Add diffusers/pipeline_cogvideox_fun_inpaint.py

diffusers/pipeline_cogvideox_fun_inpaint.py

@@ -0,0 +1,1244 @@
+# 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 Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.models.embeddings import get_1d_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 einops import rearrange
+
+from transformers import T5EncoderModel, T5Tokenizer
+from cogvideox_transformer3d import CogVideoXTransformer3DModel
+from cogvideox_vae import AutoencoderKLCogVideoX
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        pass
+        ```
+"""
+
+# Copied from diffusers.models.embeddings.get_3d_rotary_pos_embed
+def get_3d_rotary_pos_embed(
+    embed_dim,
+    crops_coords,
+    grid_size,
+    temporal_size,
+    theta: int = 10000,
+    use_real: bool = True,
+    grid_type: str = "linspace",
+    max_size: Optional[Tuple[int, int]] = None,
+) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+    """
+    RoPE for video tokens with 3D structure.
+
+    Args:
+    embed_dim: (`int`):
+        The embedding dimension size, corresponding to hidden_size_head.
+    crops_coords (`Tuple[int]`):
+        The top-left and bottom-right coordinates of the crop.
+    grid_size (`Tuple[int]`):
+        The grid size of the spatial positional embedding (height, width).
+    temporal_size (`int`):
+        The size of the temporal dimension.
+    theta (`float`):
+        Scaling factor for frequency computation.
+    grid_type (`str`):
+        Whether to use "linspace" or "slice" to compute grids.
+
+    Returns:
+        `torch.Tensor`: positional embedding with shape `(temporal_size * grid_size[0] * grid_size[1], embed_dim/2)`.
+    """
+    if use_real is not True:
+        raise ValueError(" `use_real = False` is not currently supported for get_3d_rotary_pos_embed")
+
+    if grid_type == "linspace":
+        start, stop = crops_coords
+        grid_size_h, grid_size_w = grid_size
+        grid_h = np.linspace(start[0], stop[0], grid_size_h, endpoint=False, dtype=np.float32)
+        grid_w = np.linspace(start[1], stop[1], grid_size_w, endpoint=False, dtype=np.float32)
+        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":
+        max_h, max_w = max_size
+        grid_size_h, grid_size_w = grid_size
+        grid_h = np.arange(max_h, dtype=np.float32)
+        grid_w = np.arange(max_w, dtype=np.float32)
+        grid_t = np.arange(temporal_size, dtype=np.float32)
+    else:
+        raise ValueError("Invalid value passed for `grid_type`.")
+
+    # Compute dimensions for each axis
+    dim_t = embed_dim // 4
+    dim_h = embed_dim // 8 * 3
+    dim_w = embed_dim // 8 * 3
+
+    # Temporal frequencies
+    freqs_t = get_1d_rotary_pos_embed(dim_t, grid_t, use_real=True)
+    # Spatial frequencies for height and width
+    freqs_h = get_1d_rotary_pos_embed(dim_h, grid_h, use_real=True)
+    freqs_w = get_1d_rotary_pos_embed(dim_w, grid_w, use_real=True)
+
+    # BroadCast and concatenate temporal and spaial frequencie (height and width) into a 3d tensor
+    def combine_time_height_width(freqs_t, freqs_h, freqs_w):
+        freqs_t = freqs_t[:, None, None, :].expand(
+            -1, grid_size_h, grid_size_w, -1
+        )  # temporal_size, grid_size_h, grid_size_w, dim_t
+        freqs_h = freqs_h[None, :, None, :].expand(
+            temporal_size, -1, grid_size_w, -1
+        )  # temporal_size, grid_size_h, grid_size_2, dim_h
+        freqs_w = freqs_w[None, None, :, :].expand(
+            temporal_size, grid_size_h, -1, -1
+        )  # temporal_size, grid_size_h, grid_size_2, dim_w
+
+        freqs = torch.cat(
+            [freqs_t, freqs_h, freqs_w], dim=-1
+        )  # temporal_size, grid_size_h, grid_size_w, (dim_t + dim_h + dim_w)
+        freqs = freqs.view(
+            temporal_size * grid_size_h * grid_size_w, -1
+        )  # (temporal_size * grid_size_h * grid_size_w), (dim_t + dim_h + dim_w)
+        return freqs
+
+    t_cos, t_sin = freqs_t  # both t_cos and t_sin has shape: temporal_size, dim_t
+    h_cos, h_sin = freqs_h  # both h_cos and h_sin has shape: grid_size_h, dim_h
+    w_cos, w_sin = freqs_w  # both w_cos and w_sin has shape: grid_size_w, dim_w
+
+    if grid_type == "slice":
+        t_cos, t_sin = t_cos[:temporal_size], t_sin[:temporal_size]
+        h_cos, h_sin = h_cos[:grid_size_h], h_sin[:grid_size_h]
+        w_cos, w_sin = w_cos[:grid_size_w], w_sin[:grid_size_w]
+
+    cos = combine_time_height_width(t_cos, h_cos, w_cos)
+    sin = combine_time_height_width(t_sin, h_sin, w_sin)
+    return cos, sin
+
+
+# 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
+
+
+def resize_mask(mask, latent, process_first_frame_only=True):
+    latent_size = latent.size()
+    batch_size, channels, num_frames, height, width = mask.shape
+
+    if process_first_frame_only:
+        target_size = list(latent_size[2:])
+        target_size[0] = 1
+        first_frame_resized = F.interpolate(
+            mask[:, :, 0:1, :, :],
+            size=target_size,
+            mode='trilinear',
+            align_corners=False
+        )
+        
+        target_size = list(latent_size[2:])
+        target_size[0] = target_size[0] - 1
+        if target_size[0] != 0:
+            remaining_frames_resized = F.interpolate(
+                mask[:, :, 1:, :, :],
+                size=target_size,
+                mode='trilinear',
+                align_corners=False
+            )
+            resized_mask = torch.cat([first_frame_resized, remaining_frames_resized], dim=2)
+        else:
+            resized_mask = first_frame_resized
+    else:
+        target_size = list(latent_size[2:])
+        resized_mask = F.interpolate(
+            mask,
+            size=target_size,
+            mode='trilinear',
+            align_corners=False
+        )
+    return resized_mask
+
+
+def add_noise_to_reference_video(image, ratio=None):
+    if ratio is None:
+        sigma = torch.normal(mean=-3.0, std=0.5, size=(image.shape[0],)).to(image.device)
+        sigma = torch.exp(sigma).to(image.dtype)
+    else:
+        sigma = torch.ones((image.shape[0],)).to(image.device, image.dtype) * ratio
+    
+    image_noise = torch.randn_like(image) * sigma[:, None, None, None, None]
+    image_noise = torch.where(image==-1, torch.zeros_like(image), image_noise)
+    image = image + image_noise
+    return image
+
+
+@dataclass
+class CogVideoXFunPipelineOutput(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 CogVideoXFunInpaintPipeline(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.
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. CogVideoX_Fun 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`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
+    """
+
+    _optional_components = []
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+    ]
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKLCogVideoX,
+        transformer: CogVideoXTransformer3DModel,
+        scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, 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=False, do_convert_grayscale=True
+        )
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            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`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *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.Tensor`, *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.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(
+        self, 
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        video_length,
+        dtype,
+        device,
+        generator,
+        latents=None,
+        video=None,
+        timestep=None,
+        is_strength_max=True,
+        return_noise=False,
+        return_video_latents=False,
+    ):
+        shape = (
+            batch_size,
+            (video_length - 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."
+            )
+
+        if return_video_latents or (latents is None and not is_strength_max):
+            video = video.to(device=device, dtype=self.vae.dtype)
+            
+            bs = 1
+            new_video = []
+            for i in range(0, video.shape[0], bs):
+                video_bs = video[i : i + bs]
+                video_bs = self.vae.encode(video_bs)[0]
+                video_bs = video_bs.sample()
+                new_video.append(video_bs)
+            video = torch.cat(new_video, dim = 0)
+            video = video * self.vae.config.scaling_factor
+
+            video_latents = video.repeat(batch_size // video.shape[0], 1, 1, 1, 1)
+            video_latents = video_latents.to(device=device, dtype=dtype)
+            video_latents = rearrange(video_latents, "b c f h w -> b f c h w")
+
+        if latents is None:
+            noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+            # if strength is 1. then initialise the latents to noise, else initial to image + noise
+            latents = noise if is_strength_max else self.scheduler.add_noise(video_latents, noise, timestep)
+            # if pure noise then scale the initial latents by the  Scheduler's init sigma
+            latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
+        else:
+            noise = latents.to(device)
+            latents = noise * self.scheduler.init_noise_sigma
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        outputs = (latents,)
+
+        if return_noise:
+            outputs += (noise,)
+
+        if return_video_latents:
+            outputs += (video_latents,)
+
+        return outputs
+
+    def prepare_mask_latents(
+        self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance, noise_aug_strength
+    ):
+        # 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:
+            if self.transformer.config.add_noise_in_inpaint_model:
+                masked_image = add_noise_to_reference_video(masked_image, ratio=noise_aug_strength)
+            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,
+    ) -> 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)
+
+        p = self.transformer.config.patch_size
+        p_t = self.transformer.config.patch_size_t
+
+        base_size_width = self.transformer.config.sample_width // p
+        base_size_height = self.transformer.config.sample_height // p
+
+        if p_t is None:
+            # CogVideoX 1.0
+            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,
+            )
+        else:
+            # CogVideoX 1.5
+            base_num_frames = (num_frames + p_t - 1) // p_t
+
+            freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
+                embed_dim=self.transformer.config.attention_head_dim,
+                crops_coords=None,
+                grid_size=(grid_height, grid_width),
+                temporal_size=base_num_frames,
+                grid_type="slice",
+                max_size=(base_size_height, base_size_width),
+            )
+
+        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 attention_kwargs(self):
+        return self._attention_kwargs
+
+    @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,
+        mask_video: Union[torch.FloatTensor] = None,
+        masked_video_latents: 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,
+        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,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 226,
+        strength: float = 1,
+        noise_aug_strength: float = 0.0563,
+        comfyui_progressbar: bool = False,
+        temporal_multidiffusion_stride: int = 16,
+        use_trimask: bool = False,
+        zero_out_mask_region: bool = False,
+        binarize_mask: bool = False,
+        skip_unet: bool = False,
+        use_vae_mask: bool = False,
+        stack_mask: bool = False,
+    ) -> Union[CogVideoXFunPipelineOutput, 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.CogVideoXFunPipelineOutput`] or `tuple`:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXFunPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        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_height * self.vae_scale_factor_spatial
+        width = width or self.transformer.config.sample_width * self.vae_scale_factor_spatial
+        num_frames = num_frames or self.transformer.config.sample_frames
+
+        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._attention_kwargs = attention_kwargs
+        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
+        logger.info(f'Use cfg: {do_classifier_free_guidance}, guidance_scale={guidance_scale}')
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            negative_prompt,
+            do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+
+        # 4. set timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps, num_inference_steps = self.get_timesteps(
+            num_inference_steps=num_inference_steps, strength=strength, device=device
+        )
+        self._num_timesteps = len(timesteps)
+        if comfyui_progressbar:
+            from comfy.utils import ProgressBar
+            pbar = ProgressBar(num_inference_steps + 2)
+        # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+        latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt)
+        # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+        is_strength_max = strength == 1.0
+
+        # 5. Prepare latents.
+        if video is not None:
+            video_length = video.shape[2]
+            init_video = self.image_processor.preprocess(rearrange(video, "b c f h w -> (b f) c h w"), height=height, width=width) 
+            init_video = init_video.to(dtype=torch.float32)
+            init_video = rearrange(init_video, "(b f) c h w -> b c f h w", f=video_length)
+        else:
+            video_length = num_frames
+            init_video = None
+
+        # Magvae needs the number of frames to be 4n + 1.
+        local_latent_length = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        # For CogVideoX 1.5, the latent frames should be clipped to make it divisible by patch_size_t
+        patch_size_t = self.transformer.config.patch_size_t
+        additional_frames = 0
+        if patch_size_t is not None and local_latent_length % patch_size_t != 0:
+            additional_frames = local_latent_length % patch_size_t
+            num_frames -= additional_frames * self.vae_scale_factor_temporal
+        if num_frames <= 0:
+            num_frames = 1
+
+        num_channels_latents = self.vae.config.latent_channels
+        num_channels_transformer = self.transformer.config.in_channels
+        return_image_latents = num_channels_transformer == num_channels_latents
+
+        latents_outputs = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            video_length,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+            video=init_video,
+            timestep=latent_timestep,
+            is_strength_max=is_strength_max,
+            return_noise=True,
+            return_video_latents=return_image_latents,
+        )
+        if return_image_latents:
+            latents, noise, image_latents = latents_outputs
+        else:
+            latents, noise = latents_outputs
+        if comfyui_progressbar:
+            pbar.update(1)
+
+        if mask_video is not None:
+            if (mask_video == 255).all():
+                mask_latents = torch.zeros_like(latents)[:, :, :1].to(latents.device, latents.dtype)
+                masked_video_latents = torch.zeros_like(latents).to(latents.device, latents.dtype)
+
+                mask_input = torch.cat([mask_latents] * 2) if do_classifier_free_guidance else mask_latents
+                masked_video_latents_input = (
+                    torch.cat([masked_video_latents] * 2) if do_classifier_free_guidance else masked_video_latents
+                )
+                inpaint_latents = torch.cat([mask_input, masked_video_latents_input], dim=2).to(latents.dtype)
+            else:
+                # Prepare mask latent variables
+                video_length = video.shape[2]
+                mask_condition = self.mask_processor.preprocess(rearrange(mask_video, "b c f h w -> (b f) c h w"), height=height, width=width) 
+                if use_trimask:
+                    mask_condition = torch.where(mask_condition > 0.75, 1., mask_condition)
+                    mask_condition = torch.where((mask_condition <= 0.75) * (mask_condition >= 0.25), 127. / 255., mask_condition)
+                    mask_condition = torch.where(mask_condition < 0.25, 0., mask_condition)
+                else:
+                    mask_condition = torch.where(mask_condition > 0.5, 1., 0.)
+                    
+                mask_condition = mask_condition.to(dtype=torch.float32)
+                mask_condition = rearrange(mask_condition, "(b f) c h w -> b c f h w", f=video_length)
+
+                if num_channels_transformer != num_channels_latents:
+                    mask_condition_tile = torch.tile(mask_condition, [1, 3, 1, 1, 1])
+                    if masked_video_latents is None:
+                        if zero_out_mask_region:
+                            masked_video = init_video * (mask_condition_tile < 0.75) + torch.ones_like(init_video) * (mask_condition_tile > 0.75) * -1
+                        else:
+                            masked_video = init_video
+                    else:
+                        masked_video = masked_video_latents
+
+                    mask_encoded, masked_video_latents = self.prepare_mask_latents(
+                        1 - mask_condition_tile if use_vae_mask else None,
+                        masked_video,
+                        batch_size,
+                        height,
+                        width,
+                        prompt_embeds.dtype,
+                        device,
+                        generator,
+                        do_classifier_free_guidance,
+                        noise_aug_strength=noise_aug_strength,
+                    )
+                    if not use_vae_mask and not stack_mask:
+                        mask_latents = resize_mask(1 - mask_condition, masked_video_latents)
+                        if binarize_mask:
+                            if use_trimask:
+                                mask_latents = torch.where(mask_latents > 0.75, 1., mask_latents)
+                                mask_latents = torch.where((mask_latents <= 0.75) * (mask_latents >= 0.25), 0.5, mask_latents)
+                                mask_latents = torch.where(mask_latents < 0.25, 0., mask_latents)
+                            else:
+                                mask_latents = torch.where(mask_latents < 0.9, 0., 1.).to(mask_latents.dtype)
+
+                        mask_latents = mask_latents.to(masked_video_latents.device) * self.vae.config.scaling_factor
+
+                        mask = torch.tile(mask_condition, [1, num_channels_latents, 1, 1, 1])
+                        mask = F.interpolate(mask, size=latents.size()[-3:], mode='trilinear', align_corners=True).to(latents.device, latents.dtype)
+
+                        mask_input = torch.cat([mask_latents] * 2) if do_classifier_free_guidance else mask_latents
+                        mask = rearrange(mask, "b c f h w -> b f c h w")
+                    elif stack_mask:
+                        mask_latents = torch.cat([
+                            torch.repeat_interleave(mask_condition[:, :, 0:1], repeats=4, dim=2),
+                            mask_condition[:, :, 1:],
+                        ], dim=2)
+                        mask_latents = mask_latents.view(
+                            mask_latents.shape[0],
+                            mask_latents.shape[2] // 4,
+                            4,
+                            mask_latents.shape[3],
+                            mask_latents.shape[4],
+                        )
+                        mask_latents = mask_latents.transpose(1, 2)
+                        mask_latents = resize_mask(1 - mask_latents, masked_video_latents).to(latents.device, latents.dtype)
+                        mask_input = torch.cat([mask_latents] * 2) if do_classifier_free_guidance else mask_latents
+                    else:
+                        mask_input = (
+                            torch.cat([mask_encoded] * 2) if do_classifier_free_guidance else mask_encoded
+                        )
+
+                    masked_video_latents_input = (
+                        torch.cat([masked_video_latents] * 2) if do_classifier_free_guidance else masked_video_latents
+                    )
+
+                    mask_input = rearrange(mask_input, "b c f h w -> b f c h w")
+                    masked_video_latents_input = rearrange(masked_video_latents_input, "b c f h w -> b f c h w")
+
+                    # concat(binary mask, encode(mask * video))
+                    inpaint_latents = torch.cat([mask_input, masked_video_latents_input], dim=2).to(latents.dtype)
+                else:
+                    mask = torch.tile(mask_condition, [1, num_channels_latents, 1, 1, 1])
+                    mask = F.interpolate(mask, size=latents.size()[-3:], mode='trilinear', align_corners=True).to(latents.device, latents.dtype)
+                    mask = rearrange(mask, "b c f h w -> b f c h w")
+
+                    inpaint_latents = None
+        else:
+            if num_channels_transformer != num_channels_latents:
+                mask = torch.zeros_like(latents).to(latents.device, latents.dtype)
+                masked_video_latents = torch.zeros_like(latents).to(latents.device, latents.dtype)
+
+                mask_input = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+                masked_video_latents_input = (
+                    torch.cat([masked_video_latents] * 2) if do_classifier_free_guidance else masked_video_latents
+                )
+                inpaint_latents = torch.cat([mask_input, masked_video_latents_input], dim=1).to(latents.dtype)
+            else:
+                mask = torch.zeros_like(init_video[:, :1])
+                mask = torch.tile(mask, [1, num_channels_latents, 1, 1, 1])
+                mask = F.interpolate(mask, size=latents.size()[-3:], mode='trilinear', align_corners=True).to(latents.device, latents.dtype)
+                mask = rearrange(mask, "b c f h w -> b f c h w")
+
+                inpaint_latents = None
+        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)
+        logger.debug(f'Pipeline mask {mask_condition.shape} {mask_condition.dtype} {mask_condition.min()} {mask_condition.max()}')
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        latent_temporal_window_size = (num_frames - 1) // 4 + 1
+        if latents.size(1) > latent_temporal_window_size:
+            logger.info(f'Adopt temporal multidiffusion for the latents {latents.shape} {latents.dtype}')
+
+        # VAE experiment
+        if skip_unet:
+            masked_video_latents = rearrange(masked_video_latents, "b c f h w -> b f c h w")
+            if output_type == "numpy":
+                video = self.decode_latents(masked_video_latents)
+            elif not output_type == "latent":
+                video = self.decode_latents(masked_video_latents)
+                video = self.video_processor.postprocess_video(video=video, output_type=output_type)
+            else:
+                video = masked_video_latents
+
+            # Offload all models
+            self.maybe_free_model_hooks()
+
+            if not return_dict:
+                video = torch.from_numpy(video)
+
+            return CogVideoXFunPipelineOutput(videos=video)
+
+        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
+
+                def _sample(_latents, _inpaint_latents):
+                    # 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
+                    )
+
+                    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)
+
+                    # 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,
+                        inpaint_latents=_inpaint_latents,
+                    )[0]
+                    noise_pred = noise_pred.float()
+
+                    # perform guidance
+                    if use_dynamic_cfg:
+                        self._guidance_scale = 1 + guidance_scale * (
+                            (1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
+                        )
+                    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)
+                    return _latents
+
+                if latents.size(1) <= latent_temporal_window_size:
+                    latents = _sample(latents, inpaint_latents)
+                else:
+                    # adopt temporal multidiffusion
+                    latents_canvas = torch.zeros_like(latents).float()
+                    weights_canvas = torch.zeros(1, latents.size(1), 1, 1, 1).to(latents.device).float()
+                    temporal_stride = temporal_multidiffusion_stride // 4
+                    assert latent_temporal_window_size > temporal_stride
+
+                    time_beg = 0
+                    while time_beg < latents.size(1):
+                        time_end = min(time_beg + latent_temporal_window_size, latents.size(1))
+                        
+                        latents_i = latents[:, time_beg:time_end]
+                        if inpaint_latents is not None:
+                            inpaint_latents_i = inpaint_latents[:, time_beg:time_end]
+                        else:
+                            inpaint_latents_i = None
+                        
+                        latents_i = _sample(latents_i, inpaint_latents_i)
+
+                        weights_i = torch.ones(1, time_end - time_beg, 1, 1, 1).to(latents.device).to(latents.dtype)
+                        if time_beg > 0 and temporal_stride > 0:
+                            weights_i[:, :temporal_stride] = (torch.linspace(0., 1., temporal_stride + 2)[1:-1]
+                                                                .to(latents.device)
+                                                                .to(latents.dtype)
+                                                                .reshape(1, temporal_stride, 1, 1, 1))
+                        if time_end < latents.size(1) and temporal_stride > 0:
+                            weights_i[:, -temporal_stride:] = (torch.linspace(1., 0., temporal_stride + 2)[1:-1]
+                                                                .to(latents.device)
+                                                                .to(latents.dtype)
+                                                                .reshape(1, temporal_stride, 1, 1, 1))
+
+                        latents_canvas[:, time_beg:time_end] += latents_i * weights_i
+                        weights_canvas[:, time_beg:time_end] += weights_i
+
+                        time_beg = time_end - temporal_stride
+                        if time_end >= latents.size(1):
+                            break
+                    latents = (latents_canvas / weights_canvas).to(latents.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 CogVideoXFunPipelineOutput(videos=video)