From dd60660feed8c322bc52498c7cb43a73e0bd5bd5 Mon Sep 17 00:00:00 2001
From: kijai <40791699+kijai@users.noreply.github.com>
Date: Wed, 18 Sep 2024 17:00:03 +0300
Subject: [PATCH] initial CogVideoX-Fun support

---
 cogvideox_fun/autoencoder_magvit.py         | 1296 +++++++++++++++++++
 cogvideox_fun/pipeline_cogvideox_inpaint.py |  862 ++++++++++++
 cogvideox_fun/transformer_3d.py             |  605 +++++++++
 cogvideox_fun/utils.py                      |  246 ++++
 nodes.py                                    |  162 ++-
 pipeline_cogvideox.py                       |   19 +
 6 files changed, 3177 insertions(+), 13 deletions(-)
 create mode 100644 cogvideox_fun/autoencoder_magvit.py
 create mode 100644 cogvideox_fun/pipeline_cogvideox_inpaint.py
 create mode 100644 cogvideox_fun/transformer_3d.py
 create mode 100644 cogvideox_fun/utils.py

diff --git a/cogvideox_fun/autoencoder_magvit.py b/cogvideox_fun/autoencoder_magvit.py
new file mode 100644
index 0000000..9c2b906
--- /dev/null
+++ b/cogvideox_fun/autoencoder_magvit.py
@@ -0,0 +1,1296 @@
+# 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 Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders.single_file_model import FromOriginalModelMixin
+from diffusers.utils import logging
+from diffusers.utils.accelerate_utils import apply_forward_hook
+from diffusers.models.activations import get_activation
+from diffusers.models.downsampling import CogVideoXDownsample3D
+from diffusers.models.modeling_outputs import AutoencoderKLOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.upsampling import CogVideoXUpsample3D
+from diffusers.models.autoencoders.vae import DecoderOutput, DiagonalGaussianDistribution
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class CogVideoXSafeConv3d(nn.Conv3d):
+    r"""
+    A 3D convolution layer that splits the input tensor into smaller parts to avoid OOM in CogVideoX Model.
+    """
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        memory_count = torch.prod(torch.tensor(input.shape)).item() * 2 / 1024**3
+
+        # Set to 2GB, suitable for CuDNN
+        if memory_count > 2:
+            kernel_size = self.kernel_size[0]
+            part_num = int(memory_count / 2) + 1
+            input_chunks = torch.chunk(input, part_num, dim=2)
+
+            if kernel_size > 1:
+                input_chunks = [input_chunks[0]] + [
+                    torch.cat((input_chunks[i - 1][:, :, -kernel_size + 1 :], input_chunks[i]), dim=2)
+                    for i in range(1, len(input_chunks))
+                ]
+
+            output_chunks = []
+            for input_chunk in input_chunks:
+                output_chunks.append(super().forward(input_chunk))
+            output = torch.cat(output_chunks, dim=2)
+            return output
+        else:
+            return super().forward(input)
+
+
+class CogVideoXCausalConv3d(nn.Module):
+    r"""A 3D causal convolution layer that pads the input tensor to ensure causality in CogVideoX Model.
+
+    Args:
+        in_channels (`int`): Number of channels in the input tensor.
+        out_channels (`int`): Number of output channels produced by the convolution.
+        kernel_size (`int` or `Tuple[int, int, int]`): Kernel size of the convolutional kernel.
+        stride (`int`, defaults to `1`): Stride of the convolution.
+        dilation (`int`, defaults to `1`): Dilation rate of the convolution.
+        pad_mode (`str`, defaults to `"constant"`): Padding mode.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, Tuple[int, int, int]],
+        stride: int = 1,
+        dilation: int = 1,
+        pad_mode: str = "constant",
+    ):
+        super().__init__()
+
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size,) * 3
+
+        time_kernel_size, height_kernel_size, width_kernel_size = kernel_size
+
+        self.pad_mode = pad_mode
+        time_pad = dilation * (time_kernel_size - 1) + (1 - stride)
+        height_pad = height_kernel_size // 2
+        width_pad = width_kernel_size // 2
+
+        self.height_pad = height_pad
+        self.width_pad = width_pad
+        self.time_pad = time_pad
+        self.time_causal_padding = (width_pad, width_pad, height_pad, height_pad, time_pad, 0)
+
+        self.temporal_dim = 2
+        self.time_kernel_size = time_kernel_size
+
+        stride = (stride, 1, 1)
+        dilation = (dilation, 1, 1)
+        self.conv = CogVideoXSafeConv3d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            dilation=dilation,
+        )
+
+        self.conv_cache = None
+
+    def fake_context_parallel_forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        kernel_size = self.time_kernel_size
+        if kernel_size > 1:
+            cached_inputs = (
+                [self.conv_cache] if self.conv_cache is not None else [inputs[:, :, :1]] * (kernel_size - 1)
+            )
+            inputs = torch.cat(cached_inputs + [inputs], dim=2)
+        return inputs
+
+    def _clear_fake_context_parallel_cache(self):
+        del self.conv_cache
+        self.conv_cache = None
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        inputs = self.fake_context_parallel_forward(inputs)
+
+        self._clear_fake_context_parallel_cache()
+        # Note: we could move these to the cpu for a lower maximum memory usage but its only a few
+        # hundred megabytes and so let's not do it for now
+        self.conv_cache = inputs[:, :, -self.time_kernel_size + 1 :].clone()
+
+        padding_2d = (self.width_pad, self.width_pad, self.height_pad, self.height_pad)
+        inputs = F.pad(inputs, padding_2d, mode="constant", value=0)
+
+        output = self.conv(inputs)
+        return output
+
+
+class CogVideoXSpatialNorm3D(nn.Module):
+    r"""
+    Spatially conditioned normalization as defined in https://arxiv.org/abs/2209.09002. This implementation is specific
+    to 3D-video like data.
+
+    CogVideoXSafeConv3d is used instead of nn.Conv3d to avoid OOM in CogVideoX Model.
+
+    Args:
+        f_channels (`int`):
+            The number of channels for input to group normalization layer, and output of the spatial norm layer.
+        zq_channels (`int`):
+            The number of channels for the quantized vector as described in the paper.
+        groups (`int`):
+            Number of groups to separate the channels into for group normalization.
+    """
+
+    def __init__(
+        self,
+        f_channels: int,
+        zq_channels: int,
+        groups: int = 32,
+    ):
+        super().__init__()
+        self.norm_layer = nn.GroupNorm(num_channels=f_channels, num_groups=groups, eps=1e-6, affine=True)
+        self.conv_y = CogVideoXCausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+        self.conv_b = CogVideoXCausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+
+    def forward(self, f: torch.Tensor, zq: torch.Tensor) -> torch.Tensor:
+        if f.shape[2] > 1 and f.shape[2] % 2 == 1:
+            f_first, f_rest = f[:, :, :1], f[:, :, 1:]
+            f_first_size, f_rest_size = f_first.shape[-3:], f_rest.shape[-3:]
+            z_first, z_rest = zq[:, :, :1], zq[:, :, 1:]
+            z_first = F.interpolate(z_first, size=f_first_size)
+            z_rest = F.interpolate(z_rest, size=f_rest_size)
+            zq = torch.cat([z_first, z_rest], dim=2)
+        else:
+            zq = F.interpolate(zq, size=f.shape[-3:])
+
+        norm_f = self.norm_layer(f)
+        new_f = norm_f * self.conv_y(zq) + self.conv_b(zq)
+        return new_f
+
+
+class CogVideoXResnetBlock3D(nn.Module):
+    r"""
+    A 3D ResNet block used in the CogVideoX model.
+
+    Args:
+        in_channels (`int`):
+            Number of input channels.
+        out_channels (`int`, *optional*):
+            Number of output channels. If None, defaults to `in_channels`.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        non_linearity (`str`, defaults to `"swish"`):
+            Activation function to use.
+        conv_shortcut (bool, defaults to `False`):
+            Whether or not to use a convolution shortcut.
+        spatial_norm_dim (`int`, *optional*):
+            The dimension to use for spatial norm if it is to be used instead of group norm.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        dropout: float = 0.0,
+        temb_channels: int = 512,
+        groups: int = 32,
+        eps: float = 1e-6,
+        non_linearity: str = "swish",
+        conv_shortcut: bool = False,
+        spatial_norm_dim: Optional[int] = None,
+        pad_mode: str = "first",
+    ):
+        super().__init__()
+
+        out_channels = out_channels or in_channels
+
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.nonlinearity = get_activation(non_linearity)
+        self.use_conv_shortcut = conv_shortcut
+
+        if spatial_norm_dim is None:
+            self.norm1 = nn.GroupNorm(num_channels=in_channels, num_groups=groups, eps=eps)
+            self.norm2 = nn.GroupNorm(num_channels=out_channels, num_groups=groups, eps=eps)
+        else:
+            self.norm1 = CogVideoXSpatialNorm3D(
+                f_channels=in_channels,
+                zq_channels=spatial_norm_dim,
+                groups=groups,
+            )
+            self.norm2 = CogVideoXSpatialNorm3D(
+                f_channels=out_channels,
+                zq_channels=spatial_norm_dim,
+                groups=groups,
+            )
+
+        self.conv1 = CogVideoXCausalConv3d(
+            in_channels=in_channels, out_channels=out_channels, kernel_size=3, pad_mode=pad_mode
+        )
+
+        if temb_channels > 0:
+            self.temb_proj = nn.Linear(in_features=temb_channels, out_features=out_channels)
+
+        self.dropout = nn.Dropout(dropout)
+        self.conv2 = CogVideoXCausalConv3d(
+            in_channels=out_channels, out_channels=out_channels, kernel_size=3, pad_mode=pad_mode
+        )
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = CogVideoXCausalConv3d(
+                    in_channels=in_channels, out_channels=out_channels, kernel_size=3, pad_mode=pad_mode
+                )
+            else:
+                self.conv_shortcut = CogVideoXSafeConv3d(
+                    in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, padding=0
+                )
+
+    def forward(
+        self,
+        inputs: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        zq: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        hidden_states = inputs
+
+        if zq is not None:
+            hidden_states = self.norm1(hidden_states, zq)
+        else:
+            hidden_states = self.norm1(hidden_states)
+
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.conv1(hidden_states)
+
+        if temb is not None:
+            hidden_states = hidden_states + self.temb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if zq is not None:
+            hidden_states = self.norm2(hidden_states, zq)
+        else:
+            hidden_states = self.norm2(hidden_states)
+
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.in_channels != self.out_channels:
+            inputs = self.conv_shortcut(inputs)
+
+        hidden_states = hidden_states + inputs
+        return hidden_states
+
+
+class CogVideoXDownBlock3D(nn.Module):
+    r"""
+    A downsampling block used in the CogVideoX model.
+
+    Args:
+        in_channels (`int`):
+            Number of input channels.
+        out_channels (`int`, *optional*):
+            Number of output channels. If None, defaults to `in_channels`.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        num_layers (`int`, defaults to `1`):
+            Number of resnet layers.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        resnet_eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        resnet_act_fn (`str`, defaults to `"swish"`):
+            Activation function to use.
+        resnet_groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        add_downsample (`bool`, defaults to `True`):
+            Whether or not to use a downsampling layer. If not used, output dimension would be same as input dimension.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to downsample across temporal dimension.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        add_downsample: bool = True,
+        downsample_padding: int = 0,
+        compress_time: bool = False,
+        pad_mode: str = "first",
+    ):
+        super().__init__()
+
+        resnets = []
+        for i in range(num_layers):
+            in_channel = in_channels if i == 0 else out_channels
+            resnets.append(
+                CogVideoXResnetBlock3D(
+                    in_channels=in_channel,
+                    out_channels=out_channels,
+                    dropout=dropout,
+                    temb_channels=temb_channels,
+                    groups=resnet_groups,
+                    eps=resnet_eps,
+                    non_linearity=resnet_act_fn,
+                    pad_mode=pad_mode,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.downsamplers = None
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    CogVideoXDownsample3D(
+                        out_channels, out_channels, padding=downsample_padding, compress_time=compress_time
+                    )
+                ]
+            )
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        zq: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def create_forward(*inputs):
+                        return module(*inputs)
+
+                    return create_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb, zq
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, zq)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+        return hidden_states
+
+
+class CogVideoXMidBlock3D(nn.Module):
+    r"""
+    A middle block used in the CogVideoX model.
+
+    Args:
+        in_channels (`int`):
+            Number of input channels.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        num_layers (`int`, defaults to `1`):
+            Number of resnet layers.
+        resnet_eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        resnet_act_fn (`str`, defaults to `"swish"`):
+            Activation function to use.
+        resnet_groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        spatial_norm_dim (`int`, *optional*):
+            The dimension to use for spatial norm if it is to be used instead of group norm.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        spatial_norm_dim: Optional[int] = None,
+        pad_mode: str = "first",
+    ):
+        super().__init__()
+
+        resnets = []
+        for _ in range(num_layers):
+            resnets.append(
+                CogVideoXResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    dropout=dropout,
+                    temb_channels=temb_channels,
+                    groups=resnet_groups,
+                    eps=resnet_eps,
+                    spatial_norm_dim=spatial_norm_dim,
+                    non_linearity=resnet_act_fn,
+                    pad_mode=pad_mode,
+                )
+            )
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        zq: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def create_forward(*inputs):
+                        return module(*inputs)
+
+                    return create_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb, zq
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, zq)
+
+        return hidden_states
+
+
+class CogVideoXUpBlock3D(nn.Module):
+    r"""
+    An upsampling block used in the CogVideoX model.
+
+    Args:
+        in_channels (`int`):
+            Number of input channels.
+        out_channels (`int`, *optional*):
+            Number of output channels. If None, defaults to `in_channels`.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        num_layers (`int`, defaults to `1`):
+            Number of resnet layers.
+        resnet_eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        resnet_act_fn (`str`, defaults to `"swish"`):
+            Activation function to use.
+        resnet_groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        spatial_norm_dim (`int`, defaults to `16`):
+            The dimension to use for spatial norm if it is to be used instead of group norm.
+        add_upsample (`bool`, defaults to `True`):
+            Whether or not to use a upsampling layer. If not used, output dimension would be same as input dimension.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to downsample across temporal dimension.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        spatial_norm_dim: int = 16,
+        add_upsample: bool = True,
+        upsample_padding: int = 1,
+        compress_time: bool = False,
+        pad_mode: str = "first",
+    ):
+        super().__init__()
+
+        resnets = []
+        for i in range(num_layers):
+            in_channel = in_channels if i == 0 else out_channels
+            resnets.append(
+                CogVideoXResnetBlock3D(
+                    in_channels=in_channel,
+                    out_channels=out_channels,
+                    dropout=dropout,
+                    temb_channels=temb_channels,
+                    groups=resnet_groups,
+                    eps=resnet_eps,
+                    non_linearity=resnet_act_fn,
+                    spatial_norm_dim=spatial_norm_dim,
+                    pad_mode=pad_mode,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.upsamplers = None
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [
+                    CogVideoXUpsample3D(
+                        out_channels, out_channels, padding=upsample_padding, compress_time=compress_time
+                    )
+                ]
+            )
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        zq: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        r"""Forward method of the `CogVideoXUpBlock3D` class."""
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def create_forward(*inputs):
+                        return module(*inputs)
+
+                    return create_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb, zq
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, zq)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states)
+
+        return hidden_states
+
+
+class CogVideoXEncoder3D(nn.Module):
+    r"""
+    The `CogVideoXEncoder3D` layer of a variational autoencoder that encodes its input into a latent representation.
+
+    Args:
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        down_block_types (`Tuple[str, ...]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            The types of down blocks to use. See `~diffusers.models.unet_2d_blocks.get_down_block` for available
+            options.
+        block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+            The number of output channels for each block.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+        layers_per_block (`int`, *optional*, defaults to 2):
+            The number of layers per block.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 16,
+        down_block_types: Tuple[str, ...] = (
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+        ),
+        block_out_channels: Tuple[int, ...] = (128, 256, 256, 512),
+        layers_per_block: int = 3,
+        act_fn: str = "silu",
+        norm_eps: float = 1e-6,
+        norm_num_groups: int = 32,
+        dropout: float = 0.0,
+        pad_mode: str = "first",
+        temporal_compression_ratio: float = 4,
+    ):
+        super().__init__()
+
+        # log2 of temporal_compress_times
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CogVideoXCausalConv3d(in_channels, block_out_channels[0], kernel_size=3, pad_mode=pad_mode)
+        self.down_blocks = nn.ModuleList([])
+
+        # down blocks
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            if down_block_type == "CogVideoXDownBlock3D":
+                down_block = CogVideoXDownBlock3D(
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    temb_channels=0,
+                    dropout=dropout,
+                    num_layers=layers_per_block,
+                    resnet_eps=norm_eps,
+                    resnet_act_fn=act_fn,
+                    resnet_groups=norm_num_groups,
+                    add_downsample=not is_final_block,
+                    compress_time=compress_time,
+                )
+            else:
+                raise ValueError("Invalid `down_block_type` encountered. Must be `CogVideoXDownBlock3D`")
+
+            self.down_blocks.append(down_block)
+
+        # mid block
+        self.mid_block = CogVideoXMidBlock3D(
+            in_channels=block_out_channels[-1],
+            temb_channels=0,
+            dropout=dropout,
+            num_layers=2,
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            resnet_groups=norm_num_groups,
+            pad_mode=pad_mode,
+        )
+
+        self.norm_out = nn.GroupNorm(norm_num_groups, block_out_channels[-1], eps=1e-6)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CogVideoXCausalConv3d(
+            block_out_channels[-1], 2 * out_channels, kernel_size=3, pad_mode=pad_mode
+        )
+
+        self.gradient_checkpointing = False
+
+    def forward(self, sample: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
+        r"""The forward method of the `CogVideoXEncoder3D` class."""
+        hidden_states = self.conv_in(sample)
+
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            # 1. Down
+            for down_block in self.down_blocks:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(down_block), hidden_states, temb, None
+                )
+
+            # 2. Mid
+            hidden_states = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(self.mid_block), hidden_states, temb, None
+            )
+        else:
+            # 1. Down
+            for down_block in self.down_blocks:
+                hidden_states = down_block(hidden_states, temb, None)
+
+            # 2. Mid
+            hidden_states = self.mid_block(hidden_states, temb, None)
+
+        # 3. Post-process
+        hidden_states = self.norm_out(hidden_states)
+        hidden_states = self.conv_act(hidden_states)
+        hidden_states = self.conv_out(hidden_states)
+        return hidden_states
+
+
+class CogVideoXDecoder3D(nn.Module):
+    r"""
+    The `CogVideoXDecoder3D` layer of a variational autoencoder that decodes its latent representation into an output
+    sample.
+
+    Args:
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        up_block_types (`Tuple[str, ...]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            The types of up blocks to use. See `~diffusers.models.unet_2d_blocks.get_up_block` for available options.
+        block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+            The number of output channels for each block.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+        layers_per_block (`int`, *optional*, defaults to 2):
+            The number of layers per block.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: int = 16,
+        out_channels: int = 3,
+        up_block_types: Tuple[str, ...] = (
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+        ),
+        block_out_channels: Tuple[int, ...] = (128, 256, 256, 512),
+        layers_per_block: int = 3,
+        act_fn: str = "silu",
+        norm_eps: float = 1e-6,
+        norm_num_groups: int = 32,
+        dropout: float = 0.0,
+        pad_mode: str = "first",
+        temporal_compression_ratio: float = 4,
+    ):
+        super().__init__()
+
+        reversed_block_out_channels = list(reversed(block_out_channels))
+
+        self.conv_in = CogVideoXCausalConv3d(
+            in_channels, reversed_block_out_channels[0], kernel_size=3, pad_mode=pad_mode
+        )
+
+        # mid block
+        self.mid_block = CogVideoXMidBlock3D(
+            in_channels=reversed_block_out_channels[0],
+            temb_channels=0,
+            num_layers=2,
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            resnet_groups=norm_num_groups,
+            spatial_norm_dim=in_channels,
+            pad_mode=pad_mode,
+        )
+
+        # up blocks
+        self.up_blocks = nn.ModuleList([])
+
+        output_channel = reversed_block_out_channels[0]
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        for i, up_block_type in enumerate(up_block_types):
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            if up_block_type == "CogVideoXUpBlock3D":
+                up_block = CogVideoXUpBlock3D(
+                    in_channels=prev_output_channel,
+                    out_channels=output_channel,
+                    temb_channels=0,
+                    dropout=dropout,
+                    num_layers=layers_per_block + 1,
+                    resnet_eps=norm_eps,
+                    resnet_act_fn=act_fn,
+                    resnet_groups=norm_num_groups,
+                    spatial_norm_dim=in_channels,
+                    add_upsample=not is_final_block,
+                    compress_time=compress_time,
+                    pad_mode=pad_mode,
+                )
+                prev_output_channel = output_channel
+            else:
+                raise ValueError("Invalid `up_block_type` encountered. Must be `CogVideoXUpBlock3D`")
+
+            self.up_blocks.append(up_block)
+
+        self.norm_out = CogVideoXSpatialNorm3D(reversed_block_out_channels[-1], in_channels, groups=norm_num_groups)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CogVideoXCausalConv3d(
+            reversed_block_out_channels[-1], out_channels, kernel_size=3, pad_mode=pad_mode
+        )
+
+        self.gradient_checkpointing = False
+
+    def forward(self, sample: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
+        r"""The forward method of the `CogVideoXDecoder3D` class."""
+        hidden_states = self.conv_in(sample)
+
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            # 1. Mid
+            hidden_states = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(self.mid_block), hidden_states, temb, sample
+            )
+
+            # 2. Up
+            for up_block in self.up_blocks:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(up_block), hidden_states, temb, sample
+                )
+        else:
+            # 1. Mid
+            hidden_states = self.mid_block(hidden_states, temb, sample)
+
+            # 2. Up
+            for up_block in self.up_blocks:
+                hidden_states = up_block(hidden_states, temb, sample)
+
+        # 3. Post-process
+        hidden_states = self.norm_out(hidden_states, sample)
+        hidden_states = self.conv_act(hidden_states)
+        hidden_states = self.conv_out(hidden_states)
+        return hidden_states
+
+
+class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+    r"""
+    A VAE model with KL loss for encoding images into latents and decoding latent representations into images. Used in
+    [CogVideoX](https://github.com/THUDM/CogVideo).
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+        out_channels (int,  *optional*, defaults to 3): Number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            Tuple of downsample block types.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            Tuple of upsample block types.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+            Tuple of block output channels.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+        scaling_factor (`float`, *optional*, defaults to `1.15258426`):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+        force_upcast (`bool`, *optional*, default to `True`):
+            If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+            can be fine-tuned / trained to a lower range without loosing too much precision in which case
+            `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["CogVideoXResnetBlock3D"]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str] = (
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+        ),
+        up_block_types: Tuple[str] = (
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+        ),
+        block_out_channels: Tuple[int] = (128, 256, 256, 512),
+        latent_channels: int = 16,
+        layers_per_block: int = 3,
+        act_fn: str = "silu",
+        norm_eps: float = 1e-6,
+        norm_num_groups: int = 32,
+        temporal_compression_ratio: float = 4,
+        sample_height: int = 480,
+        sample_width: int = 720,
+        scaling_factor: float = 1.15258426,
+        shift_factor: Optional[float] = None,
+        latents_mean: Optional[Tuple[float]] = None,
+        latents_std: Optional[Tuple[float]] = None,
+        force_upcast: float = True,
+        use_quant_conv: bool = False,
+        use_post_quant_conv: bool = False,
+    ):
+        super().__init__()
+
+        self.encoder = CogVideoXEncoder3D(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            down_block_types=down_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_eps=norm_eps,
+            norm_num_groups=norm_num_groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+        self.decoder = CogVideoXDecoder3D(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            up_block_types=up_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_eps=norm_eps,
+            norm_num_groups=norm_num_groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+        self.quant_conv = CogVideoXSafeConv3d(2 * out_channels, 2 * out_channels, 1) if use_quant_conv else None
+        self.post_quant_conv = CogVideoXSafeConv3d(out_channels, out_channels, 1) if use_post_quant_conv else None
+
+        self.use_slicing = False
+        self.use_tiling = False
+
+        # Can be increased to decode more latent frames at once, but comes at a reasonable memory cost and it is not
+        # recommended because the temporal parts of the VAE, here, are tricky to understand.
+        # If you decode X latent frames together, the number of output frames is:
+        #     (X + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale)) => X + 6 frames
+        #
+        # Example with num_latent_frames_batch_size = 2:
+        #     - 12 latent frames: (0, 1), (2, 3), (4, 5), (6, 7), (8, 9), (10, 11) are processed together
+        #         => (12 // 2 frame slices) * ((2 num_latent_frames_batch_size) + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale))
+        #         => 6 * 8 = 48 frames
+        #     - 13 latent frames: (0, 1, 2) (special case), (3, 4), (5, 6), (7, 8), (9, 10), (11, 12) are processed together
+        #         => (1 frame slice) * ((3 num_latent_frames_batch_size) + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale)) +
+        #            ((13 - 3) // 2) * ((2 num_latent_frames_batch_size) + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale))
+        #         => 1 * 9 + 5 * 8 = 49 frames
+        # It has been implemented this way so as to not have "magic values" in the code base that would be hard to explain. Note that
+        # setting it to anything other than 2 would give poor results because the VAE hasn't been trained to be adaptive with different
+        # number of temporal frames.
+        self.num_latent_frames_batch_size = 2
+
+        # We make the minimum height and width of sample for tiling half that of the generally supported
+        self.tile_sample_min_height = sample_height // 2
+        self.tile_sample_min_width = sample_width // 2
+        self.tile_latent_min_height = int(
+            self.tile_sample_min_height / (2 ** (len(self.config.block_out_channels) - 1))
+        )
+        self.tile_latent_min_width = int(self.tile_sample_min_width / (2 ** (len(self.config.block_out_channels) - 1)))
+
+        # These are experimental overlap factors that were chosen based on experimentation and seem to work best for
+        # 720x480 (WxH) resolution. The above resolution is the strongly recommended generation resolution in CogVideoX
+        # and so the tiling implementation has only been tested on those specific resolutions.
+        self.tile_overlap_factor_height = 1 / 6
+        self.tile_overlap_factor_width = 1 / 5
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (CogVideoXEncoder3D, CogVideoXDecoder3D)):
+            module.gradient_checkpointing = value
+
+    def _clear_fake_context_parallel_cache(self):
+        for name, module in self.named_modules():
+            if isinstance(module, CogVideoXCausalConv3d):
+                logger.debug(f"Clearing fake Context Parallel cache for layer: {name}")
+                module._clear_fake_context_parallel_cache()
+
+    def enable_tiling(
+        self,
+        tile_sample_min_height: Optional[int] = None,
+        tile_sample_min_width: Optional[int] = None,
+        tile_overlap_factor_height: Optional[float] = None,
+        tile_overlap_factor_width: Optional[float] = None,
+    ) -> None:
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+
+        Args:
+            tile_sample_min_height (`int`, *optional*):
+                The minimum height required for a sample to be separated into tiles across the height dimension.
+            tile_sample_min_width (`int`, *optional*):
+                The minimum width required for a sample to be separated into tiles across the width dimension.
+            tile_overlap_factor_height (`int`, *optional*):
+                The minimum amount of overlap between two consecutive vertical tiles. This is to ensure that there are
+                no tiling artifacts produced across the height dimension. Must be between 0 and 1. Setting a higher
+                value might cause more tiles to be processed leading to slow down of the decoding process.
+            tile_overlap_factor_width (`int`, *optional*):
+                The minimum amount of overlap between two consecutive horizontal tiles. This is to ensure that there
+                are no tiling artifacts produced across the width dimension. Must be between 0 and 1. Setting a higher
+                value might cause more tiles to be processed leading to slow down of the decoding process.
+        """
+        self.use_tiling = True
+        self.tile_sample_min_height = tile_sample_min_height or self.tile_sample_min_height
+        self.tile_sample_min_width = tile_sample_min_width or self.tile_sample_min_width
+        self.tile_latent_min_height = int(
+            self.tile_sample_min_height / (2 ** (len(self.config.block_out_channels) - 1))
+        )
+        self.tile_latent_min_width = int(self.tile_sample_min_width / (2 ** (len(self.config.block_out_channels) - 1)))
+        self.tile_overlap_factor_height = tile_overlap_factor_height or self.tile_overlap_factor_height
+        self.tile_overlap_factor_width = tile_overlap_factor_width or self.tile_overlap_factor_width
+
+    def disable_tiling(self) -> None:
+        r"""
+        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_tiling = False
+
+    def enable_slicing(self) -> None:
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.use_slicing = True
+
+    def disable_slicing(self) -> None:
+        r"""
+        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_slicing = False
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.Tensor, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+        """
+        Encode a batch of images into latents.
+
+        Args:
+            x (`torch.Tensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+                The latent representations of the encoded images. If `return_dict` is True, a
+                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+        """
+        batch_size, num_channels, num_frames, height, width = x.shape
+        if num_frames == 1:
+            h = self.encoder(x)
+            if self.quant_conv is not None:
+                h = self.quant_conv(h)
+            posterior = DiagonalGaussianDistribution(h)
+        else:
+            frame_batch_size = 4
+            h = []
+            for i in range(num_frames // frame_batch_size):
+                remaining_frames = num_frames % frame_batch_size
+                start_frame = frame_batch_size * i + (0 if i == 0 else remaining_frames)
+                end_frame = frame_batch_size * (i + 1) + remaining_frames
+                z_intermediate = x[:, :, start_frame:end_frame]
+                z_intermediate = self.encoder(z_intermediate)
+                if self.quant_conv is not None:
+                    z_intermediate = self.quant_conv(z_intermediate)
+                h.append(z_intermediate)
+            self._clear_fake_context_parallel_cache()
+            h = torch.cat(h, dim=2)
+            posterior = DiagonalGaussianDistribution(h)
+        if not return_dict:
+            return (posterior,)
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def _decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        batch_size, num_channels, num_frames, height, width = z.shape
+
+        if self.use_tiling and (width > self.tile_latent_min_width or height > self.tile_latent_min_height):
+            return self.tiled_decode(z, return_dict=return_dict)
+
+        if num_frames == 1:
+            dec = []
+            z_intermediate = z
+            if self.post_quant_conv is not None:
+                z_intermediate = self.post_quant_conv(z_intermediate)
+            z_intermediate = self.decoder(z_intermediate)
+            dec.append(z_intermediate)
+        else:
+            frame_batch_size = self.num_latent_frames_batch_size
+            dec = []
+            for i in range(num_frames // frame_batch_size):
+                remaining_frames = num_frames % frame_batch_size
+                start_frame = frame_batch_size * i + (0 if i == 0 else remaining_frames)
+                end_frame = frame_batch_size * (i + 1) + remaining_frames
+                z_intermediate = z[:, :, start_frame:end_frame]
+                if self.post_quant_conv is not None:
+                    z_intermediate = self.post_quant_conv(z_intermediate)
+                z_intermediate = self.decoder(z_intermediate)
+                dec.append(z_intermediate)
+
+        self._clear_fake_context_parallel_cache()
+        dec = torch.cat(dec, dim=2)
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    @apply_forward_hook
+    def decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        """
+        Decode a batch of images.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        if self.use_slicing and z.shape[0] > 1:
+            decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
+            decoded = torch.cat(decoded_slices)
+        else:
+            decoded = self._decode(z).sample
+
+        if not return_dict:
+            return (decoded,)
+        return DecoderOutput(sample=decoded)
+
+    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[3], b.shape[3], blend_extent)
+        for y in range(blend_extent):
+            b[:, :, :, y, :] = a[:, :, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, :, y, :] * (
+                y / blend_extent
+            )
+        return b
+
+    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[4], b.shape[4], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, :, :, x] = a[:, :, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, :, x] * (
+                x / blend_extent
+            )
+        return b
+
+    def tiled_decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        r"""
+        Decode a batch of images using a tiled decoder.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        # Rough memory assessment:
+        #   - In CogVideoX-2B, there are a total of 24 CausalConv3d layers.
+        #   - The biggest intermediate dimensions are: [1, 128, 9, 480, 720].
+        #   - Assume fp16 (2 bytes per value).
+        # Memory required: 1 * 128 * 9 * 480 * 720 * 24 * 2 / 1024**3 = 17.8 GB
+        #
+        # Memory assessment when using tiling:
+        #   - Assume everything as above but now HxW is 240x360 by tiling in half
+        # Memory required: 1 * 128 * 9 * 240 * 360 * 24 * 2 / 1024**3 = 4.5 GB
+
+        batch_size, num_channels, num_frames, height, width = z.shape
+
+        overlap_height = int(self.tile_latent_min_height * (1 - self.tile_overlap_factor_height))
+        overlap_width = int(self.tile_latent_min_width * (1 - self.tile_overlap_factor_width))
+        blend_extent_height = int(self.tile_sample_min_height * self.tile_overlap_factor_height)
+        blend_extent_width = int(self.tile_sample_min_width * self.tile_overlap_factor_width)
+        row_limit_height = self.tile_sample_min_height - blend_extent_height
+        row_limit_width = self.tile_sample_min_width - blend_extent_width
+        frame_batch_size = self.num_latent_frames_batch_size
+
+        # Split z into overlapping tiles and decode them separately.
+        # The tiles have an overlap to avoid seams between tiles.
+        rows = []
+        for i in range(0, height, overlap_height):
+            row = []
+            for j in range(0, width, overlap_width):
+                time = []
+                for k in range(num_frames // frame_batch_size):
+                    remaining_frames = num_frames % frame_batch_size
+                    start_frame = frame_batch_size * k + (0 if k == 0 else remaining_frames)
+                    end_frame = frame_batch_size * (k + 1) + remaining_frames
+                    tile = z[
+                        :,
+                        :,
+                        start_frame:end_frame,
+                        i : i + self.tile_latent_min_height,
+                        j : j + self.tile_latent_min_width,
+                    ]
+                    if self.post_quant_conv is not None:
+                        tile = self.post_quant_conv(tile)
+                    tile = self.decoder(tile)
+                    time.append(tile)
+                self._clear_fake_context_parallel_cache()
+                row.append(torch.cat(time, dim=2))
+            rows.append(row)
+
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent_height)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent_width)
+                result_row.append(tile[:, :, :, :row_limit_height, :row_limit_width])
+            result_rows.append(torch.cat(result_row, dim=4))
+
+        dec = torch.cat(result_rows, dim=3)
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+    ) -> Union[torch.Tensor, torch.Tensor]:
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z)
+        if not return_dict:
+            return (dec,)
+        return dec
diff --git a/cogvideox_fun/pipeline_cogvideox_inpaint.py b/cogvideox_fun/pipeline_cogvideox_inpaint.py
new file mode 100644
index 0000000..01fc8ac
--- /dev/null
+++ b/cogvideox_fun/pipeline_cogvideox_inpaint.py
@@ -0,0 +1,862 @@
+# 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 transformers import T5EncoderModel, T5Tokenizer
+
+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
+
+
+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
+
+
+@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_Inpaint(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 = "text_encoder->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,
+        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
+    ):
+        # 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,
+    ) -> 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,
+        )
+
+        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,
+        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,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 226,
+        strength: float = 1,
+        comfyui_progressbar: bool = False,
+    ) -> 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
+
+        self.vae.to(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. 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:
+            init_video = None
+
+        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) 
+                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:
+                        masked_video = init_video * (mask_condition_tile < 0.5) + torch.ones_like(init_video) * (mask_condition_tile > 0.5) * -1
+                    else:
+                        masked_video = masked_video_latents
+
+                    _, masked_video_latents = self.prepare_mask_latents(
+                        None,
+                        masked_video,
+                        batch_size,
+                        height,
+                        width,
+                        prompt_embeds.dtype,
+                        device,
+                        generator,
+                        do_classifier_free_guidance,
+                    )
+                    mask_latents = resize_mask(1 - mask_condition, masked_video_latents)
+                    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
+                    masked_video_latents_input = (
+                        torch.cat([masked_video_latents] * 2) if do_classifier_free_guidance else masked_video_latents
+                    )
+
+                    mask = rearrange(mask, "b c f h w -> b f c h w")
+                    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")
+
+                    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)
+
+        # 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
+        )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        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
+
+                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)
+
+                # 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
\ No newline at end of file
diff --git a/cogvideox_fun/transformer_3d.py b/cogvideox_fun/transformer_3d.py
new file mode 100644
index 0000000..b80af91
--- /dev/null
+++ b/cogvideox_fun/transformer_3d.py
@@ -0,0 +1,605 @@
+# 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
+
+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,
+    ):
+        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,
+        )
+
+    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,
+    ) -> 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
+        )
+
+        # attention
+        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,
+        )
+
+        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,
+    ):
+        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
+
+    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,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, 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)
+        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:]
+
+        # 4. Transformer blocks
+        for i, block in enumerate(self.transformer_blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states,
+                    emb,
+                    image_rotary_emb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    temb=emb,
+                    image_rotary_emb=image_rotary_emb,
+                )
+
+        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 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
\ No newline at end of file
diff --git a/cogvideox_fun/utils.py b/cogvideox_fun/utils.py
new file mode 100644
index 0000000..e9c5cc7
--- /dev/null
+++ b/cogvideox_fun/utils.py
@@ -0,0 +1,246 @@
+import os
+import gc
+import imageio
+import numpy as np
+import torch
+import torchvision
+import cv2
+from einops import rearrange
+from PIL import Image
+
+# Copyright (c) OpenMMLab. All rights reserved.
+import os
+import cv2
+import numpy as np
+import torch
+from PIL import Image
+
+
+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],
+    '0.42': [320.0, 768.0], '0.48': [352.0, 736.0], '0.5': [352.0, 704.0], '0.52': [352.0, 672.0],
+    '0.57': [384.0, 672.0], '0.6': [384.0, 640.0], '0.68': [416.0, 608.0], '0.72': [416.0, 576.0],
+    '0.78': [448.0, 576.0], '0.82': [448.0, 544.0], '0.88': [480.0, 544.0], '0.94': [480.0, 512.0],
+    '1.0': [512.0, 512.0], '1.07': [512.0, 480.0], '1.13': [544.0, 480.0], '1.21': [544.0, 448.0],
+    '1.29': [576.0, 448.0], '1.38': [576.0, 416.0], '1.46': [608.0, 416.0], '1.67': [640.0, 384.0],
+    '1.75': [672.0, 384.0], '2.0': [704.0, 352.0], '2.09': [736.0, 352.0], '2.4': [768.0, 320.0],
+    '2.5': [800.0, 320.0], '2.89': [832.0, 288.0], '3.0': [864.0, 288.0], '3.11': [896.0, 288.0],
+    '3.62': [928.0, 256.0], '3.75': [960.0, 256.0], '3.88': [992.0, 256.0], '4.0': [1024.0, 256.0]
+}
+ASPECT_RATIO_RANDOM_CROP_512 = {
+    '0.42': [320.0, 768.0], '0.5': [352.0, 704.0], 
+    '0.57': [384.0, 672.0], '0.68': [416.0, 608.0], '0.78': [448.0, 576.0], '0.88': [480.0, 544.0], 
+    '0.94': [480.0, 512.0], '1.0': [512.0, 512.0], '1.07': [512.0, 480.0], 
+    '1.13': [544.0, 480.0], '1.29': [576.0, 448.0], '1.46': [608.0, 416.0], '1.75': [672.0, 384.0], 
+    '2.0': [704.0, 352.0],  '2.4': [768.0, 320.0]
+}
+ASPECT_RATIO_RANDOM_CROP_PROB = [
+    1, 2,
+    4, 4, 4, 4,
+    8, 8, 8,
+    4, 4, 4, 4,
+    2, 1
+]
+ASPECT_RATIO_RANDOM_CROP_PROB = np.array(ASPECT_RATIO_RANDOM_CROP_PROB) / sum(ASPECT_RATIO_RANDOM_CROP_PROB)
+
+def get_closest_ratio(height: float, width: float, ratios: dict = ASPECT_RATIO_512):
+    aspect_ratio = height / width
+    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 color_transfer(sc, dc):
+    """
+    Transfer color distribution from of sc, referred to dc.
+
+    Args:
+        sc (numpy.ndarray): input image to be transfered.
+        dc (numpy.ndarray): reference image
+
+    Returns:
+        numpy.ndarray: Transferred color distribution on the sc.
+    """
+
+    def get_mean_and_std(img):
+        x_mean, x_std = cv2.meanStdDev(img)
+        x_mean = np.hstack(np.around(x_mean, 2))
+        x_std = np.hstack(np.around(x_std, 2))
+        return x_mean, x_std
+
+    sc = cv2.cvtColor(sc, cv2.COLOR_RGB2LAB)
+    s_mean, s_std = get_mean_and_std(sc)
+    dc = cv2.cvtColor(dc, cv2.COLOR_RGB2LAB)
+    t_mean, t_std = get_mean_and_std(dc)
+    img_n = ((sc - s_mean) * (t_std / s_std)) + t_mean
+    np.putmask(img_n, img_n > 255, 255)
+    np.putmask(img_n, img_n < 0, 0)
+    dst = cv2.cvtColor(cv2.convertScaleAbs(img_n), cv2.COLOR_LAB2RGB)
+    return dst
+
+def save_videos_grid(videos: torch.Tensor, path: str, rescale=False, n_rows=6, fps=12, imageio_backend=True, color_transfer_post_process=False):
+    videos = rearrange(videos, "b c t h w -> t b c h w")
+    outputs = []
+    for x in videos:
+        x = torchvision.utils.make_grid(x, nrow=n_rows)
+        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
+        if rescale:
+            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
+        x = (x * 255).numpy().astype(np.uint8)
+        outputs.append(Image.fromarray(x))
+
+    if color_transfer_post_process:
+        for i in range(1, len(outputs)):
+            outputs[i] = Image.fromarray(color_transfer(np.uint8(outputs[i]), np.uint8(outputs[0])))
+
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    if imageio_backend:
+        if path.endswith("mp4"):
+            imageio.mimsave(path, outputs, fps=fps)
+        else:
+            imageio.mimsave(path, outputs, duration=(1000 * 1/fps))
+    else:
+        if path.endswith("mp4"):
+            path = path.replace('.mp4', '.gif')
+        outputs[0].save(path, format='GIF', append_images=outputs, save_all=True, duration=100, loop=0)
+
+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):
+    if type(input_video_path) is str:
+        cap = cv2.VideoCapture(input_video_path)
+        input_video = []
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+            frame = cv2.resize(frame, (sample_size[1], sample_size[0]))
+            input_video.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
+        cap.release()
+    else:
+        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
+
+    input_video_mask = torch.zeros_like(input_video[:, :1])
+    input_video_mask[:, :, :] = 255
+
+    return  input_video, input_video_mask, None
\ No newline at end of file
diff --git a/nodes.py b/nodes.py
index 0c6ff19..849eb8b 100644
--- a/nodes.py
+++ b/nodes.py
@@ -3,11 +3,17 @@ import torch
 import folder_paths
 import comfy.model_management as mm
 from comfy.utils import ProgressBar
-from diffusers.schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
+from diffusers.schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler, DDIMScheduler, PNDMScheduler, DPMSolverMultistepScheduler, EulerDiscreteScheduler, EulerAncestralDiscreteScheduler
+
 from diffusers.models import AutoencoderKLCogVideoX, 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.utils import get_image_to_video_latent, ASPECT_RATIO_512, get_closest_ratio, to_pil
+from .cogvideox_fun.pipeline_cogvideox_inpaint import CogVideoX_Fun_Pipeline_Inpaint
+from PIL import Image
 
 import logging
 logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
@@ -24,6 +30,7 @@ class DownloadAndLoadCogVideoModel:
                         "THUDM/CogVideoX-2b",
                         "THUDM/CogVideoX-5b",
                         "bertjiazheng/KoolCogVideoX-5b",
+                        "kijai/CogVideoX-Fun-pruned"
                     ],
                 ),
 
@@ -50,10 +57,16 @@ class DownloadAndLoadCogVideoModel:
 
         dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[precision]
 
-        if "2b" in model:
+        if "Fun" in model:
+            base_path = os.path.join(folder_paths.models_dir, "CogVideoX_Fun", "CogVideoX-Fun-5b-InP")
+            if not os.path.exists(base_path):
+                base_path = os.path.join(folder_paths.models_dir, "CogVideo", "CogVideoX-Fun-5b-InP")
+            
+        elif "2b" in model:
             base_path = os.path.join(folder_paths.models_dir, "CogVideo", "CogVideo2B")
         elif "5b" in model:
             base_path = os.path.join(folder_paths.models_dir, "CogVideo", (model.split("/")[-1]))
+        
 
         if not os.path.exists(base_path):
             log.info(f"Downloading model to: {base_path}")
@@ -65,25 +78,36 @@ class DownloadAndLoadCogVideoModel:
                 local_dir=base_path,
                 local_dir_use_symlinks=False,
             )
+        
+        if "Fun" in model:
+            transformer = CogVideoXTransformer3DModelFun.from_pretrained(base_path, subfolder="transformer")
+        else:
+            transformer = CogVideoXTransformer3DModel.from_pretrained(base_path, subfolder="transformer")
+        
+        transformer = transformer.to(dtype).to(offload_device)
+        
         if fp8_transformer == "enabled" or fp8_transformer == "fastmode":
-            transformer = CogVideoXTransformer3DModel.from_pretrained(base_path, subfolder="transformer").to(offload_device)
             if "2b" in model:
                 for name, param in transformer.named_parameters():
                     if name != "pos_embedding":
                         param.data = param.data.to(torch.float8_e4m3fn)
             else:
                 transformer.to(torch.float8_e4m3fn)
-
+        
             if fp8_transformer == "fastmode":
                 from .fp8_optimization import convert_fp8_linear
                 convert_fp8_linear(transformer, dtype)
-        else:
-            transformer = CogVideoXTransformer3DModel.from_pretrained(base_path, subfolder="transformer").to(dtype).to(offload_device)
 
-        vae = AutoencoderKLCogVideoX.from_pretrained(base_path, subfolder="vae").to(dtype).to(offload_device)
+        if "Fun" in model:
+            vae = AutoencoderKLCogVideoXFun.from_pretrained(base_path, subfolder="vae").to(dtype).to(offload_device)
+        else:
+            vae = AutoencoderKLCogVideoX.from_pretrained(base_path, subfolder="vae").to(dtype).to(offload_device)
         scheduler = CogVideoXDDIMScheduler.from_pretrained(base_path, subfolder="scheduler")
 
-        pipe = CogVideoXPipeline(vae, transformer, scheduler)
+        if "Fun" in model:
+            pipe = CogVideoX_Fun_Pipeline_Inpaint(vae, transformer, scheduler)
+        else:
+            pipe = CogVideoXPipeline(vae, transformer, scheduler)
         if enable_sequential_cpu_offload:
             pipe.enable_sequential_cpu_offload()
 
@@ -92,7 +116,7 @@ class DownloadAndLoadCogVideoModel:
             pipe.transformer.to(memory_format=torch.channels_last)
             pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
         elif compile == "onediff":
-            from onediffx import compile_pipe, quantize_pipe
+            from onediffx import compile_pipe
             os.environ['NEXFORT_FX_FORCE_TRITON_SDPA'] = '1'
             
             pipe = compile_pipe(
@@ -280,6 +304,7 @@ class CogVideoSampler:
             "optional": {
                 "samples": ("LATENT", ),
                 "denoise_strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "image_cond_latents": ("LATENT", ),
             }
         }
 
@@ -288,7 +313,8 @@ class CogVideoSampler:
     FUNCTION = "process"
     CATEGORY = "CogVideoWrapper"
 
-    def process(self, pipeline, positive, negative, steps, cfg, seed, height, width, num_frames, scheduler, t_tile_length, t_tile_overlap, samples=None, denoise_strength=1.0):
+    def process(self, pipeline, positive, negative, steps, cfg, seed, height, width, num_frames, scheduler, t_tile_length, t_tile_overlap, samples=None, 
+                denoise_strength=1.0, image_cond_latents=None):
         mm.soft_empty_cache()
 
         assert t_tile_length > t_tile_overlap, "t_tile_length must be greater than t_tile_overlap"
@@ -328,6 +354,7 @@ class CogVideoSampler:
                 t_tile_overlap = t_tile_overlap,
                 guidance_scale=cfg,
                 latents=samples["samples"] if samples is not None else None,
+                image_cond_latents=image_cond_latents["samples"] if image_cond_latents is not None else None,
                 denoise_strength=denoise_strength,
                 prompt_embeds=positive.to(dtype).to(device),
                 negative_prompt_embeds=negative.to(dtype).to(device),
@@ -387,7 +414,7 @@ class CogVideoDecode:
         latents = latents.to(vae.dtype)
         latents = latents.permute(0, 2, 1, 3, 4)  # [batch_size, num_channels, num_frames, height, width]
         latents = 1 / vae.config.scaling_factor * latents
-
+       
         frames = vae.decode(latents).sample
         if not pipeline["cpu_offloading"]:
             vae.to(offload_device)
@@ -399,18 +426,127 @@ class CogVideoDecode:
 
         return (video,)
 
+class CogVideoXFunSampler:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "pipeline": ("COGVIDEOPIPE",),
+                "positive": ("CONDITIONING", ),
+                "negative": ("CONDITIONING", ),
+                "video_length": ("INT", {"default": 49, "min": 5, "max": 49, "step": 4}),
+                "base_resolution": (
+                    [ 
+                        512,
+                        768,
+                        960,
+                        1024,
+                    ], {"default": 768}
+                ),
+                "seed": ("INT", {"default": 43, "min": 0, "max": 0xffffffffffffffff}),
+                "steps": ("INT", {"default": 50, "min": 1, "max": 200, "step": 1}),
+                "cfg": ("FLOAT", {"default": 6.0, "min": 1.0, "max": 20.0, "step": 0.01}),
+                "scheduler": (
+                    [ 
+                        "Euler",
+                        "Euler A",
+                        "DPM++",
+                        "PNDM",
+                        "DDIM",
+                    ],
+                    {
+                        "default": 'DDIM'
+                    }
+                )
+            },
+            "optional":{
+                "start_img": ("IMAGE",),
+                "end_img": ("IMAGE",),
+            },
+        }
+    
+    RETURN_TYPES = ("COGVIDEOPIPE", "LATENT",)
+    RETURN_NAMES = ("cogvideo_pipe", "samples",)
+    FUNCTION = "process"
+    CATEGORY = "CogVideoWrapper"
+
+    def process(self, pipeline,  positive, negative, video_length, base_resolution, seed, steps, cfg, scheduler, start_img=None, end_img=None):
+        device = mm.get_torch_device()
+        offload_device = mm.unet_offload_device()
+        pipe = pipeline["pipe"]
+        dtype = pipeline["dtype"]
+
+        pipe.enable_model_cpu_offload()
+
+        mm.soft_empty_cache()
+
+        start_img = [to_pil(_start_img) for _start_img in start_img] if start_img is not None else None
+        end_img = [to_pil(_end_img) for _end_img in end_img] if end_img is not None else None
+        # Count most suitable height and width
+        aspect_ratio_sample_size    = {key : [x / 512 * base_resolution for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()}
+        original_width, original_height = start_img[0].size if type(start_img) is list else Image.open(start_img).size
+        closest_size, closest_ratio = get_closest_ratio(original_height, original_width, ratios=aspect_ratio_sample_size)
+        height, width = [int(x / 16) * 16 for x in closest_size]
+        
+        base_path = pipeline["base_path"]
+
+        # Load Sampler
+        if scheduler == "DPM++":
+            noise_scheduler = DPMSolverMultistepScheduler.from_pretrained(base_path, subfolder= 'scheduler')
+        elif scheduler == "Euler":
+            noise_scheduler = EulerDiscreteScheduler.from_pretrained(base_path, subfolder= 'scheduler')
+        elif scheduler == "Euler A":
+            noise_scheduler = EulerAncestralDiscreteScheduler.from_pretrained(base_path, subfolder= 'scheduler')
+        elif scheduler == "PNDM":
+            noise_scheduler = PNDMScheduler.from_pretrained(base_path, subfolder= 'scheduler')
+        elif scheduler == "DDIM":
+            noise_scheduler = DDIMScheduler.from_pretrained(base_path, subfolder= 'scheduler')
+        pipe.scheduler = noise_scheduler
+
+        #if not pipeline["cpu_offloading"]:
+        #    pipe.transformer.to(device)
+        generator= torch.Generator(device=device).manual_seed(seed)
+
+        autocastcondition = not pipeline["onediff"]
+        autocast_context = torch.autocast(mm.get_autocast_device(device)) if autocastcondition else nullcontext()
+        with autocast_context:
+            video_length = int((video_length - 1) // pipe.vae.config.temporal_compression_ratio * pipe.vae.config.temporal_compression_ratio) + 1 if video_length != 1 else 1
+            input_video, input_video_mask, clip_image = get_image_to_video_latent(start_img, end_img, video_length=video_length, sample_size=(height, width))
+
+            latents = pipe(
+                prompt_embeds=positive.to(dtype).to(device),
+                negative_prompt_embeds=negative.to(dtype).to(device),
+                num_frames = video_length,
+                height      = height,
+                width       = width,
+                generator   = generator,
+                guidance_scale = cfg,
+                num_inference_steps = steps,
+
+                video        = input_video,
+                mask_video   = input_video_mask,
+                comfyui_progressbar = True,
+            )
+        #if not pipeline["cpu_offloading"]:
+       #     pipe.transformer.to(offload_device)
+        mm.soft_empty_cache()
+        print(latents.shape)
+
+        return (pipeline, {"samples": latents})
 
 NODE_CLASS_MAPPINGS = {
     "DownloadAndLoadCogVideoModel": DownloadAndLoadCogVideoModel,
     "CogVideoSampler": CogVideoSampler,
     "CogVideoDecode": CogVideoDecode,
     "CogVideoTextEncode": CogVideoTextEncode,
-    "CogVideoImageEncode": CogVideoImageEncode
+    "CogVideoImageEncode": CogVideoImageEncode,
+    "CogVideoXFunSampler": CogVideoXFunSampler
 }
 NODE_DISPLAY_NAME_MAPPINGS = {
     "DownloadAndLoadCogVideoModel": "(Down)load CogVideo Model",
     "CogVideoSampler": "CogVideo Sampler",
     "CogVideoDecode": "CogVideo Decode",
     "CogVideoTextEncode": "CogVideo TextEncode",
-    "CogVideoImageEncode": "CogVideo ImageEncode"
+    "CogVideoImageEncode": "CogVideo ImageEncode",
+    "CogVideoXFunSampler": "CogVideoXFun Sampler"
     }
\ No newline at end of file
diff --git a/pipeline_cogvideox.py b/pipeline_cogvideox.py
index 09eeeec..2bae2f8 100644
--- a/pipeline_cogvideox.py
+++ b/pipeline_cogvideox.py
@@ -333,6 +333,7 @@ class CogVideoXPipeline(DiffusionPipeline):
         eta: float = 0.0,
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
         latents: Optional[torch.Tensor] = None,
+        image_cond_latents: Optional[torch.Tensor] = None,
         prompt_embeds: Optional[torch.Tensor] = None,
         negative_prompt_embeds: Optional[torch.Tensor] = None,
         device = torch.device("cuda"),
@@ -442,6 +443,20 @@ class CogVideoXPipeline(DiffusionPipeline):
             latents
         )
         latents = latents.to(self.transformer.dtype)
+
+        # 5.5.
+        if image_cond_latents is not None:
+            image_cond_latents = torch.cat(image_cond_latents, dim=0).to(self.transformer.dtype)#.permute(0, 2, 1, 3, 4)  # [B, F, C, H, W]
+
+            padding_shape = (
+                batch_size,
+                num_frames - 1,
+                latent_channels,
+                height // self.vae_scale_factor_spatial,
+                width // self.vae_scale_factor_spatial,
+            )
+            latent_padding = torch.zeros(padding_shape, device=device, dtype=self.transformer.dtype)
+            image_latents = torch.cat([image_latents, latent_padding], dim=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)
@@ -582,6 +597,10 @@ class CogVideoXPipeline(DiffusionPipeline):
                     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)
 
+                    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])