# 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 Dict, 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.modeling_outputs import AutoencoderKLOutput from diffusers.models.modeling_utils import ModelMixin from diffusers.models.autoencoders.vae import DecoderOutput, DiagonalGaussianDistribution from diffusers.models.autoencoders.autoencoder_kl_cogvideox import CogVideoXCausalConv3d, CogVideoXDownBlock3D, CogVideoXMidBlock3D # from diffusers.models.autoencoders.autoencoder_kl_cogvideox import CogVideoXEncoder3D, CogVideoXCausalConv3d, CogVideoXDownBlock3D, CogVideoXMidBlock3D logger = logging.get_logger(__name__) # pylint: disable=invalid-name 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, conv_cache: Optional[Dict[str, torch.Tensor]] = None, ) -> torch.Tensor: r"""The forward method of the `CogVideoXEncoder3D` class.""" new_conv_cache = {} conv_cache = conv_cache or {} hidden_states, new_conv_cache["conv_in"] = self.conv_in(sample, conv_cache=conv_cache.get("conv_in")) if torch.is_grad_enabled() and self.gradient_checkpointing: def create_custom_forward(module): def custom_forward(*inputs): return module(*inputs) return custom_forward # 1. Down for i, down_block in enumerate(self.down_blocks): conv_cache_key = f"down_block_{i}" hidden_states, new_conv_cache[conv_cache_key] = torch.utils.checkpoint.checkpoint( create_custom_forward(down_block), hidden_states, temb, None, conv_cache.get(conv_cache_key), use_reentrant=False ) # 2. Mid hidden_states, new_conv_cache["mid_block"] = torch.utils.checkpoint.checkpoint( create_custom_forward(self.mid_block), hidden_states, temb, None, conv_cache.get("mid_block"), use_reentrant=False ) else: # 1. Down for i, down_block in enumerate(self.down_blocks): conv_cache_key = f"down_block_{i}" hidden_states, new_conv_cache[conv_cache_key] = down_block( hidden_states, temb, None, conv_cache.get(conv_cache_key) ) # 2. Mid hidden_states, new_conv_cache["mid_block"] = self.mid_block( hidden_states, temb, None, conv_cache=conv_cache.get("mid_block") ) # 3. Post-process hidden_states = self.norm_out(hidden_states) hidden_states = self.conv_act(hidden_states) hidden_states, new_conv_cache["conv_out"] = self.conv_out(hidden_states, conv_cache=conv_cache.get("conv_out")) return hidden_states, new_conv_cache class ControlnetXsVaeEncoderCogVideoX(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, down_block_types: Tuple[str] = ( "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", ), 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, ): 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.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 self.num_sample_frames_batch_size = 8 # 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): module.gradient_checkpointing = value 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 def _encode(self, x: torch.Tensor) -> torch.Tensor: batch_size, num_channels, num_frames, height, width = x.shape if self.use_tiling and (width > self.tile_sample_min_width or height > self.tile_sample_min_height): return self.tiled_encode(x) frame_batch_size = self.num_sample_frames_batch_size # Note: We expect the number of frames to be either `1` or `frame_batch_size * k` or `frame_batch_size * k + 1` for some k. # As the extra single frame is handled inside the loop, it is not required to round up here. num_batches = max(num_frames // frame_batch_size, 1) conv_cache = None enc = [] for i in range(num_batches): 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 x_intermediate = x[:, :, start_frame:end_frame] x_intermediate, conv_cache = self.encoder(x_intermediate, conv_cache=conv_cache) # if self.quant_conv is not None: # x_intermediate = self.quant_conv(x_intermediate) enc.append(x_intermediate) enc = torch.cat(enc, dim=2) return enc @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 videos. If `return_dict` is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned. """ if self.use_slicing and x.shape[0] > 1: encoded_slices = [self._encode(x_slice) for x_slice in x.split(1)] h = torch.cat(encoded_slices) else: h = self._encode(x) posterior = DiagonalGaussianDistribution(h) if not return_dict: return (posterior,) return AutoencoderKLOutput(latent_dist=posterior) 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_encode(self, x: torch.Tensor) -> torch.Tensor: r"""Encode a batch of images using a tiled encoder. When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the output, but they should be much less noticeable. Args: x (`torch.Tensor`): Input batch of videos. Returns: `torch.Tensor`: The latent representation of the encoded videos. """ # For a rough memory estimate, take a look at the `tiled_decode` method. batch_size, num_channels, num_frames, height, width = x.shape overlap_height = int(self.tile_sample_min_height * (1 - self.tile_overlap_factor_height)) overlap_width = int(self.tile_sample_min_width * (1 - self.tile_overlap_factor_width)) blend_extent_height = int(self.tile_latent_min_height * self.tile_overlap_factor_height) blend_extent_width = int(self.tile_latent_min_width * self.tile_overlap_factor_width) row_limit_height = self.tile_latent_min_height - blend_extent_height row_limit_width = self.tile_latent_min_width - blend_extent_width frame_batch_size = self.num_sample_frames_batch_size # Split x into overlapping tiles and encode 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): # Note: We expect the number of frames to be either `1` or `frame_batch_size * k` or `frame_batch_size * k + 1` for some k. # As the extra single frame is handled inside the loop, it is not required to round up here. num_batches = max(num_frames // frame_batch_size, 1) conv_cache = None time = [] for k in range(num_batches): 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 = x[ :, :, start_frame:end_frame, i : i + self.tile_sample_min_height, j : j + self.tile_sample_min_width, ] tile, conv_cache = self.encoder(tile, conv_cache=conv_cache) # if self.quant_conv is not None: # tile = self.quant_conv(tile) time.append(tile) 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)) enc = torch.cat(result_rows, dim=3) return enc 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() return z