPE",
-          "link": 124
-        },
-        {
-          "name": "image",
-          "type": "IMAGE",
-          "link": 125
-        },
-        {
-          "name": "mask",
-          "type": "MASK",
-          "link": null
-        }
-      ],
-      "outputs": [
-        {
-          "name": "samples",
-          "type": "LATENT",
-          "links": [
-            129
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoImageEncode"
-      },
-      "widgets_values": [
-        16,
-        true
-      ]
-    },
-    {
-      "id": 56,
-      "type": "CogVideoDecode",
-      "pos": {
-        "0": 1581,
-        "1": 148
-      },
-      "size": {
-        "0": 300.396484375,
-        "1": 198
-      },
-      "flags": {},
-      "order": 8,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "pipeline",
-          "type": "COGVIDEOPIPE",
-          "link": 128
-        },
-        {
-          "name": "samples",
-          "type": "LATENT",
-          "link": 127
-        }
-      ],
-      "outputs": [
-        {
-          "name": "images",
-          "type": "IMAGE",
-          "links": [
-            118
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoDecode"
-      },
-      "widgets_values": [
-        false,
-        240,
-        360,
-        0.2,
-        0.2,
-        true
-      ]
-    },
-    {
-      "id": 44,
-      "type": "VHS_VideoCombine",
-      "pos": {
-        "0": 1927,
-        "1": 146
-      },
-      "size": [
-        605.3909912109375,
-        714.2606608072917
-      ],
-      "flags": {},
-      "order": 9,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "images",
-          "type": "IMAGE",
-          "link": 118
-        },
-        {
-          "name": "audio",
-          "type": "AUDIO",
-          "link": null
-        },
-        {
-          "name": "meta_batch",
-          "type": "VHS_BatchManager",
-          "link": null
-        },
-        {
-          "name": "vae",
-          "type": "VAE",
-          "link": null
-        }
-      ],
-      "outputs": [
-        {
-          "name": "Filenames",
-          "type": "VHS_FILENAMES",
-          "links": null,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "VHS_VideoCombine"
-      },
-      "widgets_values": {
-        "frame_rate": 8,
-        "loop_count": 0,
-        "filename_prefix": "CogVideoX-I2V",
-        "format": "video/h264-mp4",
-        "pix_fmt": "yuv420p",
-        "crf": 19,
-        "save_metadata": true,
-        "pingpong": false,
-        "save_output": false,
-        "videopreview": {
-          "hidden": false,
-          "paused": false,
-          "params": {
-            "filename": "CogVideoX-I2V_00001.mp4",
-            "subfolder": "",
-            "type": "temp",
-            "format": "video/h264-mp4",
-            "frame_rate": 8
-          },
-          "muted": false
-        }
-      }
-    },
     {
       "id": 36,
       "type": "LoadImage",
       "pos": {
-        "0": 365,
-        "1": 685
+        "0": 335,
+        "1": 731
       },
       "size": {
         "0": 402.06353759765625,
@@ -409,138 +420,159 @@
       ]
     },
     {
-      "id": 57,
-      "type": "CogVideoSampler",
+      "id": 20,
+      "type": "CLIPLoader",
       "pos": {
-        "0": 1138,
-        "1": 150
-      },
-      "size": [
-        399.878095897654,
-        350
-      ],
-      "flags": {},
-      "order": 7,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "pipeline",
-          "type": "COGVIDEOPIPE",
-          "link": 121
-        },
-        {
-          "name": "positive",
-          "type": "CONDITIONING",
-          "link": 122
-        },
-        {
-          "name": "negative",
-          "type": "CONDITIONING",
-          "link": 123
-        },
-        {
-          "name": "samples",
-          "type": "LATENT",
-          "link": null
-        },
-        {
-          "name": "image_cond_latents",
-          "type": "LATENT",
-          "link": 129
-        },
-        {
-          "name": "context_options",
-          "type": "COGCONTEXT",
-          "link": null
-        }
-      ],
-      "outputs": [
-        {
-          "name": "cogvideo_pipe",
-          "type": "COGVIDEOPIPE",
-          "links": [
-            128
-          ],
-          "slot_index": 0,
-          "shape": 3
-        },
-        {
-          "name": "samples",
-          "type": "LATENT",
-          "links": [
-            127
-          ],
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoSampler"
-      },
-      "widgets_values": [
-        480,
-        720,
-        49,
-        20,
-        6,
-        65334758276105,
-        "fixed",
-        "CogVideoXDPMScheduler",
-        1
-      ]
-    },
-    {
-      "id": 1,
-      "type": "DownloadAndLoadCogVideoModel",
-      "pos": {
-        "0": 633,
-        "1": 44
+        "0": -2,
+        "1": 304
       },
       "size": {
-        "0": 337.8885192871094,
-        "1": 194
+        "0": 451.30548095703125,
+        "1": 82
       },
       "flags": {},
       "order": 2,
       "mode": 0,
-      "inputs": [
-        {
-          "name": "pab_config",
-          "type": "PAB_CONFIG",
-          "link": null
-        },
-        {
-          "name": "block_edit",
-          "type": "TRANSFORMERBLOCKS",
-          "link": null
-        },
-        {
-          "name": "lora",
-          "type": "COGLORA",
-          "link": null
-        }
-      ],
+      "inputs": [],
       "outputs": [
         {
-          "name": "cogvideo_pipe",
-          "type": "COGVIDEOPIPE",
+          "name": "CLIP",
+          "type": "CLIP",
           "links": [
-            121,
-            124
+            54
           ],
           "slot_index": 0,
           "shape": 3
         }
       ],
       "properties": {
-        "Node name for S&R": "DownloadAndLoadCogVideoModel"
+        "Node name for S&R": "CLIPLoader"
       },
       "widgets_values": [
-        "THUDM/CogVideoX-5b-I2V",
-        "bf16",
-        "disabled",
-        "disabled",
-        false
+        "t5\\google_t5-v1_1-xxl_encoderonly-fp8_e4m3fn.safetensors",
+        "sd3"
       ]
+    },
+    {
+      "id": 60,
+      "type": "CogVideoDecode",
+      "pos": {
+        "0": 1523,
+        "1": -6
+      },
+      "size": {
+        "0": 315,
+        "1": 198
+      },
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 132
+        },
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 148
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            134
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "CogVideoDecode"
+      },
+      "widgets_values": [
+        true,
+        240,
+        360,
+        0.2,
+        0.2,
+        true
+      ]
+    },
+    {
+      "id": 44,
+      "type": "VHS_VideoCombine",
+      "pos": {
+        "0": 1884,
+        "1": -6
+      },
+      "size": [
+        605.3909912109375,
+        714.2606608072917
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 134
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "meta_batch",
+          "type": "VHS_BatchManager",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "Filenames",
+          "type": "VHS_FILENAMES",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VHS_VideoCombine"
+      },
+      "widgets_values": {
+        "frame_rate": 8,
+        "loop_count": 0,
+        "filename_prefix": "CogVideoX-I2V",
+        "format": "video/h264-mp4",
+        "pix_fmt": "yuv420p",
+        "crf": 19,
+        "save_metadata": true,
+        "pingpong": false,
+        "save_output": true,
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {
+            "filename": "CogVideoX-I2V_00001.mp4",
+            "subfolder": "",
+            "type": "temp",
+            "format": "video/h264-mp4",
+            "frame_rate": 8
+          },
+          "muted": false
+        }
+      }
     }
   ],
   "links": [
@@ -552,14 +584,6 @@
       0,
       "CLIP"
     ],
-    [
-      56,
-      20,
-      0,
-      31,
-      0,
-      "CLIP"
-    ],
     [
       71,
       36,
@@ -569,86 +593,94 @@
       "IMAGE"
     ],
     [
-      118,
-      56,
+      132,
+      59,
+      1,
+      60,
+      0,
+      "VAE"
+    ],
+    [
+      134,
+      60,
       0,
       44,
       0,
       "IMAGE"
     ],
     [
-      121,
+      141,
+      59,
       1,
+      62,
       0,
-      57,
-      0,
-      "COGVIDEOPIPE"
+      "VAE"
     ],
     [
-      122,
-      30,
-      0,
-      57,
-      1,
-      "CONDITIONING"
-    ],
-    [
-      123,
-      31,
-      0,
-      57,
-      2,
-      "CONDITIONING"
-    ],
-    [
-      124,
-      1,
-      0,
-      58,
-      0,
-      "COGVIDEOPIPE"
-    ],
-    [
-      125,
+      142,
       37,
       0,
-      58,
+      62,
       1,
       "IMAGE"
     ],
     [
-      127,
-      57,
-      1,
-      56,
-      1,
-      "LATENT"
+      144,
+      59,
+      0,
+      63,
+      0,
+      "COGVIDEOMODEL"
     ],
     [
-      128,
-      57,
+      145,
+      30,
       0,
-      56,
-      0,
-      "COGVIDEOPIPE"
+      63,
+      1,
+      "CONDITIONING"
     ],
     [
-      129,
-      58,
+      146,
+      31,
       0,
-      57,
+      63,
+      2,
+      "CONDITIONING"
+    ],
+    [
+      147,
+      62,
+      0,
+      63,
       4,
       "LATENT"
+    ],
+    [
+      148,
+      63,
+      0,
+      60,
+      1,
+      "LATENT"
+    ],
+    [
+      149,
+      30,
+      1,
+      31,
+      0,
+      "CLIP"
     ]
   ],
   "groups": [],
   "config": {},
   "extra": {
     "ds": {
-      "scale": 0.6934334949442514,
+      "scale": 0.7627768444387059,
       "offset": [
-        -24.154349208343916,
-        155.20539218330134
+        648.7113591814891,
+        185.9907078691075
       ]
     }
   },

example_workflows/cogvideox_1_0_5b_T2V_02.json

@@ -1,48 +1,7 @@
 {
-  "last_node_id": 34,
-  "last_link_id": 64,
+  "last_node_id": 37,
+  "last_link_id": 72,
   "nodes": [
-    {
-      "id": 31,
-      "type": "CogVideoTextEncode",
-      "pos": {
-        "0": 503,
-        "1": 521
-      },
-      "size": {
-        "0": 463.01251220703125,
-        "1": 124
-      },
-      "flags": {},
-      "order": 3,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "clip",
-          "type": "CLIP",
-          "link": 56
-        }
-      ],
-      "outputs": [
-        {
-          "name": "conditioning",
-          "type": "CONDITIONING",
-          "links": [
-            62
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoTextEncode"
-      },
-      "widgets_values": [
-        "",
-        1,
-        true
-      ]
-    },
     {
       "id": 30,
       "type": "CogVideoTextEncode",
@@ -50,12 +9,12 @@
         "0": 500,
         "1": 308
       },
-      "size": {
-        "0": 471.90142822265625,
-        "1": 168.08047485351562
-      },
+      "size": [
+        470.99399664051055,
+        237.5088638951354
+      ],
       "flags": {},
-      "order": 2,
+      "order": 3,
       "mode": 0,
       "inputs": [
         {
@@ -69,10 +28,18 @@
           "name": "conditioning",
           "type": "CONDITIONING",
           "links": [
-            61
+            67
           ],
           "slot_index": 0,
           "shape": 3
+        },
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "links": [
+            65
+          ],
+          "slot_index": 1
         }
       ],
       "properties": {
@@ -81,192 +48,79 @@
       "widgets_values": [
         "A golden retriever, sporting sleek black sunglasses, with its lengthy fur flowing in the breeze, sprints playfully across a rooftop terrace, recently refreshed by a light rain. The scene unfolds from a distance, the dog's energetic bounds growing larger as it approaches the camera, its tail wagging with unrestrained joy, while droplets of water glisten on the concrete behind it. The overcast sky provides a dramatic backdrop, emphasizing the vibrant golden coat of the canine as it dashes towards the viewer.\n\n",
         1,
-        true
+        false
       ]
     },
     {
-      "id": 33,
-      "type": "VHS_VideoCombine",
+      "id": 31,
+      "type": "CogVideoTextEncode",
       "pos": {
-        "0": 1441,
-        "1": 129
+        "0": 503,
+        "1": 602
       },
       "size": [
-        778.7022705078125,
-        310
+        464.4980515341475,
+        169.87479027400514
       ],
       "flags": {},
-      "order": 6,
+      "order": 4,
       "mode": 0,
       "inputs": [
         {
-          "name": "images",
-          "type": "IMAGE",
-          "link": 59
-        },
-        {
-          "name": "audio",
-          "type": "AUDIO",
-          "link": null
-        },
-        {
-          "name": "meta_batch",
-          "type": "VHS_BatchManager",
-          "link": null
-        },
-        {
-          "name": "vae",
-          "type": "VAE",
-          "link": null
-        }
-      ],
-      "outputs": [
-        {
-          "name": "Filenames",
-          "type": "VHS_FILENAMES",
-          "links": null,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "VHS_VideoCombine"
-      },
-      "widgets_values": {
-        "frame_rate": 8,
-        "loop_count": 0,
-        "filename_prefix": "CogVideoX5B",
-        "format": "video/h264-mp4",
-        "pix_fmt": "yuv420p",
-        "crf": 19,
-        "save_metadata": true,
-        "pingpong": false,
-        "save_output": false,
-        "videopreview": {
-          "hidden": false,
-          "paused": false,
-          "params": {
-            "filename": "CogVideoX5B_00009.mp4",
-            "subfolder": "",
-            "type": "temp",
-            "format": "video/h264-mp4",
-            "frame_rate": 8
-          },
-          "muted": false
-        }
-      }
-    },
-    {
-      "id": 20,
-      "type": "CLIPLoader",
-      "pos": {
-        "0": -26,
-        "1": 400
-      },
-      "size": {
-        "0": 451.30548095703125,
-        "1": 82
-      },
-      "flags": {},
-      "order": 0,
-      "mode": 0,
-      "inputs": [],
-      "outputs": [
-        {
-          "name": "CLIP",
+          "name": "clip",
           "type": "CLIP",
-          "links": [
-            54,
-            56
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CLIPLoader"
-      },
-      "widgets_values": [
-        "t5\\google_t5-v1_1-xxl_encoderonly-fp8_e4m3fn.safetensors",
-        "sd3"
-      ]
-    },
-    {
-      "id": 1,
-      "type": "DownloadAndLoadCogVideoModel",
-      "pos": {
-        "0": 642,
-        "1": 90
-      },
-      "size": {
-        "0": 315,
-        "1": 194
-      },
-      "flags": {},
-      "order": 1,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "pab_config",
-          "type": "PAB_CONFIG",
-          "link": null
-        },
-        {
-          "name": "block_edit",
-          "type": "TRANSFORMERBLOCKS",
-          "link": null
-        },
-        {
-          "name": "lora",
-          "type": "COGLORA",
-          "link": null
+          "link": 65
         }
       ],
       "outputs": [
         {
-          "name": "cogvideo_pipe",
-          "type": "COGVIDEOPIPE",
+          "name": "conditioning",
+          "type": "CONDITIONING",
           "links": [
-            60
+            68
           ],
           "slot_index": 0,
           "shape": 3
+        },
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "links": null
         }
       ],
       "properties": {
-        "Node name for S&R": "DownloadAndLoadCogVideoModel"
+        "Node name for S&R": "CogVideoTextEncode"
       },
       "widgets_values": [
-        "THUDM/CogVideoX-5b",
-        "bf16",
-        "disabled",
-        "disabled",
-        false
+        "",
+        1,
+        true
       ]
     },
     {
       "id": 11,
       "type": "CogVideoDecode",
       "pos": {
-        "0": 1051,
-        "1": 748
+        "0": 1416,
+        "1": 40
       },
       "size": {
         "0": 300.396484375,
         "1": 198
       },
       "flags": {},
-      "order": 5,
+      "order": 6,
       "mode": 0,
       "inputs": [
         {
-          "name": "pipeline",
-          "type": "COGVIDEOPIPE",
-          "link": 63
+          "name": "vae",
+          "type": "VAE",
+          "link": 71
         },
         {
           "name": "samples",
           "type": "LATENT",
-          "link": 64
+          "link": 69
         }
       ],
       "outputs": [
@@ -293,83 +147,297 @@
       ]
     },
     {
-      "id": 34,
-      "type": "CogVideoSampler",
+      "id": 36,
+      "type": "DownloadAndLoadCogVideoModel",
       "pos": {
-        "0": 1041,
-        "1": 342
+        "0": 645,
+        "1": 17
       },
       "size": {
-        "0": 315.8404846191406,
-        "1": 358
+        "0": 315,
+        "1": 218
       },
       "flags": {},
-      "order": 4,
+      "order": 0,
       "mode": 0,
       "inputs": [
         {
-          "name": "pipeline",
-          "type": "COGVIDEOPIPE",
-          "link": 60
+          "name": "block_edit",
+          "type": "TRANSFORMERBLOCKS",
+          "link": null,
+          "shape": 7
         },
         {
-          "name": "positive",
-          "type": "CONDITIONING",
-          "link": 61
+          "name": "lora",
+          "type": "COGLORA",
+          "link": null,
+          "shape": 7
         },
         {
-          "name": "negative",
-          "type": "CONDITIONING",
-          "link": 62
-        },
-        {
-          "name": "samples",
-          "type": "LATENT",
-          "link": null
-        },
-        {
-          "name": "image_cond_latents",
-          "type": "LATENT",
-          "link": null
-        },
-        {
-          "name": "context_options",
-          "type": "COGCONTEXT",
-          "link": null
+          "name": "compile_args",
+          "type": "COMPILEARGS",
+          "link": null,
+          "shape": 7
         }
       ],
       "outputs": [
         {
-          "name": "cogvideo_pipe",
-          "type": "COGVIDEOPIPE",
+          "name": "model",
+          "type": "COGVIDEOMODEL",
           "links": [
-            63
+            70
+          ]
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "links": [
+            71
           ],
+          "slot_index": 1
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DownloadAndLoadCogVideoModel"
+      },
+      "widgets_values": [
+        "THUDM/CogVideoX-5b",
+        "bf16",
+        "disabled",
+        false,
+        "sdpa",
+        "main_device"
+      ]
+    },
+    {
+      "id": 20,
+      "type": "CLIPLoader",
+      "pos": {
+        "0": 5,
+        "1": 308
+      },
+      "size": {
+        "0": 451.30548095703125,
+        "1": 82
+      },
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "CLIP",
+          "type": "CLIP",
+          "links": [
+            54
+          ],
+          "slot_index": 0,
           "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "CLIPLoader"
+      },
+      "widgets_values": [
+        "t5\\google_t5-v1_1-xxl_encoderonly-fp8_e4m3fn.safetensors",
+        "sd3"
+      ]
+    },
+    {
+      "id": 37,
+      "type": "EmptyLatentImage",
+      "pos": {
+        "0": 643,
+        "1": 827
+      },
+      "size": {
+        "0": 315,
+        "1": 106
+      },
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "LATENT",
+          "type": "LATENT",
+          "links": [
+            72
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "EmptyLatentImage"
+      },
+      "widgets_values": [
+        720,
+        480,
+        1
+      ]
+    },
+    {
+      "id": 35,
+      "type": "CogVideoSampler",
+      "pos": {
+        "0": 1042,
+        "1": 291
+      },
+      "size": [
+        330,
+        574
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "COGVIDEOMODEL",
+          "link": 70
+        },
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 67
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 68
         },
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 72,
+          "shape": 7
+        },
+        {
+          "name": "image_cond_latents",
+          "type": "LATENT",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "context_options",
+          "type": "COGCONTEXT",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "controlnet",
+          "type": "COGVIDECONTROLNET",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "tora_trajectory",
+          "type": "TORAFEATURES",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "fastercache",
+          "type": "FASTERCACHEARGS",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
         {
           "name": "samples",
           "type": "LATENT",
           "links": [
-            64
-          ],
-          "shape": 3
+            69
+          ]
         }
       ],
       "properties": {
         "Node name for S&R": "CogVideoSampler"
       },
       "widgets_values": [
-        480,
-        720,
         49,
         50,
         6,
-        806286757407563,
+        0,
         "fixed",
-        "DPM++",
+        "CogVideoXDDIM",
         1
       ]
+    },
+    {
+      "id": 33,
+      "type": "VHS_VideoCombine",
+      "pos": {
+        "0": 1767,
+        "1": 39
+      },
+      "size": [
+        778.7022705078125,
+        829.801513671875
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 59
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "meta_batch",
+          "type": "VHS_BatchManager",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "Filenames",
+          "type": "VHS_FILENAMES",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VHS_VideoCombine"
+      },
+      "widgets_values": {
+        "frame_rate": 8,
+        "loop_count": 0,
+        "filename_prefix": "CogVideoX5B-T2V",
+        "format": "video/h264-mp4",
+        "pix_fmt": "yuv420p",
+        "crf": 19,
+        "save_metadata": true,
+        "pingpong": false,
+        "save_output": false,
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {
+            "filename": "CogVideoX5B_00001.mp4",
+            "subfolder": "",
+            "type": "temp",
+            "format": "video/h264-mp4",
+            "frame_rate": 8
+          },
+          "muted": false
+        }
+      }
     }
   ],
   "links": [
@@ -381,14 +449,6 @@
       0,
       "CLIP"
     ],
-    [
-      56,
-      20,
-      0,
-      31,
-      0,
-      "CLIP"
-    ],
     [
       59,
       11,
@@ -398,43 +458,59 @@
       "IMAGE"
     ],
     [
-      60,
+      65,
+      30,
       1,
+      31,
       0,
-      34,
-      0,
-      "COGVIDEOPIPE"
+      "CLIP"
     ],
     [
-      61,
+      67,
       30,
       0,
-      34,
+      35,
       1,
       "CONDITIONING"
     ],
     [
-      62,
+      68,
       31,
       0,
-      34,
+      35,
       2,
       "CONDITIONING"
     ],
     [
-      63,
-      34,
+      69,
+      35,
       0,
       11,
-      0,
-      "COGVIDEOPIPE"
+      1,
+      "LATENT"
     ],
     [
-      64,
-      34,
+      70,
+      36,
+      0,
+      35,
+      0,
+      "COGVIDEOMODEL"
+    ],
+    [
+      71,
+      36,
       1,
       11,
-      1,
+      0,
+      "VAE"
+    ],
+    [
+      72,
+      37,
+      0,
+      35,
+      3,
       "LATENT"
     ]
   ],
@@ -442,10 +518,10 @@
   "config": {},
   "extra": {
     "ds": {
-      "scale": 0.6934334949442514,
+      "scale": 0.7627768444387061,
       "offset": [
-        -24.154349208343916,
-        155.20539218330134
+        734.1791945221892,
+        237.29437844909364
       ]
     }
   },

example_workflows/cogvideox_1_5_5b_I2V_01.json

@@ -1,60 +1,164 @@
 {
-  "last_node_id": 51,
-  "last_link_id": 114,
+  "last_node_id": 64,
+  "last_link_id": 149,
   "nodes": [
     {
-      "id": 20,
-      "type": "CLIPLoader",
+      "id": 63,
+      "type": "CogVideoSampler",
       "pos": {
-        "0": -26,
-        "1": 400
+        "0": 1142,
+        "1": 74
       },
       "size": {
-        "0": 451.30548095703125,
-        "1": 82
+        "0": 330,
+        "1": 574
       },
       "flags": {},
-      "order": 0,
+      "order": 7,
       "mode": 0,
-      "inputs": [],
+      "inputs": [
+        {
+          "name": "model",
+          "type": "COGVIDEOMODEL",
+          "link": 144
+        },
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 145
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 146
+        },
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "image_cond_latents",
+          "type": "LATENT",
+          "link": 147,
+          "shape": 7
+        },
+        {
+          "name": "context_options",
+          "type": "COGCONTEXT",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "controlnet",
+          "type": "COGVIDECONTROLNET",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "tora_trajectory",
+          "type": "TORAFEATURES",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "fastercache",
+          "type": "FASTERCACHEARGS",
+          "link": null,
+          "shape": 7
+        }
+      ],
       "outputs": [
         {
-          "name": "CLIP",
-          "type": "CLIP",
+          "name": "samples",
+          "type": "LATENT",
           "links": [
-            54
-          ],
-          "slot_index": 0,
-          "shape": 3
+            148
+          ]
         }
       ],
       "properties": {
-        "Node name for S&R": "CLIPLoader"
+        "Node name for S&R": "CogVideoSampler"
       },
       "widgets_values": [
-        "t5\\google_t5-v1_1-xxl_encoderonly-fp8_e4m3fn.safetensors",
-        "sd3"
+        49,
+        25,
+        6,
+        0,
+        "fixed",
+        "CogVideoXDDIM",
+        1
       ]
     },
     {
-      "id": 31,
-      "type": "CogVideoTextEncode",
+      "id": 62,
+      "type": "CogVideoImageEncode",
       "pos": {
-        "0": 497,
-        "1": 520
+        "0": 1149,
+        "1": 711
       },
       "size": {
-        "0": 463.01251220703125,
-        "1": 144
+        "0": 315,
+        "1": 122
       },
       "flags": {},
       "order": 5,
       "mode": 0,
+      "inputs": [
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 141
+        },
+        {
+          "name": "start_image",
+          "type": "IMAGE",
+          "link": 142
+        },
+        {
+          "name": "end_image",
+          "type": "IMAGE",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "links": [
+            147
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "CogVideoImageEncode"
+      },
+      "widgets_values": [
+        false,
+        0
+      ]
+    },
+    {
+      "id": 30,
+      "type": "CogVideoTextEncode",
+      "pos": {
+        "0": 493,
+        "1": 303
+      },
+      "size": {
+        "0": 471.90142822265625,
+        "1": 168.08047485351562
+      },
+      "flags": {},
+      "order": 4,
+      "mode": 0,
       "inputs": [
         {
           "name": "clip",
           "type": "CLIP",
-          "link": 108
+          "link": 54
         }
       ],
       "outputs": [
@@ -62,7 +166,7 @@
           "name": "conditioning",
           "type": "CONDITIONING",
           "links": [
-            111
+            145
           ],
           "slot_index": 0,
           "shape": 3
@@ -70,105 +174,34 @@
         {
           "name": "clip",
           "type": "CLIP",
-          "links": null
+          "links": [
+            149
+          ],
+          "slot_index": 1
         }
       ],
       "properties": {
         "Node name for S&R": "CogVideoTextEncode"
       },
       "widgets_values": [
-        "The video is not of a high quality, it has a low resolution. Watermark present in each frame. Strange motion trajectory. ",
+        "a majestic stag is grazing in an enhanced forest, basking in the setting sun filtered by the trees",
         1,
-        true
+        false
       ]
     },
-    {
-      "id": 44,
-      "type": "VHS_VideoCombine",
-      "pos": {
-        "0": 1842,
-        "1": 345
-      },
-      "size": [
-        855.81494140625,
-        881.2099609375
-      ],
-      "flags": {},
-      "order": 8,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "images",
-          "type": "IMAGE",
-          "link": 97
-        },
-        {
-          "name": "audio",
-          "type": "AUDIO",
-          "link": null,
-          "shape": 7
-        },
-        {
-          "name": "meta_batch",
-          "type": "VHS_BatchManager",
-          "link": null,
-          "shape": 7
-        },
-        {
-          "name": "vae",
-          "type": "VAE",
-          "link": null,
-          "shape": 7
-        }
-      ],
-      "outputs": [
-        {
-          "name": "Filenames",
-          "type": "VHS_FILENAMES",
-          "links": null,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "VHS_VideoCombine"
-      },
-      "widgets_values": {
-        "frame_rate": 16,
-        "loop_count": 0,
-        "filename_prefix": "CogVideoX_Fun",
-        "format": "video/h264-mp4",
-        "pix_fmt": "yuv420p",
-        "crf": 19,
-        "save_metadata": true,
-        "pingpong": false,
-        "save_output": false,
-        "videopreview": {
-          "hidden": false,
-          "paused": false,
-          "params": {
-            "filename": "CogVideoX_Fun_00003.mp4",
-            "subfolder": "",
-            "type": "temp",
-            "format": "video/h264-mp4",
-            "frame_rate": 16
-          },
-          "muted": false
-        }
-      }
-    },
     {
       "id": 36,
       "type": "LoadImage",
       "pos": {
-        "0": 227,
-        "1": 700
+        "0": 335,
+        "1": 731
       },
       "size": {
-        "0": 391.3421325683594,
-        "1": 456.8497009277344
+        "0": 402.06353759765625,
+        "1": 396.6225891113281
       },
       "flags": {},
-      "order": 1,
+      "order": 0,
       "mode": 0,
       "inputs": [],
       "outputs": [
@@ -196,19 +229,100 @@
         "image"
       ]
     },
+    {
+      "id": 20,
+      "type": "CLIPLoader",
+      "pos": {
+        "0": -2,
+        "1": 304
+      },
+      "size": {
+        "0": 451.30548095703125,
+        "1": 82
+      },
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "CLIP",
+          "type": "CLIP",
+          "links": [
+            54
+          ],
+          "slot_index": 0,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "CLIPLoader"
+      },
+      "widgets_values": [
+        "t5\\google_t5-v1_1-xxl_encoderonly-fp8_e4m3fn.safetensors",
+        "sd3"
+      ]
+    },
+    {
+      "id": 60,
+      "type": "CogVideoDecode",
+      "pos": {
+        "0": 1523,
+        "1": -6
+      },
+      "size": {
+        "0": 315,
+        "1": 198
+      },
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 132
+        },
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 148
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            134
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "CogVideoDecode"
+      },
+      "widgets_values": [
+        true,
+        240,
+        360,
+        0.2,
+        0.2,
+        true
+      ]
+    },
     {
       "id": 37,
       "type": "ImageResizeKJ",
       "pos": {
-        "0": 688,
-        "1": 708
+        "0": 784,
+        "1": 731
       },
       "size": {
         "0": 315,
         "1": 266
       },
       "flags": {},
-      "order": 4,
+      "order": 3,
       "mode": 0,
       "inputs": [
         {
@@ -244,7 +358,7 @@
           "name": "IMAGE",
           "type": "IMAGE",
           "links": [
-            112
+            142
           ],
           "slot_index": 0,
           "shape": 3
@@ -266,10 +380,10 @@
         "Node name for S&R": "ImageResizeKJ"
       },
       "widgets_values": [
-        720,
-        480,
+        1360,
+        768,
         "lanczos",
-        true,
+        false,
         16,
         0,
         0,
@@ -277,73 +391,24 @@
       ]
     },
     {
-      "id": 11,
-      "type": "CogVideoDecode",
-      "pos": {
-        "0": 1477,
-        "1": 344
-      },
-      "size": {
-        "0": 300.396484375,
-        "1": 198
-      },
-      "flags": {},
-      "order": 7,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "pipeline",
-          "type": "COGVIDEOPIPE",
-          "link": 113
-        },
-        {
-          "name": "samples",
-          "type": "LATENT",
-          "link": 114
-        }
-      ],
-      "outputs": [
-        {
-          "name": "images",
-          "type": "IMAGE",
-          "links": [
-            97
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoDecode"
-      },
-      "widgets_values": [
-        true,
-        240,
-        360,
-        0.2,
-        0.2,
-        true
-      ]
-    },
-    {
-      "id": 30,
+      "id": 31,
       "type": "CogVideoTextEncode",
       "pos": {
-        "0": 493,
-        "1": 303
+        "0": 497,
+        "1": 520
       },
       "size": {
-        "0": 471.90142822265625,
-        "1": 168.08047485351562
+        "0": 463.01251220703125,
+        "1": 144
       },
       "flags": {},
-      "order": 3,
+      "order": 6,
       "mode": 0,
       "inputs": [
         {
           "name": "clip",
           "type": "CLIP",
-          "link": 54
+          "link": 149
         }
       ],
       "outputs": [
@@ -351,7 +416,7 @@
           "name": "conditioning",
           "type": "CONDITIONING",
           "links": [
-            110
+            146
           ],
           "slot_index": 0,
           "shape": 3
@@ -359,170 +424,155 @@
         {
           "name": "clip",
           "type": "CLIP",
-          "links": [
-            108
-          ],
-          "slot_index": 1
+          "links": null
         }
       ],
       "properties": {
         "Node name for S&R": "CogVideoTextEncode"
       },
       "widgets_values": [
-        "majestic stag grazing in a forest and basking in the setting sun",
+        "",
         1,
-        false
+        true
       ]
     },
     {
-      "id": 51,
-      "type": "CogVideoXFunSampler",
+      "id": 59,
+      "type": "DownloadAndLoadCogVideoModel",
       "pos": {
-        "0": 1058,
-        "1": 345
+        "0": 622,
+        "1": -25
       },
       "size": {
-        "0": 367.79998779296875,
-        "1": 434
-      },
-      "flags": {},
-      "order": 6,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "pipeline",
-          "type": "COGVIDEOPIPE",
-          "link": 109
-        },
-        {
-          "name": "positive",
-          "type": "CONDITIONING",
-          "link": 110
-        },
-        {
-          "name": "negative",
-          "type": "CONDITIONING",
-          "link": 111
-        },
-        {
-          "name": "start_img",
-          "type": "IMAGE",
-          "link": 112,
-          "shape": 7
-        },
-        {
-          "name": "end_img",
-          "type": "IMAGE",
-          "link": null,
-          "shape": 7
-        },
-        {
-          "name": "context_options",
-          "type": "COGCONTEXT",
-          "link": null,
-          "shape": 7
-        },
-        {
-          "name": "tora_trajectory",
-          "type": "TORAFEATURES",
-          "link": null,
-          "shape": 7
-        },
-        {
-          "name": "fastercache",
-          "type": "FASTERCACHEARGS",
-          "link": null,
-          "shape": 7
-        },
-        {
-          "name": "vid2vid_images",
-          "type": "IMAGE",
-          "link": null,
-          "shape": 7
-        }
-      ],
-      "outputs": [
-        {
-          "name": "cogvideo_pipe",
-          "type": "COGVIDEOPIPE",
-          "links": [
-            113
-          ]
-        },
-        {
-          "name": "samples",
-          "type": "LATENT",
-          "links": [
-            114
-          ]
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoXFunSampler"
-      },
-      "widgets_values": [
-        49,
-        720,
-        480,
-        43,
-        "randomize",
-        50,
-        6,
-        "DDIM",
-        0.0563,
-        1
-      ]
-    },
-    {
-      "id": 48,
-      "type": "DownloadAndLoadCogVideoGGUFModel",
-      "pos": {
-        "0": 585,
-        "1": 34
-      },
-      "size": {
-        "0": 378,
-        "1": 198
+        "0": 315,
+        "1": 218
       },
       "flags": {},
       "order": 2,
       "mode": 0,
       "inputs": [
         {
-          "name": "pab_config",
-          "type": "PAB_CONFIG",
+          "name": "block_edit",
+          "type": "TRANSFORMERBLOCKS",
           "link": null,
           "shape": 7
         },
         {
-          "name": "block_edit",
-          "type": "TRANSFORMERBLOCKS",
+          "name": "lora",
+          "type": "COGLORA",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "compile_args",
+          "type": "COMPILEARGS",
           "link": null,
           "shape": 7
         }
       ],
       "outputs": [
         {
-          "name": "cogvideo_pipe",
-          "type": "COGVIDEOPIPE",
+          "name": "model",
+          "type": "COGVIDEOMODEL",
           "links": [
-            109
+            144
+          ]
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "links": [
+            132,
+            141
           ],
-          "slot_index": 0,
+          "slot_index": 1
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DownloadAndLoadCogVideoModel"
+      },
+      "widgets_values": [
+        "kijai/CogVideoX-5b-1.5-I2V",
+        "bf16",
+        "disabled",
+        false,
+        "sdpa",
+        "main_device"
+      ]
+    },
+    {
+      "id": 44,
+      "type": "VHS_VideoCombine",
+      "pos": {
+        "0": 1884,
+        "1": -6
+      },
+      "size": [
+        605.3909912109375,
+        310
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 134
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "meta_batch",
+          "type": "VHS_BatchManager",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "Filenames",
+          "type": "VHS_FILENAMES",
+          "links": null,
           "shape": 3
         }
       ],
       "properties": {
-        "Node name for S&R": "DownloadAndLoadCogVideoGGUFModel"
+        "Node name for S&R": "VHS_VideoCombine"
       },
-      "widgets_values": [
-        "CogVideoX_5b_fun_1_1_GGUF_Q4_0.safetensors",
-        "bf16",
-        false,
-        "offload_device",
-        false,
-        "disabled"
-      ]
+      "widgets_values": {
+        "frame_rate": 16,
+        "loop_count": 0,
+        "filename_prefix": "CogVideoX_1_5_I2V",
+        "format": "video/h264-mp4",
+        "pix_fmt": "yuv420p",
+        "crf": 19,
+        "save_metadata": true,
+        "pingpong": false,
+        "save_output": true,
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {
+            "filename": "CogVideoX-I2V_00004.mp4",
+            "subfolder": "",
+            "type": "temp",
+            "format": "video/h264-mp4",
+            "frame_rate": 8
+          },
+          "muted": false
+        }
+      }
     }
   ],
   "links": [
@@ -543,78 +593,94 @@
       "IMAGE"
     ],
     [
-      97,
-      11,
+      132,
+      59,
+      1,
+      60,
+      0,
+      "VAE"
+    ],
+    [
+      134,
+      60,
       0,
       44,
       0,
       "IMAGE"
     ],
     [
-      108,
+      141,
+      59,
+      1,
+      62,
+      0,
+      "VAE"
+    ],
+    [
+      142,
+      37,
+      0,
+      62,
+      1,
+      "IMAGE"
+    ],
+    [
+      144,
+      59,
+      0,
+      63,
+      0,
+      "COGVIDEOMODEL"
+    ],
+    [
+      145,
+      30,
+      0,
+      63,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      146,
+      31,
+      0,
+      63,
+      2,
+      "CONDITIONING"
+    ],
+    [
+      147,
+      62,
+      0,
+      63,
+      4,
+      "LATENT"
+    ],
+    [
+      148,
+      63,
+      0,
+      60,
+      1,
+      "LATENT"
+    ],
+    [
+      149,
       30,
       1,
       31,
       0,
       "CLIP"
-    ],
-    [
-      109,
-      48,
-      0,
-      51,
-      0,
-      "COGVIDEOPIPE"
-    ],
-    [
-      110,
-      30,
-      0,
-      51,
-      1,
-      "CONDITIONING"
-    ],
-    [
-      111,
-      31,
-      0,
-      51,
-      2,
-      "CONDITIONING"
-    ],
-    [
-      112,
-      37,
-      0,
-      51,
-      3,
-      "IMAGE"
-    ],
-    [
-      113,
-      51,
-      0,
-      11,
-      0,
-      "COGVIDEOPIPE"
-    ],
-    [
-      114,
-      51,
-      1,
-      11,
-      1,
-      "LATENT"
     ]
   ],
   "groups": [],
   "config": {},
   "extra": {
     "ds": {
-      "scale": 0.7513148009015784,
+      "scale": 0.7627768444387097,
       "offset": [
-        724.7448506313632,
-        128.336592104936
+        716.7143770104391,
+        291.75859557289965
       ]
     }
   },

examples/cogvideo_2b_context_schedule_test_01.json

@@ -1,561 +0,0 @@
-{
-  "last_node_id": 34,
-  "last_link_id": 61,
-  "nodes": [
-    {
-      "id": 33,
-      "type": "GetImageSizeAndCount",
-      "pos": {
-        "0": 1176,
-        "1": 122
-      },
-      "size": {
-        "0": 210,
-        "1": 86
-      },
-      "flags": {},
-      "order": 7,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "image",
-          "type": "IMAGE",
-          "link": 59
-        }
-      ],
-      "outputs": [
-        {
-          "name": "image",
-          "type": "IMAGE",
-          "links": [
-            60
-          ],
-          "slot_index": 0,
-          "shape": 3
-        },
-        {
-          "name": "720 width",
-          "type": "INT",
-          "links": null,
-          "shape": 3
-        },
-        {
-          "name": "480 height",
-          "type": "INT",
-          "links": null,
-          "shape": 3
-        },
-        {
-          "name": "104 count",
-          "type": "INT",
-          "links": null,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "GetImageSizeAndCount"
-      },
-      "widgets_values": []
-    },
-    {
-      "id": 30,
-      "type": "CogVideoTextEncode",
-      "pos": {
-        "0": 500,
-        "1": 308
-      },
-      "size": [
-        474.8035864085422,
-        211.10369504535595
-      ],
-      "flags": {},
-      "order": 3,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "clip",
-          "type": "CLIP",
-          "link": 54
-        }
-      ],
-      "outputs": [
-        {
-          "name": "conditioning",
-          "type": "CONDITIONING",
-          "links": [
-            55
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoTextEncode"
-      },
-      "widgets_values": [
-        "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature\nacoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters\nthrough the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The\nbackground includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical\nperformance.",
-        1,
-        true
-      ]
-    },
-    {
-      "id": 31,
-      "type": "CogVideoTextEncode",
-      "pos": {
-        "0": 508,
-        "1": 576
-      },
-      "size": {
-        "0": 463.01251220703125,
-        "1": 124
-      },
-      "flags": {},
-      "order": 4,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "clip",
-          "type": "CLIP",
-          "link": 56
-        }
-      ],
-      "outputs": [
-        {
-          "name": "conditioning",
-          "type": "CONDITIONING",
-          "links": [
-            57
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoTextEncode"
-      },
-      "widgets_values": [
-        "",
-        1,
-        true
-      ]
-    },
-    {
-      "id": 20,
-      "type": "CLIPLoader",
-      "pos": {
-        "0": -37,
-        "1": 443
-      },
-      "size": {
-        "0": 451.30548095703125,
-        "1": 82
-      },
-      "flags": {},
-      "order": 0,
-      "mode": 0,
-      "inputs": [],
-      "outputs": [
-        {
-          "name": "CLIP",
-          "type": "CLIP",
-          "links": [
-            54,
-            56
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CLIPLoader"
-      },
-      "widgets_values": [
-        "t5\\google_t5-v1_1-xxl_encoderonly-fp8_e4m3fn.safetensors",
-        "sd3"
-      ]
-    },
-    {
-      "id": 11,
-      "type": "CogVideoDecode",
-      "pos": {
-        "0": 1045,
-        "1": 776
-      },
-      "size": {
-        "0": 295.70111083984375,
-        "1": 198
-      },
-      "flags": {},
-      "order": 6,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "pipeline",
-          "type": "COGVIDEOPIPE",
-          "link": 37
-        },
-        {
-          "name": "samples",
-          "type": "LATENT",
-          "link": 38
-        }
-      ],
-      "outputs": [
-        {
-          "name": "images",
-          "type": "IMAGE",
-          "links": [
-            59
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoDecode"
-      },
-      "widgets_values": [
-        true,
-        96,
-        96,
-        0.083,
-        0.083,
-        true
-      ]
-    },
-    {
-      "id": 1,
-      "type": "DownloadAndLoadCogVideoModel",
-      "pos": {
-        "0": 652,
-        "1": 43
-      },
-      "size": {
-        "0": 315,
-        "1": 194
-      },
-      "flags": {},
-      "order": 1,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "pab_config",
-          "type": "PAB_CONFIG",
-          "link": null
-        },
-        {
-          "name": "block_edit",
-          "type": "TRANSFORMERBLOCKS",
-          "link": null
-        },
-        {
-          "name": "lora",
-          "type": "COGLORA",
-          "link": null
-        }
-      ],
-      "outputs": [
-        {
-          "name": "cogvideo_pipe",
-          "type": "COGVIDEOPIPE",
-          "links": [
-            36
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "DownloadAndLoadCogVideoModel"
-      },
-      "widgets_values": [
-        "THUDM/CogVideoX-2b",
-        "fp16",
-        "enabled",
-        "disabled",
-        false
-      ]
-    },
-    {
-      "id": 32,
-      "type": "VHS_VideoCombine",
-      "pos": {
-        "0": 1439,
-        "1": 122
-      },
-      "size": [
-        563.3333740234375,
-        686.2222493489583
-      ],
-      "flags": {},
-      "order": 8,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "images",
-          "type": "IMAGE",
-          "link": 60,
-          "slot_index": 0
-        },
-        {
-          "name": "audio",
-          "type": "VHS_AUDIO",
-          "link": null
-        },
-        {
-          "name": "meta_batch",
-          "type": "VHS_BatchManager",
-          "link": null
-        },
-        {
-          "name": "vae",
-          "type": "VAE",
-          "link": null
-        }
-      ],
-      "outputs": [
-        {
-          "name": "Filenames",
-          "type": "VHS_FILENAMES",
-          "links": null,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "VHS_VideoCombine"
-      },
-      "widgets_values": {
-        "frame_rate": 8,
-        "loop_count": 0,
-        "filename_prefix": "CogVideo2B_long",
-        "format": "video/h264-mp4",
-        "pix_fmt": "yuv420p",
-        "crf": 19,
-        "save_metadata": true,
-        "pingpong": false,
-        "save_output": false,
-        "videopreview": {
-          "hidden": false,
-          "paused": false,
-          "params": {
-            "filename": "CogVideo2B_long_00005.mp4",
-            "subfolder": "",
-            "type": "temp",
-            "format": "video/h264-mp4",
-            "frame_rate": 8
-          }
-        }
-      }
-    },
-    {
-      "id": 34,
-      "type": "CogVideoContextOptions",
-      "pos": {
-        "0": 1053,
-        "1": -84
-      },
-      "size": {
-        "0": 315,
-        "1": 154
-      },
-      "flags": {},
-      "order": 2,
-      "mode": 0,
-      "inputs": [],
-      "outputs": [
-        {
-          "name": "context_options",
-          "type": "COGCONTEXT",
-          "links": [
-            61
-          ],
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoContextOptions"
-      },
-      "widgets_values": [
-        "uniform_standard",
-        52,
-        4,
-        8,
-        true
-      ]
-    },
-    {
-      "id": 22,
-      "type": "CogVideoSampler",
-      "pos": {
-        "0": 1041,
-        "1": 342
-      },
-      "size": {
-        "0": 315,
-        "1": 382
-      },
-      "flags": {},
-      "order": 5,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "pipeline",
-          "type": "COGVIDEOPIPE",
-          "link": 36
-        },
-        {
-          "name": "positive",
-          "type": "CONDITIONING",
-          "link": 55,
-          "slot_index": 1
-        },
-        {
-          "name": "negative",
-          "type": "CONDITIONING",
-          "link": 57
-        },
-        {
-          "name": "samples",
-          "type": "LATENT",
-          "link": null
-        },
-        {
-          "name": "image_cond_latents",
-          "type": "LATENT",
-          "link": null
-        },
-        {
-          "name": "context_options",
-          "type": "COGCONTEXT",
-          "link": 61
-        }
-      ],
-      "outputs": [
-        {
-          "name": "cogvideo_pipe",
-          "type": "COGVIDEOPIPE",
-          "links": [
-            37
-          ],
-          "shape": 3
-        },
-        {
-          "name": "samples",
-          "type": "LATENT",
-          "links": [
-            38
-          ],
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "CogVideoSampler"
-      },
-      "widgets_values": [
-        480,
-        720,
-        104,
-        32,
-        6,
-        42,
-        "fixed",
-        "CogVideoXDDIM",
-        1
-      ]
-    }
-  ],
-  "links": [
-    [
-      36,
-      1,
-      0,
-      22,
-      0,
-      "COGVIDEOPIPE"
-    ],
-    [
-      37,
-      22,
-      0,
-      11,
-      0,
-      "COGVIDEOPIPE"
-    ],
-    [
-      38,
-      22,
-      1,
-      11,
-      1,
-      "LATENT"
-    ],
-    [
-      54,
-      20,
-      0,
-      30,
-      0,
-      "CLIP"
-    ],
-    [
-      55,
-      30,
-      0,
-      22,
-      1,
-      "CONDITIONING"
-    ],
-    [
-      56,
-      20,
-      0,
-      31,
-      0,
-      "CLIP"
-    ],
-    [
-      57,
-      31,
-      0,
-      22,
-      2,
-      "CONDITIONING"
-    ],
-    [
-      59,
-      11,
-      0,
-      33,
-      0,
-      "IMAGE"
-    ],
-    [
-      60,
-      33,
-      0,
-      32,
-      0,
-      "IMAGE"
-    ],
-    [
-      61,
-      34,
-      0,
-      22,
-      5,
-      "COGCONTEXT"
-    ]
-  ],
-  "groups": [],
-  "config": {},
-  "extra": {
-    "ds": {
-      "scale": 0.8390545288825444,
-      "offset": [
-        -14.198557467892236,
-        144.90015432747748
-      ]
-    }
-  },
-  "version": 0.4
-}

latent_preview.py

@@ -1,79 +0,0 @@
-import io
-
-import torch
-from PIL import Image
-import struct
-import numpy as np
-from comfy.cli_args import args, LatentPreviewMethod
-from comfy.taesd.taesd import TAESD
-import comfy.model_management
-import folder_paths
-import comfy.utils
-import logging
-
-MAX_PREVIEW_RESOLUTION = args.preview_size
-
-def preview_to_image(latent_image):
-        latents_ubyte = (((latent_image + 1.0) / 2.0).clamp(0, 1)  # change scale from -1..1 to 0..1
-                            .mul(0xFF)  # to 0..255
-                            ).to(device="cpu", dtype=torch.uint8, non_blocking=comfy.model_management.device_supports_non_blocking(latent_image.device))
-
-        return Image.fromarray(latents_ubyte.numpy())
-
-class LatentPreviewer:
-    def decode_latent_to_preview(self, x0):
-        pass
-
-    def decode_latent_to_preview_image(self, preview_format, x0):
-        preview_image = self.decode_latent_to_preview(x0)
-        return ("GIF", preview_image, MAX_PREVIEW_RESOLUTION)
-
-class Latent2RGBPreviewer(LatentPreviewer):
-    def __init__(self):
-        latent_rgb_factors = [[0.11945946736445662, 0.09919175788574555, -0.004832707433877734], [-0.0011977028264356232, 0.05496505130267682, 0.021321622433638193], [-0.014088548986590666, -0.008701477861945644, -0.020991313281459367], [0.03063921972519621, 0.12186477097625073, 0.0139593690235148], [0.0927403067854673, 0.030293187650929136, 0.05083134241694003], [0.0379112441305742, 0.04935199882777209, 0.058562766246777774], [0.017749911959153715, 0.008839453404921545, 0.036005638019226294], [0.10610119248526109, 0.02339855688237826, 0.057154257614084596], [0.1273639464837117, -0.010959856130713416, 0.043268631260428896], [-0.01873510946881321, 0.08220930648486932, 0.10613256772247093], [0.008429116376722327, 0.07623856561000408, 0.09295712117576727], [0.12938137079617007, 0.12360403483892413, 0.04478930933220116], [0.04565908794779364, 0.041064156741596365, -0.017695041535528512], [0.00019003240570281826, -0.013965147883381978, 0.05329669529635849], [0.08082391586738358, 0.11548306825496074, -0.021464170006615893], [-0.01517932393230994, -0.0057985555313003236, 0.07216646476618871]]
-
-        self.latent_rgb_factors = torch.tensor(latent_rgb_factors, device="cpu").transpose(0, 1)
-        self.latent_rgb_factors_bias = None
-        # if latent_rgb_factors_bias is not None:
-        #     self.latent_rgb_factors_bias = torch.tensor(latent_rgb_factors_bias, device="cpu")
-
-    def decode_latent_to_preview(self, x0):
-        self.latent_rgb_factors = self.latent_rgb_factors.to(dtype=x0.dtype, device=x0.device)
-        if self.latent_rgb_factors_bias is not None:
-            self.latent_rgb_factors_bias = self.latent_rgb_factors_bias.to(dtype=x0.dtype, device=x0.device)
-
-        latent_image = torch.nn.functional.linear(x0[0].permute(1, 2, 0), self.latent_rgb_factors,
-                                                    bias=self.latent_rgb_factors_bias)
-        return preview_to_image(latent_image)
-
-
-def get_previewer():
-    previewer = None
-    method = args.preview_method
-    if method != LatentPreviewMethod.NoPreviews:
-        # TODO previewer method
-
-        if method == LatentPreviewMethod.Auto:
-            method = LatentPreviewMethod.Latent2RGB
-
-        if previewer is None:
-            previewer = Latent2RGBPreviewer()
-    return previewer
-
-def prepare_callback(model, steps, x0_output_dict=None):
-    preview_format = "JPEG"
-    if preview_format not in ["JPEG", "PNG"]:
-        preview_format = "JPEG"
-
-    previewer = get_previewer()
-
-    pbar = comfy.utils.ProgressBar(steps)
-    def callback(step, x0, x, total_steps):
-        if x0_output_dict is not None:
-            x0_output_dict["x0"] = x0
-        preview_bytes = None
-        if previewer:
-            preview_bytes = previewer.decode_latent_to_preview_image(preview_format, x0)
-        pbar.update_absolute(step + 1, total_steps, preview_bytes)
-    return callback
-

lora_utils.py

@@ -406,20 +406,23 @@ def merge_lora(transformer, lora_path, multiplier, device='cpu', dtype=torch.flo
                 else:
                     temp_name = layer_infos.pop(0)
 
-        weight_up = elems['lora_up.weight'].to(dtype)
-        weight_down = elems['lora_down.weight'].to(dtype)
+        weight_up = elems['lora_up.weight'].to(dtype).to(device)
+        weight_down = elems['lora_down.weight'].to(dtype).to(device)
         if 'alpha' in elems.keys():
             alpha = elems['alpha'].item() / weight_up.shape[1]
         else:
             alpha = 1.0
 
         curr_layer.weight.data = curr_layer.weight.data.to(device)
-        if len(weight_up.shape) == 4:
-            curr_layer.weight.data += multiplier * alpha * torch.mm(weight_up.squeeze(3).squeeze(2),
-                                                                    weight_down.squeeze(3).squeeze(2)).unsqueeze(
-                2).unsqueeze(3)
-        else:
-            curr_layer.weight.data += multiplier * alpha * torch.mm(weight_up, weight_down)
+        try:
+            if len(weight_up.shape) == 4:
+                curr_layer.weight.data += multiplier * alpha * torch.mm(weight_up.squeeze(3).squeeze(2),
+                                                                        weight_down.squeeze(3).squeeze(2)).unsqueeze(
+                    2).unsqueeze(3)
+            else:
+                curr_layer.weight.data += multiplier * alpha * torch.mm(weight_up, weight_down)
+        except:
+            print(f"Could not apply LoRA weight in layer {layer}")
 
     return transformer
 

model_loading.py

@@ -43,11 +43,8 @@ from .custom_cogvideox_transformer_3d import CogVideoXTransformer3DModel
 from .pipeline_cogvideox import CogVideoXPipeline
 from contextlib import nullcontext
 
-from .cogvideox_fun.transformer_3d import CogVideoXTransformer3DModel as CogVideoXTransformer3DModelFun
-from .cogvideox_fun.autoencoder_magvit import AutoencoderKLCogVideoX as AutoencoderKLCogVideoXFun
-
-from .cogvideox_fun.pipeline_cogvideox_inpaint import CogVideoX_Fun_Pipeline_Inpaint
-from .cogvideox_fun.pipeline_cogvideox_control import CogVideoX_Fun_Pipeline_Control
+from accelerate import init_empty_weights
+from accelerate.utils import set_module_tensor_to_device
 
 from .utils import remove_specific_blocks, log
 from comfy.utils import load_torch_file
@@ -73,6 +70,7 @@ class CogVideoLoraSelect:
     RETURN_NAMES = ("lora", )
     FUNCTION = "getlorapath"
     CATEGORY = "CogVideoWrapper"
+    DESCRIPTION = "Select a LoRA model from ComfyUI/models/CogVideo/loras"
 
     def getlorapath(self, lora, strength, prev_lora=None, fuse_lora=False):
         cog_loras_list = []
@@ -89,6 +87,43 @@ class CogVideoLoraSelect:
         cog_loras_list.append(cog_lora)
         print(cog_loras_list)
         return (cog_loras_list,)
+
+class CogVideoLoraSelectComfy:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+               "lora": (folder_paths.get_filename_list("loras"), 
+                {"tooltip": "LORA models are expected to be in ComfyUI/models/loras with .safetensors extension"}),
+                "strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.0001, "tooltip": "LORA strength, set to 0.0 to unmerge the LORA"}),
+            },
+            "optional": {
+                "prev_lora":("COGLORA", {"default": None, "tooltip": "For loading multiple LoRAs"}),
+                "fuse_lora": ("BOOLEAN", {"default": False, "tooltip": "Fuse the LoRA weights into the transformer"}),
+            }
+        }
+
+    RETURN_TYPES = ("COGLORA",)
+    RETURN_NAMES = ("lora", )
+    FUNCTION = "getlorapath"
+    CATEGORY = "CogVideoWrapper"
+    DESCRIPTION = "Select a LoRA model from ComfyUI/models/loras"
+
+    def getlorapath(self, lora, strength, prev_lora=None, fuse_lora=False):
+        cog_loras_list = []
+
+        cog_lora = {
+            "path": folder_paths.get_full_path("loras", lora),
+            "strength": strength,
+            "name": lora.split(".")[0],
+            "fuse_lora": fuse_lora
+        }
+        if prev_lora is not None:
+            cog_loras_list.extend(prev_lora)
+            
+        cog_loras_list.append(cog_lora)
+        print(cog_loras_list)
+        return (cog_loras_list,)
     
 #region DownloadAndLoadCogVideoModel
 class DownloadAndLoadCogVideoModel:
@@ -111,6 +146,8 @@ class DownloadAndLoadCogVideoModel:
                         "alibaba-pai/CogVideoX-Fun-V1.1-5b-InP",
                         "alibaba-pai/CogVideoX-Fun-V1.1-2b-Pose",
                         "alibaba-pai/CogVideoX-Fun-V1.1-5b-Pose",
+                        "alibaba-pai/CogVideoX-Fun-V1.1-5b-Control",
+                        "alibaba-pai/CogVideoX-Fun-V1.5-5b-InP",
                         "feizhengcong/CogvideoX-Interpolation",
                         "NimVideo/cogvideox-2b-img2vid"
                     ],
@@ -121,27 +158,51 @@ class DownloadAndLoadCogVideoModel:
                 "precision": (["fp16", "fp32", "bf16"],
                     {"default": "bf16", "tooltip": "official recommendation is that 2b model should be fp16, 5b model should be bf16"}
                 ),
-                "fp8_transformer": (['disabled', 'enabled', 'fastmode', 'torchao_fp8dq', "torchao_fp8dqrow", "torchao_int8dq", "torchao_fp6"], {"default": 'disabled', "tooltip": "enabled casts the transformer to torch.float8_e4m3fn, fastmode is only for latest nvidia GPUs and requires torch 2.4.0 and cu124 minimum"}),
-                "compile": (["disabled","onediff","torch"], {"tooltip": "compile the model for faster inference, these are advanced options only available on Linux, see readme for more info"}),
+                "quantization": (['disabled', 'fp8_e4m3fn', 'fp8_e4m3fn_fastmode', 'torchao_fp8dq', "torchao_fp8dqrow", "torchao_int8dq", "torchao_fp6"], {"default": 'disabled', "tooltip": "enabled casts the transformer to torch.float8_e4m3fn, fastmode is only for latest nvidia GPUs and requires torch 2.4.0 and cu124 minimum"}),
                 "enable_sequential_cpu_offload": ("BOOLEAN", {"default": False, "tooltip": "significantly reducing memory usage and slows down the inference"}),
                 "block_edit": ("TRANSFORMERBLOCKS", {"default": None}),
                 "lora": ("COGLORA", {"default": None}),
                 "compile_args":("COMPILEARGS", ),
-                "attention_mode": (["sdpa", "sageattn", "fused_sdpa", "fused_sageattn"], {"default": "sdpa"}),
+                "attention_mode": ([
+                    "sdpa",
+                    "fused_sdpa",
+                    "sageattn",
+                    "fused_sageattn", 
+                    "sageattn_qk_int8_pv_fp8_cuda",
+                    "sageattn_qk_int8_pv_fp16_cuda",
+                    "sageattn_qk_int8_pv_fp16_triton",
+                    "fused_sageattn_qk_int8_pv_fp8_cuda",
+                    "fused_sageattn_qk_int8_pv_fp16_cuda",
+                    "fused_sageattn_qk_int8_pv_fp16_triton",
+                    "comfy"
+                    ], {"default": "sdpa"}),
                 "load_device": (["main_device", "offload_device"], {"default": "main_device"}),
             }
         }
 
-    RETURN_TYPES = ("COGVIDEOPIPE",)
-    RETURN_NAMES = ("cogvideo_pipe", )
+    RETURN_TYPES = ("COGVIDEOMODEL", "VAE",)
+    RETURN_NAMES = ("model", "vae", )
     FUNCTION = "loadmodel"
     CATEGORY = "CogVideoWrapper"
     DESCRIPTION = "Downloads and loads the selected CogVideo model from Huggingface to 'ComfyUI/models/CogVideo'"
 
-    def loadmodel(self, model, precision, fp8_transformer="disabled", compile="disabled", 
+    def loadmodel(self, model, precision, quantization="disabled", compile="disabled", 
                   enable_sequential_cpu_offload=False, block_edit=None, lora=None, compile_args=None, 
                   attention_mode="sdpa", load_device="main_device"):
         
+        transformer = None
+
+        if "sage" in attention_mode:
+            try:
+                from sageattention import sageattn
+            except Exception as e:
+                raise ValueError(f"Can't import SageAttention: {str(e)}")
+            if "qk_int8" in attention_mode:
+                try:
+                    from sageattention import sageattn_qk_int8_pv_fp16_cuda
+                except Exception as e:
+                    raise ValueError(f"Can't import SageAttention 2.0.0: {str(e)}")
+        
         if precision == "fp16" and "1.5" in model:
             raise ValueError("1.5 models do not currently work in fp16")
 
@@ -155,7 +216,7 @@ class DownloadAndLoadCogVideoModel:
         download_path = folder_paths.get_folder_paths("CogVideo")[0]
         
         if "Fun" in model:
-            if not "1.1" in model:
+            if "1.1" not in model and "1.5" not in model:
                 repo_id = "kijai/CogVideoX-Fun-pruned"
                 if "2b" in model:
                     base_path = os.path.join(folder_paths.models_dir, "CogVideoX_Fun", "CogVideoX-Fun-2b-InP") # location of the official model
@@ -165,7 +226,7 @@ class DownloadAndLoadCogVideoModel:
                     base_path = os.path.join(folder_paths.models_dir, "CogVideoX_Fun", "CogVideoX-Fun-5b-InP") # location of the official model
                     if not os.path.exists(base_path):
                         base_path = os.path.join(download_path, "CogVideoX-Fun-5b-InP")
-            elif "1.1" in model:
+            else:
                 repo_id = model
                 base_path = os.path.join(folder_paths.models_dir, "CogVideoX_Fun", (model.split("/")[-1])) # location of the official model
                 if not os.path.exists(base_path):
@@ -215,15 +276,10 @@ class DownloadAndLoadCogVideoModel:
                 local_dir_use_symlinks=False,
             )
 
-        #transformer
-        if "Fun" in model:
-            transformer = CogVideoXTransformer3DModelFun.from_pretrained(base_path, subfolder=subfolder)
-        else:
-            transformer = CogVideoXTransformer3DModel.from_pretrained(base_path, subfolder=subfolder)
-        
+        transformer = CogVideoXTransformer3DModel.from_pretrained(base_path, subfolder=subfolder, attention_mode=attention_mode)
         transformer = transformer.to(dtype).to(transformer_load_device)
 
-        if "1.5" in model:
+        if "1.5" in model and not "fun" in model:
             transformer.config.sample_height = 300
             transformer.config.sample_width = 300
 
@@ -235,54 +291,73 @@ class DownloadAndLoadCogVideoModel:
         scheduler = CogVideoXDDIMScheduler.from_config(scheduler_config)     
 
         # VAE
-        if "Fun" in model:
-            vae = AutoencoderKLCogVideoXFun.from_pretrained(base_path, subfolder="vae").to(dtype).to(offload_device)
-            if "Pose" in model:
-                pipe = CogVideoX_Fun_Pipeline_Control(vae, transformer, scheduler)
-            else:
-                pipe = CogVideoX_Fun_Pipeline_Inpaint(vae, transformer, scheduler)
-        else:
-            vae = AutoencoderKLCogVideoX.from_pretrained(base_path, subfolder="vae").to(dtype).to(offload_device)
-            pipe = CogVideoXPipeline(vae, transformer, scheduler)
-            if "cogvideox-2b-img2vid" in model:
-                pipe.input_with_padding = False
+        vae = AutoencoderKLCogVideoX.from_pretrained(base_path, subfolder="vae").to(dtype).to(offload_device)
+
+        #pipeline
+        pipe = CogVideoXPipeline(
+            transformer, 
+            scheduler, 
+            dtype=dtype, 
+            is_fun_inpaint="fun" in model.lower() and not ("pose" in model.lower() or "control" in model.lower())
+            )
+        if "cogvideox-2b-img2vid" in model:
+            pipe.input_with_padding = False
         
         #LoRAs
         if lora is not None:
-            from .lora_utils import merge_lora#, load_lora_into_transformer
-            if "fun" in model.lower():
-                for l in lora:
-                    log.info(f"Merging LoRA weights from {l['path']} with strength {l['strength']}")
-                    transformer = merge_lora(transformer, l["path"], l["strength"])
-            else:
-                adapter_list = []
-                adapter_weights = []
-                for l in lora:
-                    fuse = True if l["fuse_lora"] else False
-                    lora_sd = load_torch_file(l["path"])             
-                    for key, val in lora_sd.items():
-                        if "lora_B" in key:
-                            lora_rank = val.shape[1]
-                            break
+            dimensionx_loras = ["orbit", "dimensionx"] # for now dimensionx loras need scaling
+            dimensionx_lora = False
+            adapter_list = []
+            adapter_weights = []
+            for l in lora:
+                if any(item in l["path"].lower() for item in dimensionx_loras):
+                    dimensionx_lora = True
+                fuse = True if l["fuse_lora"] else False
+                lora_sd = load_torch_file(l["path"])
+                lora_rank = None            
+                for key, val in lora_sd.items():
+                    if "lora_B" in key:
+                        lora_rank = val.shape[1]
+                        break
+                if lora_rank is not None:
                     log.info(f"Merging rank {lora_rank} LoRA weights from {l['path']} with strength {l['strength']}")
                     adapter_name = l['path'].split("/")[-1].split(".")[0]
                     adapter_weight = l['strength']
                     pipe.load_lora_weights(l['path'], weight_name=l['path'].split("/")[-1], lora_rank=lora_rank, adapter_name=adapter_name)
                     
-                    #transformer = load_lora_into_transformer(lora, transformer)
                     adapter_list.append(adapter_name)
                     adapter_weights.append(adapter_weight)
-                for l in lora:
-                    pipe.set_adapters(adapter_list, adapter_weights=adapter_weights)
+                else:
+                    try: #Fun trainer LoRAs are loaded differently
+                        from .lora_utils import merge_lora
+                        log.info(f"Merging LoRA weights from {l['path']} with strength {l['strength']}")
+                        pipe.transformer = merge_lora(pipe.transformer, l["path"], l["strength"], device=transformer_load_device, state_dict=lora_sd)
+                    except:
+                        raise ValueError(f"Can't recognize LoRA {l['path']}")
+                del lora_sd
+                mm.soft_empty_cache()
+            if adapter_list:
+                pipe.set_adapters(adapter_list, adapter_weights=adapter_weights)
                 if fuse:
                     lora_scale = 1
-                    dimension_loras = ["orbit", "dimensionx"] # for now dimensionx loras need scaling
-                    if any(item in lora[-1]["path"].lower() for item in dimension_loras):
+                    if dimensionx_lora:
                         lora_scale = lora_scale / lora_rank
                     pipe.fuse_lora(lora_scale=lora_scale, components=["transformer"])
+                    pipe.delete_adapters(adapter_list)
+           
+
+        if "fused" in attention_mode:
+            from diffusers.models.attention import Attention
+            pipe.transformer.fuse_qkv_projections = True
+            for module in pipe.transformer.modules():
+                if isinstance(module, Attention):
+                    module.fuse_projections(fuse=True)
+
+        if compile_args is not None:
+            pipe.transformer.to(memory_format=torch.channels_last)
 
         #fp8
-        if fp8_transformer == "enabled" or fp8_transformer == "fastmode":
+        if quantization == "fp8_e4m3fn" or quantization == "fp8_e4m3fn_fastmode":
             params_to_keep = {"patch_embed", "lora", "pos_embedding", "time_embedding", "norm_k", "norm_q", "to_k.bias", "to_q.bias", "to_v.bias"}
             if "1.5" in model:
                     params_to_keep.update({"norm1.linear.weight", "ofs_embedding", "norm_final", "norm_out", "proj_out"}) 
@@ -290,13 +365,20 @@ class DownloadAndLoadCogVideoModel:
                 if not any(keyword in name for keyword in params_to_keep):
                     param.data = param.data.to(torch.float8_e4m3fn)
         
-            if fp8_transformer == "fastmode":
+            if quantization == "fp8_e4m3fn_fastmode":
                 from .fp8_optimization import convert_fp8_linear
                 if "1.5" in model:
                     params_to_keep.update({"ff"}) #otherwise NaNs
                 convert_fp8_linear(pipe.transformer, dtype, params_to_keep=params_to_keep)
+        
+        # compilation
+        if compile_args is not None:
+            torch._dynamo.config.cache_size_limit = compile_args["dynamo_cache_size_limit"]
+            for i, block in enumerate(pipe.transformer.transformer_blocks):
+                if "CogVideoXBlock" in str(block):
+                    pipe.transformer.transformer_blocks[i] = torch.compile(block, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
 
-        elif "torchao" in fp8_transformer:
+        if "torchao" in quantization:
             try:
                 from torchao.quantization import (
                 quantize_,
@@ -313,14 +395,14 @@ class DownloadAndLoadCogVideoModel:
                     return isinstance(module, nn.Linear)
                 return False
             
-            if "fp6" in fp8_transformer: #slower for some reason on 4090
+            if "fp6" in quantization: #slower for some reason on 4090
                 quant_func = fpx_weight_only(3, 2)
-            elif "fp8dq" in fp8_transformer: #very fast on 4090 when compiled
+            elif "fp8dq" in quantization: #very fast on 4090 when compiled
                 quant_func = float8_dynamic_activation_float8_weight()
-            elif 'fp8dqrow' in fp8_transformer:
+            elif 'fp8dqrow' in quantization:
                 from torchao.quantization.quant_api import PerRow
                 quant_func = float8_dynamic_activation_float8_weight(granularity=PerRow())
-            elif 'int8dq' in fp8_transformer:
+            elif 'int8dq' in quantization:
                 quant_func = int8_dynamic_activation_int8_weight()
         
             for i, block in enumerate(pipe.transformer.transformer_blocks):
@@ -365,45 +447,24 @@ class DownloadAndLoadCogVideoModel:
         #         (3): Dropout(p=0.0, inplace=False)
         #         )
         #     )
-        #     )       
+        #     )               
+       
+        # if compile == "onediff":
+        #     from onediffx import compile_pipe
+        #     os.environ['NEXFORT_FX_FORCE_TRITON_SDPA'] = '1'
             
-        # compilation
-        if compile == "torch":
-            #pipe.transformer.to(memory_format=torch.channels_last)
-            if compile_args is not None:
-                torch._dynamo.config.cache_size_limit = compile_args["dynamo_cache_size_limit"]
-                for i, block in enumerate(pipe.transformer.transformer_blocks):
-                    if "CogVideoXBlock" in str(block):
-                        pipe.transformer.transformer_blocks[i] = torch.compile(block, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
-            else:
-                for i, block in enumerate(pipe.transformer.transformer_blocks):
-                    if "CogVideoXBlock" in str(block):
-                        pipe.transformer.transformer_blocks[i] = torch.compile(block, fullgraph=False, dynamic=False, backend="inductor")
-        
-        transformer.attention_mode = attention_mode
-
-        if "fused" in attention_mode:
-            from diffusers.models.attention import Attention
-            transformer.fuse_qkv_projections = True
-            for module in transformer.modules():
-                if isinstance(module, Attention):
-                    module.fuse_projections(fuse=True)
-
-        elif compile == "onediff":
-            from onediffx import compile_pipe
-            os.environ['NEXFORT_FX_FORCE_TRITON_SDPA'] = '1'
-            
-            pipe = compile_pipe(
-            pipe,
-            backend="nexfort",
-            options= {"mode": "max-optimize:max-autotune:max-autotune", "memory_format": "channels_last", "options": {"inductor.optimize_linear_epilogue": False, "triton.fuse_attention_allow_fp16_reduction": False}},
-            ignores=["vae"],
-            fuse_qkv_projections= False,
-            )          
+        #     pipe = compile_pipe(
+        #     pipe,
+        #     backend="nexfort",
+        #     options= {"mode": "max-optimize:max-autotune:max-autotune", "memory_format": "channels_last", "options": {"inductor.optimize_linear_epilogue": False, "triton.fuse_attention_allow_fp16_reduction": False}},
+        #     ignores=["vae"],
+        #     fuse_qkv_projections= False,
+        #     )          
         
         pipeline = {
             "pipe": pipe,
             "dtype": dtype,
+            "quantization": quantization,
             "base_path": base_path,
             "onediff": True if compile == "onediff" else False,
             "cpu_offloading": enable_sequential_cpu_offload,
@@ -412,7 +473,7 @@ class DownloadAndLoadCogVideoModel:
             "model_name": model,
         }
 
-        return (pipeline,)
+        return (pipeline, vae)
 #region GGUF
 class DownloadAndLoadCogVideoGGUFModel:
     @classmethod
@@ -438,19 +499,24 @@ class DownloadAndLoadCogVideoGGUFModel:
             },
             "optional": {
                 "block_edit": ("TRANSFORMERBLOCKS", {"default": None}),
-                #"lora": ("COGLORA", {"default": None}),
-                "compile": (["disabled","torch"], {"tooltip": "compile the model for faster inference, these are advanced options only available on Linux, see readme for more info"}),
+                #"compile_args":("COMPILEARGS", ),
                 "attention_mode": (["sdpa", "sageattn"], {"default": "sdpa"}),
             }
         }
 
-    RETURN_TYPES = ("COGVIDEOPIPE",)
-    RETURN_NAMES = ("cogvideo_pipe", )
+    RETURN_TYPES = ("COGVIDEOMODEL", "VAE",)
+    RETURN_NAMES = ("model", "vae",)
     FUNCTION = "loadmodel"
     CATEGORY = "CogVideoWrapper"
 
     def loadmodel(self, model, vae_precision, fp8_fastmode, load_device, enable_sequential_cpu_offload, 
-                  block_edit=None, compile="disabled", attention_mode="sdpa"):
+                  block_edit=None, compile_args=None, attention_mode="sdpa"):
+        
+        if "sage" in attention_mode:
+            try:
+                from sageattention import sageattn
+            except Exception as e:
+                raise ValueError(f"Can't import SageAttention: {str(e)}")
 
         device = mm.get_torch_device()
         offload_device = mm.unet_offload_device()
@@ -486,7 +552,6 @@ class DownloadAndLoadCogVideoGGUFModel:
         with open(transformer_path) as f:
             transformer_config = json.load(f)
 
-        
 
         from . import mz_gguf_loader
         import importlib
@@ -498,7 +563,6 @@ class DownloadAndLoadCogVideoGGUFModel:
                     transformer_config["in_channels"] = 32
                 else:
                     transformer_config["in_channels"] = 33
-                transformer = CogVideoXTransformer3DModelFun.from_config(transformer_config)
             elif "I2V" in model or "Interpolation" in model:
                 transformer_config["in_channels"] = 32
                 if "1_5" in model:
@@ -508,11 +572,11 @@ class DownloadAndLoadCogVideoGGUFModel:
                     transformer_config["patch_bias"] = False
                     transformer_config["sample_height"] = 300
                     transformer_config["sample_width"] = 300
-                transformer = CogVideoXTransformer3DModel.from_config(transformer_config)
             else:
                 transformer_config["in_channels"] = 16
-                transformer = CogVideoXTransformer3DModel.from_config(transformer_config)
 
+            transformer = CogVideoXTransformer3DModel.from_config(transformer_config, attention_mode=attention_mode)
+            cast_dtype = vae_dtype
             params_to_keep = {"patch_embed", "pos_embedding", "time_embedding"}
             if "2b" in model:
                 cast_dtype = torch.float16
@@ -539,10 +603,6 @@ class DownloadAndLoadCogVideoGGUFModel:
            from .fp8_optimization import convert_fp8_linear
            convert_fp8_linear(transformer, vae_dtype, params_to_keep=params_to_keep)
 
-        if compile == "torch":
-            # compilation
-            for i, block in enumerate(transformer.transformer_blocks):
-                    transformer.transformer_blocks[i] = torch.compile(block, fullgraph=False, dynamic=False, backend="inductor")
         with open(scheduler_path) as f:
             scheduler_config = json.load(f)
         
@@ -564,63 +624,34 @@ class DownloadAndLoadCogVideoGGUFModel:
         with open(os.path.join(script_directory, 'configs', 'vae_config.json')) as f:
             vae_config = json.load(f)
         
+        #VAE
         vae_sd = load_torch_file(vae_path)
-        if "fun" in model:
-            vae = AutoencoderKLCogVideoXFun.from_config(vae_config).to(vae_dtype).to(offload_device)
-            vae.load_state_dict(vae_sd)
-            if "Pose" in model:
-                pipe = CogVideoX_Fun_Pipeline_Control(vae, transformer, scheduler)
-            else:
-                pipe = CogVideoX_Fun_Pipeline_Inpaint(vae, transformer, scheduler)
-        else:
-            vae = AutoencoderKLCogVideoX.from_config(vae_config).to(vae_dtype).to(offload_device)
-            vae.load_state_dict(vae_sd)
-            pipe = CogVideoXPipeline(vae, transformer, scheduler)
+        vae = AutoencoderKLCogVideoX.from_config(vae_config).to(vae_dtype).to(offload_device)
+        vae.load_state_dict(vae_sd)
+        del vae_sd
+        pipe = CogVideoXPipeline(
+            transformer, 
+            scheduler, 
+            dtype=vae_dtype,
+            is_fun_inpaint="fun" in model.lower() and not ("pose" in model.lower() or "control" in model.lower())
+            )
 
         if enable_sequential_cpu_offload:
             pipe.enable_sequential_cpu_offload()
 
         sd = load_torch_file(gguf_path)
-
-        # #LoRAs
-        # if lora is not None:        
-        #     if "fun" in model.lower():
-        #         raise NotImplementedError("LoRA with GGUF is not supported for Fun models")
-        #         from .lora_utils import merge_lora#, load_lora_into_transformer
-        #         #for l in lora:
-        #         #    log.info(f"Merging LoRA weights from {l['path']} with strength {l['strength']}")
-        #         #    pipe.transformer = merge_lora(pipe.transformer, l["path"], l["strength"])
-        #     else:
-        #         adapter_list = []
-        #         adapter_weights = []
-        #         for l in lora:
-        #             lora_sd = load_torch_file(l["path"])             
-        #             for key, val in lora_sd.items():
-        #                 if "lora_B" in key:
-        #                     lora_rank = val.shape[1]
-        #                     break
-        #             log.info(f"Loading rank {lora_rank} LoRA weights from {l['path']} with strength {l['strength']}")
-        #             adapter_name = l['path'].split("/")[-1].split(".")[0]
-        #             adapter_weight = l['strength']
-        #             pipe.load_lora_weights(l['path'], weight_name=l['path'].split("/")[-1], lora_rank=lora_rank, adapter_name=adapter_name)
-                    
-        #             #transformer = load_lora_into_transformer(lora, transformer)
-        #             adapter_list.append(adapter_name)
-        #             adapter_weights.append(adapter_weight)
-        #         for l in lora:
-        #             pipe.set_adapters(adapter_list, adapter_weights=adapter_weights)
-        #         #pipe.fuse_lora(lora_scale=1 / lora_rank, components=["transformer"])
-        
         pipe.transformer = mz_gguf_loader.quantize_load_state_dict(pipe.transformer, sd, device="cpu")
+        del sd
+
         if load_device == "offload_device":
             pipe.transformer.to(offload_device)
         else:
             pipe.transformer.to(device)
 
-
         pipeline = {
             "pipe": pipe,
             "dtype": vae_dtype,
+            "quantization": "GGUF",
             "base_path": model,
             "onediff": False,
             "cpu_offloading": enable_sequential_cpu_offload,
@@ -629,21 +660,283 @@ class DownloadAndLoadCogVideoGGUFModel:
             "manual_offloading": True,
         }
 
+        return (pipeline, vae)
+
+#region ModelLoader
+class CogVideoXModelLoader:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "model": (folder_paths.get_filename_list("diffusion_models"), {"tooltip": "These models are loaded from the 'ComfyUI/models/diffusion_models' -folder",}),
+            
+            "base_precision": (["fp16", "fp32", "bf16"], {"default": "bf16"}),
+            "quantization": (['disabled', 'fp8_e4m3fn', 'fp8_e4m3fn_fast', 'torchao_fp8dq', "torchao_fp8dqrow", "torchao_int8dq", "torchao_fp6"], {"default": 'disabled', "tooltip": "optional quantization method"}),
+            "load_device": (["main_device", "offload_device"], {"default": "main_device"}),
+            "enable_sequential_cpu_offload": ("BOOLEAN", {"default": False, "tooltip": "significantly reducing memory usage and slows down the inference"}),
+            },
+            "optional": {
+                "block_edit": ("TRANSFORMERBLOCKS", {"default": None}),
+                "lora": ("COGLORA", {"default": None}),
+                "compile_args":("COMPILEARGS", ),
+                "attention_mode": ([
+                    "sdpa",
+                    "fused_sdpa",
+                    "sageattn",
+                    "fused_sageattn", 
+                    "sageattn_qk_int8_pv_fp8_cuda",
+                    "sageattn_qk_int8_pv_fp16_cuda",
+                    "sageattn_qk_int8_pv_fp16_triton",
+                    "fused_sageattn_qk_int8_pv_fp8_cuda",
+                    "fused_sageattn_qk_int8_pv_fp16_cuda",
+                    "fused_sageattn_qk_int8_pv_fp16_triton",
+                    "comfy"
+                    ], {"default": "sdpa"}),
+            }
+        }
+
+    RETURN_TYPES = ("COGVIDEOMODEL",)
+    RETURN_NAMES = ("model", )
+    FUNCTION = "loadmodel"
+    CATEGORY = "CogVideoWrapper"
+
+    def loadmodel(self, model, base_precision, load_device, enable_sequential_cpu_offload, 
+                  block_edit=None, compile_args=None, lora=None, attention_mode="sdpa", quantization="disabled"):
+        transformer = None
+        if "sage" in attention_mode:
+            try:
+                from sageattention import sageattn
+            except Exception as e:
+                raise ValueError(f"Can't import SageAttention: {str(e)}")
+
+        device = mm.get_torch_device()
+        offload_device = mm.unet_offload_device()
+        manual_offloading = True
+        transformer_load_device = device if load_device == "main_device" else offload_device
+        mm.soft_empty_cache()
+
+        base_dtype = {"fp8_e4m3fn": torch.float8_e4m3fn, "fp8_e4m3fn_fast": torch.float8_e4m3fn, "bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[base_precision]
+
+        model_path = folder_paths.get_full_path_or_raise("diffusion_models", model)
+        sd = load_torch_file(model_path, device=transformer_load_device)
+
+        model_type = ""
+        if sd["patch_embed.proj.weight"].shape == (3072, 33, 2, 2):
+            model_type = "fun_5b"
+        elif sd["patch_embed.proj.weight"].shape == (3072, 16, 2, 2):
+            model_type = "5b"
+        elif sd["patch_embed.proj.weight"].shape == (3072, 128):
+            model_type = "5b_1_5"
+        elif sd["patch_embed.proj.weight"].shape == (3072, 256):
+            model_type = "5b_I2V_1_5"
+        elif sd["patch_embed.proj.weight"].shape == (1920, 33, 2, 2):
+            model_type = "fun_2b"
+        elif sd["patch_embed.proj.weight"].shape == (1920, 32, 2, 2):
+            model_type = "cogvideox-2b-img2vid"
+        elif sd["patch_embed.proj.weight"].shape == (1920, 16, 2, 2):
+            model_type = "2b"
+        elif sd["patch_embed.proj.weight"].shape == (3072, 32, 2, 2):
+            if "pos_embedding" in sd:
+                model_type = "fun_5b_pose"
+            else:
+                model_type = "I2V_5b"
+        else:
+            raise Exception("Selected model is not recognized")
+        log.info(f"Detected CogVideoX model type: {model_type}")
+
+        if "5b" in model_type:
+            scheduler_config_path = os.path.join(script_directory, 'configs', 'scheduler_config_5b.json')
+            transformer_config_path = os.path.join(script_directory, 'configs', 'transformer_config_5b.json')
+        elif "2b" in model_type:
+            scheduler_config_path = os.path.join(script_directory, 'configs', 'scheduler_config_2b.json')
+            transformer_config_path = os.path.join(script_directory, 'configs', 'transformer_config_2b.json')
+    
+        with open(transformer_config_path) as f:
+            transformer_config = json.load(f)
+
+            if model_type in ["I2V", "I2V_5b", "fun_5b_pose", "5b_I2V_1_5", "cogvideox-2b-img2vid"]:
+                transformer_config["in_channels"] = 32
+                if "1_5" in model_type:
+                    transformer_config["ofs_embed_dim"] = 512
+            elif "fun" in model_type:
+                transformer_config["in_channels"] = 33
+            else:
+                transformer_config["in_channels"] = 16
+            if "1_5" in model_type:
+                    transformer_config["use_learned_positional_embeddings"] = False
+                    transformer_config["patch_size_t"] = 2
+                    transformer_config["patch_bias"] = False
+                    transformer_config["sample_height"] = 300
+                    transformer_config["sample_width"] = 300
+
+        with init_empty_weights():
+            transformer = CogVideoXTransformer3DModel.from_config(transformer_config, attention_mode=attention_mode)
+
+        #load weights
+        #params_to_keep = {}
+        log.info("Using accelerate to load and assign model weights to device...")
+        
+        for name, param in transformer.named_parameters():
+            #dtype_to_use = base_dtype if any(keyword in name for keyword in params_to_keep) else dtype
+            set_module_tensor_to_device(transformer, name, device=transformer_load_device, dtype=base_dtype, value=sd[name])
+        del sd
+        # TODO fix for transformer model patch_embed.pos_embedding dtype 
+        #   or at add line  ComfyUI-CogVideoXWrapper/embeddings.py:129 code
+        #   pos_embedding = pos_embedding.to(embeds.device, dtype=embeds.dtype)
+        transformer = transformer.to(base_dtype).to(transformer_load_device)
+
+        #scheduler
+        with open(scheduler_config_path) as f:
+            scheduler_config = json.load(f)
+        scheduler = CogVideoXDDIMScheduler.from_config(scheduler_config, subfolder="scheduler")
+
+        if block_edit is not None:
+            transformer = remove_specific_blocks(transformer, block_edit)
+
+        if "fused" in attention_mode:
+            from diffusers.models.attention import Attention
+            transformer.fuse_qkv_projections = True
+            for module in transformer.modules():
+                if isinstance(module, Attention):
+                    module.fuse_projections(fuse=True)
+        transformer.attention_mode = attention_mode
+
+        pipe = CogVideoXPipeline(
+            transformer, 
+            scheduler, 
+            dtype=base_dtype, 
+            is_fun_inpaint="fun" in model.lower() and not ("pose" in model.lower() or "control" in model.lower())
+            )
+        if "cogvideox-2b-img2vid" == model_type:
+            pipe.input_with_padding = False
+        if enable_sequential_cpu_offload:
+            pipe.enable_sequential_cpu_offload()
+
+        #LoRAs
+        if lora is not None:
+            dimensionx_loras = ["orbit", "dimensionx"] # for now dimensionx loras need scaling
+            dimensionx_lora = False
+            adapter_list = []
+            adapter_weights = []
+            for l in lora:
+                if any(item in l["path"].lower() for item in dimensionx_loras):
+                    dimensionx_lora = True
+                fuse = True if l["fuse_lora"] else False
+                lora_sd = load_torch_file(l["path"])
+                lora_rank = None            
+                for key, val in lora_sd.items():
+                    if "lora_B" in key:
+                        lora_rank = val.shape[1]
+                        break
+                if lora_rank is not None:
+                    log.info(f"Merging rank {lora_rank} LoRA weights from {l['path']} with strength {l['strength']}")
+                    adapter_name = l['path'].split("/")[-1].split(".")[0]
+                    adapter_weight = l['strength']
+                    pipe.load_lora_weights(l['path'], weight_name=l['path'].split("/")[-1], lora_rank=lora_rank, adapter_name=adapter_name)
+                    
+                    adapter_list.append(adapter_name)
+                    adapter_weights.append(adapter_weight)
+                else:
+                    try: #Fun trainer LoRAs are loaded differently
+                        from .lora_utils import merge_lora
+                        log.info(f"Merging LoRA weights from {l['path']} with strength {l['strength']}")
+                        pipe.transformer = merge_lora(pipe.transformer, l["path"], l["strength"], device=transformer_load_device, state_dict=lora_sd)
+                    except:
+                        raise ValueError(f"Can't recognize LoRA {l['path']}")
+            if adapter_list:
+                pipe.set_adapters(adapter_list, adapter_weights=adapter_weights)
+                if fuse:
+                    lora_scale = 1
+                    if dimensionx_lora:
+                        lora_scale = lora_scale / lora_rank
+                    pipe.fuse_lora(lora_scale=lora_scale, components=["transformer"])
+
+        if compile_args is not None:
+            pipe.transformer.to(memory_format=torch.channels_last)
+
+        #quantization
+        if quantization == "fp8_e4m3fn" or quantization == "fp8_e4m3fn_fast":
+            params_to_keep = {"patch_embed", "lora", "pos_embedding", "time_embedding", "norm_k", "norm_q", "to_k.bias", "to_q.bias", "to_v.bias"}
+            if "1.5" in model:
+                    params_to_keep.update({"norm1.linear.weight", "ofs_embedding", "norm_final", "norm_out", "proj_out"}) 
+            for name, param in pipe.transformer.named_parameters():
+                if not any(keyword in name for keyword in params_to_keep):
+                    param.data = param.data.to(torch.float8_e4m3fn)
+        
+            if quantization == "fp8_e4m3fn_fast":
+                from .fp8_optimization import convert_fp8_linear
+                if "1.5" in model:
+                    params_to_keep.update({"ff"}) #otherwise NaNs
+                convert_fp8_linear(pipe.transformer, base_dtype, params_to_keep=params_to_keep)
+        
+        #compile
+        if compile_args is not None:
+            torch._dynamo.config.cache_size_limit = compile_args["dynamo_cache_size_limit"]
+            for i, block in enumerate(pipe.transformer.transformer_blocks):
+                if "CogVideoXBlock" in str(block):
+                    pipe.transformer.transformer_blocks[i] = torch.compile(block, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
+
+        if "torchao" in quantization:
+            try:
+                from torchao.quantization import (
+                quantize_,
+                fpx_weight_only,
+                float8_dynamic_activation_float8_weight,
+                int8_dynamic_activation_int8_weight
+            )
+            except:
+                raise ImportError("torchao is not installed, please install torchao to use fp8dq")
+
+            def filter_fn(module: nn.Module, fqn: str) -> bool:
+                target_submodules = {'attn1', 'ff'} # avoid norm layers, 1.5 at least won't work with quantized norm1 #todo: test other models
+                if any(sub in fqn for sub in target_submodules):
+                    return isinstance(module, nn.Linear)
+                return False
+            
+            if "fp6" in quantization: #slower for some reason on 4090
+                quant_func = fpx_weight_only(3, 2)
+            elif "fp8dq" in quantization: #very fast on 4090 when compiled
+                quant_func = float8_dynamic_activation_float8_weight()
+            elif 'fp8dqrow' in quantization:
+                from torchao.quantization.quant_api import PerRow
+                quant_func = float8_dynamic_activation_float8_weight(granularity=PerRow())
+            elif 'int8dq' in quantization:
+                quant_func = int8_dynamic_activation_int8_weight()
+        
+            for i, block in enumerate(pipe.transformer.transformer_blocks):
+                if "CogVideoXBlock" in str(block):
+                    quantize_(block, quant_func, filter_fn=filter_fn)
+            
+            manual_offloading = False # to disable manual .to(device) calls
+            log.info(f"Quantized transformer blocks to {quantization}")
+
+        pipeline = {
+            "pipe": pipe,
+            "dtype": base_dtype,
+            "quantization": quantization,
+            "base_path": model,
+            "onediff": False,
+            "cpu_offloading": enable_sequential_cpu_offload,
+            "scheduler_config": scheduler_config,
+            "model_name": model,
+            "manual_offloading": manual_offloading,
+        }
         return (pipeline,)
     
-#revion VAE
+#region VAE
 
 class CogVideoXVAELoader:
     @classmethod
     def INPUT_TYPES(s):
         return {
             "required": {
-                "model_name": (folder_paths.get_filename_list("vae"), {"tooltip": "The name of the checkpoint (vae) to load."}),
+                "model_name": (folder_paths.get_filename_list("vae"), {"tooltip": "These models are loaded from 'ComfyUI/models/vae'"}),
             },
             "optional": {
                 "precision": (["fp16", "fp32", "bf16"],
                     {"default": "bf16"}
                 ),
+                "compile_args":("COMPILEARGS", ),
             }
         }
 
@@ -653,7 +946,7 @@ class CogVideoXVAELoader:
     CATEGORY = "CogVideoWrapper"
     DESCRIPTION = "Loads CogVideoX VAE model from 'ComfyUI/models/vae'"
 
-    def loadmodel(self, model_name, precision):
+    def loadmodel(self, model_name, precision, compile_args=None):
         device = mm.get_torch_device()
         offload_device = mm.unet_offload_device()
 
@@ -665,6 +958,10 @@ class CogVideoXVAELoader:
 
         vae = AutoencoderKLCogVideoX.from_config(vae_config).to(dtype).to(offload_device)
         vae.load_state_dict(vae_sd)
+        #compile
+        if compile_args is not None:
+            torch._dynamo.config.cache_size_limit = compile_args["dynamo_cache_size_limit"]
+            vae = torch.compile(vae, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
 
         return (vae,)
     
@@ -677,6 +974,7 @@ class DownloadAndLoadToraModel:
                 "model": (
                     [
                         "kijai/CogVideoX-5b-Tora",
+                        "kijai/CogVideoX-5b-Tora-I2V",
                     ],
                 ),
             },
@@ -706,14 +1004,17 @@ class DownloadAndLoadToraModel:
             pass
 
         download_path = os.path.join(folder_paths.models_dir, 'CogVideo', "CogVideoX-5b-Tora")
-        fuser_path = os.path.join(download_path, "fuser", "fuser.safetensors")
+
+        
+        fuser_model = "fuser.safetensors" if not "I2V" in model else "fuser_I2V.safetensors"
+        fuser_path = os.path.join(download_path, "fuser", fuser_model)
         if not os.path.exists(fuser_path):
             log.info(f"Downloading Fuser model to: {fuser_path}")
             from huggingface_hub import snapshot_download
 
             snapshot_download(
                 repo_id=model,
-                allow_patterns=["*fuser.safetensors*"],
+                allow_patterns=[fuser_model],
                 local_dir=download_path,
                 local_dir_use_symlinks=False,
             )
@@ -735,14 +1036,15 @@ class DownloadAndLoadToraModel:
                     param.data = param.data.to(torch.bfloat16).to(device)
         del fuser_sd
 
-        traj_extractor_path = os.path.join(download_path, "traj_extractor", "traj_extractor.safetensors")
+        traj_extractor_model = "traj_extractor.safetensors" if not "I2V" in model else "traj_extractor_I2V.safetensors"
+        traj_extractor_path = os.path.join(download_path, "traj_extractor", traj_extractor_model)
         if not os.path.exists(traj_extractor_path):
             log.info(f"Downloading trajectory extractor model to: {traj_extractor_path}")
             from huggingface_hub import snapshot_download
 
             snapshot_download(
                 repo_id="kijai/CogVideoX-5b-Tora",
-                allow_patterns=["*traj_extractor.safetensors*"],
+                allow_patterns=[traj_extractor_model],
                 local_dir=download_path,
                 local_dir_use_symlinks=False,
             )
@@ -829,6 +1131,8 @@ NODE_CLASS_MAPPINGS = {
     "DownloadAndLoadToraModel": DownloadAndLoadToraModel,
     "CogVideoLoraSelect": CogVideoLoraSelect,
     "CogVideoXVAELoader": CogVideoXVAELoader,
+    "CogVideoXModelLoader": CogVideoXModelLoader,
+    "CogVideoLoraSelectComfy": CogVideoLoraSelectComfy
 }
 NODE_DISPLAY_NAME_MAPPINGS = {
     "DownloadAndLoadCogVideoModel": "(Down)load CogVideo Model",
@@ -837,4 +1141,6 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "DownloadAndLoadToraModel": "(Down)load Tora Model",
     "CogVideoLoraSelect": "CogVideo LoraSelect",
     "CogVideoXVAELoader": "CogVideoX VAE Loader",
+    "CogVideoXModelLoader": "CogVideoX Model Loader",
+    "CogVideoLoraSelectComfy": "CogVideo LoraSelect Comfy"
     }

pipeline_cogvideox.py

@@ -17,20 +17,19 @@ import inspect
 from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import torch
-import torch.nn.functional as F
 import math
 
-from diffusers.models import AutoencoderKLCogVideoX
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
 from diffusers.utils import logging
 from diffusers.utils.torch_utils import randn_tensor
-from diffusers.video_processor import VideoProcessor
+
 #from diffusers.models.embeddings import get_3d_rotary_pos_embed
 from diffusers.loaders import CogVideoXLoraLoaderMixin
 
 from .embeddings import get_3d_rotary_pos_embed
 from .custom_cogvideox_transformer_3d import CogVideoXTransformer3DModel
+from .enhance_a_video.globals import enable_enhance, disable_enhance, set_enhance_weight
 
 from comfy.utils import ProgressBar
 
@@ -112,6 +111,34 @@ def retrieve_timesteps(
         timesteps = scheduler.timesteps
     return timesteps, num_inference_steps
 
+class CogVideoXLatentFormat():
+    latent_channels = 16
+    latent_dimensions = 3
+    scale_factor = 0.7
+    taesd_decoder_name = None
+   
+    latent_rgb_factors = [[0.03197404301362048, 0.04091260743347359, 0.0015679806301828524], 
+                          [0.005517101026578029, 0.0052348639043457755, -0.005613441650464035], 
+                          [0.0012485338264583965, -0.016096744206117782, 0.025023940031635054], 
+                          [0.01760126794276171, 0.0036818415416642893, -0.0006019202528157255], 
+                          [0.000444954842288864, 0.006102128982092191, 0.0008457999272962447], 
+                          [-0.010531904354560697, -0.0032275501924977175, -0.00886595780267917], 
+                          [-0.0001454543946122991, 0.010199210750845965, -0.00012702234832386188], 
+                          [0.02078497279904325, -0.001669617778939972, 0.006712703698951264], 
+                          [0.005529571599763264, 0.009733929789086743, 0.001887302765339838], 
+                          [0.012138415094654218, 0.024684961927224837, 0.037211249767461915], 
+                          [0.0010364484570000384, 0.01983636315929172, 0.009864602025627755], 
+                          [0.006802862648143341, -0.0010509255113510681, -0.007026003345126021], 
+                          [0.0003532208468418043, 0.005351971582801936, -0.01845912126717106], 
+                          [-0.009045079994694397, -0.01127941143183089, 0.0042294057970470806], 
+                          [0.002548289972720752, 0.025224244654428216, -0.0006086130121693347], 
+                          [-0.011135669222532816, 0.0018181308593668505, 0.02794541485349922]]
+    latent_rgb_factors_bias = [ -0.023, 0.0, -0.017]
+
+class CogVideoXModelPlaceholder():
+    def __init__(self):
+        self.latent_format = CogVideoXLatentFormat
+
 class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
     r"""
     Pipeline for text-to-video generation using CogVideoX.
@@ -120,15 +147,6 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
     library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
 
     Args:
-        vae ([`AutoencoderKL`]):
-            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
-        text_encoder ([`T5EncoderModel`]):
-            Frozen text-encoder. CogVideoX uses
-            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
-            [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
-        tokenizer (`T5Tokenizer`):
-            Tokenizer of class
-            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
         transformer ([`CogVideoXTransformer3DModel`]):
             A text conditioned `CogVideoXTransformer3DModel` to denoise the encoded video latents.
         scheduler ([`SchedulerMixin`]):
@@ -140,31 +158,25 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
 
     def __init__(
         self,
-        vae: AutoencoderKLCogVideoX,
         transformer: CogVideoXTransformer3DModel,
         scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
-        original_mask = None,
+        dtype: torch.dtype = torch.bfloat16,
+        is_fun_inpaint: bool = False,
     ):
         super().__init__()
 
-        self.register_modules(
-            vae=vae, transformer=transformer, scheduler=scheduler
-        )
-        self.vae_scale_factor_spatial = (
-            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
-        )
-        self.vae_scale_factor_temporal = (
-            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
-        )        
-        self.original_mask = original_mask
-        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
-        self.video_processor.config.do_resize = False
+        self.register_modules(transformer=transformer, scheduler=scheduler)
+        self.vae_scale_factor_spatial = 8
+        self.vae_scale_factor_temporal = 4
+        self.vae_latent_channels = 16
+        self.vae_dtype = dtype
+        self.is_fun_inpaint = is_fun_inpaint
 
         self.input_with_padding = True
 
 
     def prepare_latents(
-        self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, timesteps, denoise_strength,
+        self, batch_size, num_channels_latents, num_frames, height, width, device, generator, timesteps, denoise_strength,
          num_inference_steps, latents=None, freenoise=True, context_size=None, context_overlap=None
     ):
         shape = (
@@ -174,14 +186,10 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
             height // self.vae_scale_factor_spatial,
             width // self.vae_scale_factor_spatial,
         )
-        if isinstance(generator, list) and len(generator) != batch_size:
-            raise ValueError(
-                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
-                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
-            )
-        noise = randn_tensor(shape, generator=generator, device=torch.device("cpu"), dtype=self.vae.dtype)
+        
+        noise = randn_tensor(shape, generator=generator, device=torch.device("cpu"), dtype=self.vae_dtype)
         if freenoise:
-            print("Applying FreeNoise")
+            logger.info("Applying FreeNoise")
             # code and comments from AnimateDiff-Evolved by Kosinkadink (https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved)
             video_length = num_frames // 4
             delta = context_size - context_overlap
@@ -216,25 +224,27 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                 noise[:, place_idx:place_idx + delta, :, :, :] = noise[:, list_idx, :, :, :]
         if latents is None:
             latents = noise.to(device)
-        else:
+        elif denoise_strength < 1.0:
             latents = latents.to(device)
             timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, denoise_strength, device)
             latent_timestep = timesteps[:1]
             
-            noise = randn_tensor(shape, generator=generator, device=device, dtype=self.vae.dtype)
             frames_needed = noise.shape[1]
             current_frames = latents.shape[1]
             
             if frames_needed > current_frames:
-                repeat_factor = frames_needed // current_frames
+                repeat_factor = frames_needed - current_frames
                 additional_frame = torch.randn((latents.size(0), repeat_factor, latents.size(2), latents.size(3), latents.size(4)), dtype=latents.dtype, device=latents.device)
-                latents = torch.cat((latents, additional_frame), dim=1)
+                latents = torch.cat((additional_frame, latents), dim=1)
+                self.additional_frames = repeat_factor
             elif frames_needed < current_frames:
                 latents = latents[:, :frames_needed, :, :, :]
 
-            latents = self.scheduler.add_noise(latents, noise, latent_timestep)
+            latents = self.scheduler.add_noise(latents, noise.to(device), latent_timestep)
+        else:
+            latents = latents.to(device)
         latents = latents * self.scheduler.init_noise_sigma # scale the initial noise by the standard deviation required by the scheduler
-        return latents, timesteps, noise
+        return latents, timesteps
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
     def prepare_extra_step_kwargs(self, generator, eta):
@@ -355,10 +365,10 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
         guidance_scale: float = 6,
         denoise_strength: float = 1.0,
         sigmas: Optional[List[float]] = None,
-        num_videos_per_prompt: int = 1,
         eta: float = 0.0,
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
         latents: Optional[torch.Tensor] = None,
+        fun_mask: Optional[torch.Tensor] = None,
         image_cond_latents: Optional[torch.Tensor] = None,
         prompt_embeds: Optional[torch.Tensor] = None,
         negative_prompt_embeds: Optional[torch.Tensor] = None,
@@ -370,6 +380,9 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
         freenoise: Optional[bool] = True,
         controlnet: Optional[dict] = None,
         tora: Optional[dict] = None,
+        image_cond_start_percent: float = 0.0,
+        image_cond_end_percent: float = 1.0,
+        feta_args: Optional[dict] = None,
         
     ):
         """
@@ -398,8 +411,6 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                 Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                 usually at the expense of lower image quality.
-            num_videos_per_prompt (`int`, *optional*, defaults to 1):
-                The number of videos to generate per prompt.
             generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                 One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                 to make generation deterministic.
@@ -418,7 +429,6 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
         
         height = height or self.transformer.config.sample_size * self.vae_scale_factor_spatial
         width = width or self.transformer.config.sample_size * self.vae_scale_factor_spatial
-        num_videos_per_prompt = 1
 
         self.num_frames = num_frames
 
@@ -443,7 +453,7 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
 
         if do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
-        prompt_embeds = prompt_embeds.to(self.vae.dtype)
+        prompt_embeds = prompt_embeds.to(self.vae_dtype)
 
         # 4. Prepare timesteps
         if sigmas is None:
@@ -453,7 +463,7 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
         self._num_timesteps = len(timesteps)
 
         # 5. Prepare latents.
-        latent_channels = self.vae.config.latent_channels
+        latent_channels = self.vae_latent_channels
         latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
 
         # For CogVideoX 1.5, the latent frames should be padded to make it divisible by patch_size_t
@@ -469,18 +479,12 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
             self.additional_frames = patch_size_t - latent_frames % patch_size_t
             num_frames += self.additional_frames * self.vae_scale_factor_temporal
 
-
-        if self.original_mask is not None:
-            image_latents = latents
-            original_image_latents = image_latents
-
-        latents, timesteps, noise = self.prepare_latents(
-            batch_size * num_videos_per_prompt,
+        latents, timesteps = self.prepare_latents(
+            batch_size,
             latent_channels,
             num_frames,
             height,
             width,
-            self.vae.dtype,
             device,
             generator,
             timesteps,
@@ -491,70 +495,73 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
             context_overlap=context_overlap,
             freenoise=freenoise,
         )
-        latents = latents.to(self.vae.dtype)
+        latents = latents.to(self.vae_dtype)
+
+        if self.is_fun_inpaint and fun_mask is None: # For FUN inpaint vid2vid, we need to mask all the latents
+            fun_mask = torch.zeros_like(latents[:, :, :1, :, :], device=latents.device, dtype=latents.dtype)
+            fun_masked_video_latents = torch.zeros_like(latents, device=latents.device, dtype=latents.dtype)
 
         # 5.5.
         if image_cond_latents is not None:
-            if image_cond_latents.shape[1] > 1:
+            image_cond_frame_count = image_cond_latents.size(1)
+            patch_size_t = self.transformer.config.patch_size_t
+            if image_cond_frame_count == 2:
                 logger.info("More than one image conditioning frame received, interpolating")
                 padding_shape = (
-                batch_size,
-                (latents.shape[1] - 2),
-                self.vae.config.latent_channels,
-                height // self.vae_scale_factor_spatial,
-                width // self.vae_scale_factor_spatial,
+                    batch_size,
+                    (latents.shape[1] - 2),
+                    self.vae_latent_channels,
+                    height // self.vae_scale_factor_spatial,
+                    width // self.vae_scale_factor_spatial,
                 )
-                latent_padding = torch.zeros(padding_shape, device=device, dtype=self.vae.dtype)
+                latent_padding = torch.zeros(padding_shape, device=device, dtype=self.vae_dtype)
                 image_cond_latents = torch.cat([image_cond_latents[:, 0, :, :, :].unsqueeze(1), latent_padding, image_cond_latents[:, -1, :, :, :].unsqueeze(1)], dim=1)
-                if self.transformer.config.patch_size_t is not None:
-                        first_frame = image_cond_latents[:, : image_cond_latents.size(1) % self.transformer.config.patch_size_t, ...]
-                        image_cond_latents = torch.cat([first_frame, image_cond_latents], dim=1)
+                if patch_size_t:
+                    first_frame = image_cond_latents[:, : image_cond_latents.size(1) % patch_size_t, ...]
+                    image_cond_latents = torch.cat([first_frame, image_cond_latents], dim=1)
 
                 logger.info(f"image cond latents shape: {image_cond_latents.shape}")
-            else:
+            elif image_cond_frame_count == 1:
                 logger.info("Only one image conditioning frame received, img2vid")
                 if self.input_with_padding:
                     padding_shape = (
                         batch_size,
                         (latents.shape[1] - 1),
-                        self.vae.config.latent_channels,
+                        self.vae_latent_channels,
                         height // self.vae_scale_factor_spatial,
                         width // self.vae_scale_factor_spatial,
                     )
-                    latent_padding = torch.zeros(padding_shape, device=device, dtype=self.vae.dtype)
+                    latent_padding = torch.zeros(padding_shape, device=device, dtype=self.vae_dtype)
                     image_cond_latents = torch.cat([image_cond_latents, latent_padding], dim=1)
                     # Select the first frame along the second dimension
-                    if self.transformer.config.patch_size_t is not None:
-                        first_frame = image_cond_latents[:, : image_cond_latents.size(1) % self.transformer.config.patch_size_t, ...]
+                    if patch_size_t:
+                        first_frame = image_cond_latents[:, : image_cond_latents.size(1) % patch_size_t, ...]
                         image_cond_latents = torch.cat([first_frame, image_cond_latents], dim=1)
                 else:
                     image_cond_latents = image_cond_latents.repeat(1, latents.shape[1], 1, 1, 1)
+            else:
+                logger.info(f"Received {image_cond_latents.shape[1]} image conditioning frames")
+                if fun_mask is not None and patch_size_t:
+                    logger.info(f"1.5 model received {fun_mask.shape[1]} masks")
+                    first_frame = image_cond_latents[:, : image_cond_frame_count % patch_size_t, ...]
+                    image_cond_latents = torch.cat([first_frame, image_cond_latents], dim=1)
+                    fun_mask_first_frame = fun_mask[:, : image_cond_frame_count % patch_size_t, ...]
+                    fun_mask = torch.cat([fun_mask_first_frame, fun_mask], dim=1)
+                    fun_mask[:, 1:, ...] = 0
+            image_cond_latents = image_cond_latents.to(self.vae_dtype)
+
         # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
-
-        # masks
-        if self.original_mask is not None:
-            mask = self.original_mask.to(device)
-            logger.info(f"self.original_mask: {self.original_mask.shape}")
-            
-            mask = F.interpolate(self.original_mask.unsqueeze(1), size=(latents.shape[-2], latents.shape[-1]), mode='bilinear', align_corners=False)
-            if mask.shape[0] != latents.shape[1]:
-                mask = mask.unsqueeze(1).repeat(1, latents.shape[1], 16, 1, 1)
-            else:
-                mask = mask.unsqueeze(0).repeat(1, 1, 16, 1, 1)
-            logger.info(f"latents: {latents.shape}")
-            logger.info(f"mask: {mask.shape}")
-
       
         num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
 
         # 7. context schedule
         if context_schedule is not None:
-            if image_cond_latents is not None:
-                raise NotImplementedError("Context schedule not currently supported with image conditioning")
+            # if image_cond_latents is not None:
+            #     raise NotImplementedError("Context schedule not currently supported with image conditioning")
             logger.info(f"Context schedule enabled: {context_frames} frames, {context_stride} stride, {context_overlap} overlap")
             use_context_schedule = True
-            from .cogvideox_fun.context import get_context_scheduler
+            from .context import get_context_scheduler
             context = get_context_scheduler(context_schedule)
             #todo ofs embeds?
 
@@ -598,7 +605,7 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
         else:
             controlnet_states = None
             control_weights= None
-
+        # 9. Tora
         if tora is not None:
             trajectory_length = tora["video_flow_features"].shape[1]
             logger.info(f"Tora trajectory length: {trajectory_length}")
@@ -606,36 +613,58 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
             #    raise ValueError(f"Tora trajectory length {trajectory_length} does not match inpaint_latents count {latents.shape[2]}")
             for module in self.transformer.fuser_list:
                 for param in module.parameters():
-                    param.data = param.data.to(device)
+                    param.data = param.data.to(self.vae_dtype).to(device)
 
         logger.info(f"Sampling {num_frames} frames in {latent_frames} latent frames at {width}x{height} with {num_inference_steps} inference steps")
 
-        from .latent_preview import prepare_callback
-        callback = prepare_callback(self.transformer, num_inference_steps)
+        if feta_args is not None:
+            set_enhance_weight(feta_args["weight"])
+            feta_start_percent = feta_args["start_percent"]
+            feta_end_percent = feta_args["end_percent"]
+            enable_enhance()
+        else:
+            disable_enhance()
+
+        # reset TeaCache
+        if hasattr(self.transformer, 'accumulated_rel_l1_distance'):
+            delattr(self.transformer, 'accumulated_rel_l1_distance')
+        self.transformer.teacache_counter = 0
+
+        # 11. Denoising loop
+        #from .latent_preview import prepare_callback
+        #callback = prepare_callback(self.transformer, num_inference_steps)
+        from latent_preview import prepare_callback
+        self.model = CogVideoXModelPlaceholder()
+        self.load_device = device
+        callback = prepare_callback(self, num_inference_steps)
 
-        # 9. Denoising loop
         comfy_pbar = ProgressBar(len(timesteps))
         with self.progress_bar(total=len(timesteps)) as progress_bar:    
             old_pred_original_sample = None # for DPM-solver++
             for i, t in enumerate(timesteps):
                 if self.interrupt:
                     continue
+
+                current_step_percentage = i / num_inference_steps
+
+                if feta_args is not None:
+                    if feta_start_percent <= current_step_percentage <= feta_end_percent:
+                        enable_enhance()
+                    else:
+                        disable_enhance()
                 # region context schedule sampling
                 if use_context_schedule:
                     latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                     latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
                     counter = torch.zeros_like(latent_model_input)
                     noise_pred = torch.zeros_like(latent_model_input)
-                    
+
                     if image_cond_latents is not None:
                         latent_image_input = torch.cat([image_cond_latents] * 2) if do_classifier_free_guidance else image_cond_latents
-                        latent_model_input = torch.cat([latent_model_input, latent_image_input], dim=2)
 
                     # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                     timestep = t.expand(latent_model_input.shape[0])
 
-                    current_step_percentage = i / num_inference_steps
-
                     # use same rotary embeddings for all context windows
                     image_rotary_emb = (
                             self._prepare_rotary_positional_embeddings(height, width, context_frames, device)
@@ -694,7 +723,14 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                             noise_pred = noise_pred.float()
                     else:
                         for c in context_queue:
+                            print("c:", c)
+
                             partial_latent_model_input = latent_model_input[:, c, :, :, :]
+                            if image_cond_latents is not None:
+                                partial_latent_image_input = latent_image_input[:, :len(c), :, :, :]
+                                partial_latent_model_input = torch.cat([partial_latent_model_input,partial_latent_image_input], dim=2)
+                            
+                            print(partial_latent_model_input.shape)
                             if (tora is not None and tora["start_percent"] <= current_step_percentage <= tora["end_percent"]):
                                 if do_classifier_free_guidance:
                                     partial_video_flow_features = tora["video_flow_features"][:, c, :, :, :].repeat(1, 2, 1, 1, 1).contiguous()
@@ -738,7 +774,13 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
 
                     if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                         progress_bar.update()
-                        comfy_pbar.update(1)
+                        if callback is not None: 
+                            alpha_prod_t = self.scheduler.alphas_cumprod[t]
+                            beta_prod_t = 1 - alpha_prod_t
+                            callback_tensor = (alpha_prod_t**0.5) * latent_model_input[0][:, :16, :, :] - (beta_prod_t**0.5) * noise_pred.detach()[0]
+                            callback(i, callback_tensor * 5, None, num_inference_steps)
+                        else:
+                            comfy_pbar.update(1)
 
                 # region sampling
                 else:
@@ -746,14 +788,26 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                     latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
                     if image_cond_latents is not None:
-                        latent_image_input = torch.cat([image_cond_latents] * 2) if do_classifier_free_guidance else image_cond_latents
-                        latent_model_input = torch.cat([latent_model_input, latent_image_input], dim=2)
+                        if not image_cond_start_percent <= current_step_percentage <= image_cond_end_percent:
+                            latent_image_input = torch.zeros_like(latent_model_input)
+                        else:
+                            latent_image_input = torch.cat([image_cond_latents] * 2) if do_classifier_free_guidance else image_cond_latents
+                        if fun_mask is not None: #for fun img2vid and interpolation
+                            fun_inpaint_mask = torch.cat([fun_mask] * 2) if do_classifier_free_guidance else fun_mask
+                            masks_input = torch.cat([fun_inpaint_mask, latent_image_input], dim=2)
+                            latent_model_input = torch.cat([latent_model_input, masks_input], dim=2)
+                        else:
+                            latent_model_input = torch.cat([latent_model_input, latent_image_input], dim=2)
+                    else: # for Fun inpaint vid2vid
+                        if fun_mask is not None:
+                            fun_inpaint_mask = torch.cat([fun_mask] * 2) if do_classifier_free_guidance else fun_mask
+                            fun_inpaint_masked_video_latents = torch.cat([fun_masked_video_latents] * 2) if do_classifier_free_guidance else fun_masked_video_latents
+                            fun_inpaint_latents = torch.cat([fun_inpaint_mask, fun_inpaint_masked_video_latents], dim=2).to(latents.dtype)
+                            latent_model_input = torch.cat([latent_model_input, fun_inpaint_latents], dim=2)
 
                     # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                     timestep = t.expand(latent_model_input.shape[0])
 
-                    current_step_percentage = i / num_inference_steps
-
                     if controlnet is not None:
                         controlnet_states = None
                         if (control_start <= current_step_percentage <= control_end):
@@ -767,10 +821,9 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                                 return_dict=False,
                             )[0]
                             if isinstance(controlnet_states, (tuple, list)):
-                                controlnet_states = [x.to(dtype=self.vae.dtype) for x in controlnet_states]
+                                controlnet_states = [x.to(dtype=self.vae_dtype) for x in controlnet_states]
                             else:
-                                controlnet_states = controlnet_states.to(dtype=self.vae.dtype)
-
+                                controlnet_states = controlnet_states.to(dtype=self.vae_dtype)
 
                     # predict noise model_output
                     noise_pred = self.transformer(
@@ -796,35 +849,26 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
 
                     # compute the previous noisy sample x_t -> x_t-1
                     if not isinstance(self.scheduler, CogVideoXDPMScheduler):
-                        latents = self.scheduler.step(noise_pred, t, latents.to(self.vae.dtype), **extra_step_kwargs, return_dict=False)[0]
+                        latents = self.scheduler.step(noise_pred, t, latents.to(self.vae_dtype), **extra_step_kwargs, return_dict=False)[0]
                     else:
                         latents, old_pred_original_sample = self.scheduler.step(
                             noise_pred,
                             old_pred_original_sample,
                             t,
                             timesteps[i - 1] if i > 0 else None,
-                            latents.to(self.vae.dtype),
+                            latents.to(self.vae_dtype),
                             **extra_step_kwargs,
                             return_dict=False,
                         )
                     latents = latents.to(prompt_embeds.dtype)
-                    # start diff diff
-                    if i < len(timesteps) - 1 and self.original_mask is not None:
-                        noise_timestep = timesteps[i + 1]
-                        image_latent = self.scheduler.add_noise(original_image_latents, noise, torch.tensor([noise_timestep])
-                        )
-                        mask = mask.to(latents)
-                        ts_from = timesteps[0]
-                        ts_to = timesteps[-1]
-                        threshold = (t - ts_to) / (ts_from - ts_to)
-                        mask = torch.where(mask >= threshold, mask, torch.zeros_like(mask))
-                        latents = image_latent * mask + latents * (1 - mask)
-                        # end diff diff
 
                     if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                         progress_bar.update()
-                        if callback is not None:
-                            callback(i, latents.detach()[-1], None, num_inference_steps)
+                        if callback is not None: 
+                            alpha_prod_t = self.scheduler.alphas_cumprod[t]
+                            beta_prod_t = 1 - alpha_prod_t
+                            callback_tensor = (alpha_prod_t**0.5) * latent_model_input[0][:, :16, :, :] - (beta_prod_t**0.5) * noise_pred.detach()[0]
+                            callback(i, callback_tensor * 5, None, num_inference_steps)
                         else:
                             comfy_pbar.update(1)
             

pyproject.toml

@@ -1,9 +1,9 @@
 [project]
 name = "comfyui-cogvideoxwrapper"
-description = "Diffusers wrapper for CogVideoX -models: [a/https://github.com/THUDM/CogVideo](https://github.com/THUDM/CogVideo)"
-version = "1.1.0"
+description = "Diffusers wrapper for CogVideoX -models: https://github.com/THUDM/CogVideo"
+version = "1.5.1"
 license = {file = "LICENSE"}
-dependencies = ["huggingface_hub", "diffusers>=0.30.1", "accelerate>=0.33.0"]
+dependencies = ["huggingface_hub", "diffusers>=0.31.0", "accelerate>=0.33.0"]
 
 [project.urls]
 Repository = "https://github.com/kijai/ComfyUI-CogVideoXWrapper"

readme.md

@@ -1,5 +1,47 @@
 # WORK IN PROGRESS
 
+Spreadsheet (WIP) of supported models and their supported features: https://docs.google.com/spreadsheets/d/16eA6mSL8XkTcu9fSWkPSHfRIqyAKJbR1O99xnuGdCKY/edit?usp=sharing
+
+## Update 9
+Added preliminary support for [Go-with-the-Flow](https://github.com/VGenAI-Netflix-Eyeline-Research/Go-with-the-Flow)
+
+This uses LoRA weights available here: https://huggingface.co/Eyeline-Research/Go-with-the-Flow/tree/main
+
+To create the input videos for the NoiseWarp process, I've added a node to KJNodes that works alongside my SplineEditor, and either [comfyui-inpaint-nodes](https://github.com/Acly/comfyui-inpaint-nodes) or just cv2 inpainting to create the cut and drag input videos.
+
+The workflows are in the example_workflows -folder.
+
+Quick video to showcase: First mask the subject, then use the cut and drag -workflow to create a video as seen here, then that video is used as input to the NoiseWarp node in the main workflow.
+
+https://github.com/user-attachments/assets/112706b0-a38b-4c3c-b779-deba0827af4f
+
+## BREAKING Update8
+
+This is big one, and unfortunately to do the necessary cleanup and refactoring this will break every old workflow as they are.
+I apologize for the inconvenience, if I don't do this now I'll keep making it worse until maintaining becomes too much of a chore, so from my pov there was no choice.
+
+*Please either use the new workflows or fix the nodes in your old ones before posting issue reports!*
+
+Old version will be kept in a legacy branch, but not maintained
+
+- Support CogVideoX 1.5 models
+- Major code cleanup (it was bad, still isn't great, wip)
+- Merge Fun -model functionality into main pipeline:
+    - All Fun specific nodes, besides image encode node for Fun -InP models are gone
+    - Main CogVideo Sampler works with Fun models
+    - DimensionX LoRAs now work with Fun models as well
+
+- Remove width/height from the sampler widgets and detect from input instead, this meanst text2vid now requires using empty latents
+- Separate VAE from the model, allow using fp32 VAE
+- Add ability to load some of the non-GGUF models as single files (only few available for now: https://huggingface.co/Kijai/CogVideoX-comfy)
+- Add some torchao quantizations as options
+- Add interpolation as option for the main encode node, old interpolation specific node is gone
+- torch.compile optimizations
+- Remove PAB in favor of FasterCache and cleaner code
+- other smaller things I forgot about at this point
+
+For Fun -model based workflows it's more drastic change, for others migrating generally means re-setting many of the nodes.
+
 ## Update7
 
 - Refactored the Fun version's sampler to accept any resolution, this should make it lot simpler to use with Tora. **BREAKS OLD WORKFLOWS**, old FunSampler nodes need to be remade.

requirements.txt

@@ -1,5 +1,5 @@
 huggingface_hub
-diffusers>=0.31.0
+diffusers>=0.33.1
 accelerate>=0.33.0
 einops
 peft