diffusers 0.30.2__py3-none-any.whl → 0.30.3__py3-none-any.whl

diffusers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.30.2"
+__version__ = "0.30.3"
 
 from typing import TYPE_CHECKING
 
@@ -252,7 +252,9 @@ else:
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
             "CLIPImageProjection",
+            "CogVideoXImageToVideoPipeline",
             "CogVideoXPipeline",
+            "CogVideoXVideoToVideoPipeline",
             "CycleDiffusionPipeline",
             "FluxPipeline",
             "HunyuanDiTControlNetPipeline",
@@ -691,7 +693,9 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDMPipeline,
             AuraFlowPipeline,
             CLIPImageProjection,
+            CogVideoXImageToVideoPipeline,
             CogVideoXPipeline,
+            CogVideoXVideoToVideoPipeline,
             CycleDiffusionPipeline,
             FluxPipeline,
             HunyuanDiTControlNetPipeline,

diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -999,6 +999,7 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         # 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
@@ -1081,6 +1082,31 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         """
         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.
+        num_batches = num_frames // frame_batch_size if num_frames > 1 else 1
+        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 = self.encoder(x_intermediate)
+            if self.quant_conv is not None:
+                x_intermediate = self.quant_conv(x_intermediate)
+            enc.append(x_intermediate)
+
+        self._clear_fake_context_parallel_cache()
+        enc = torch.cat(enc, dim=2)
+
+        return enc
+
     @apply_forward_hook
     def encode(
         self, x: torch.Tensor, return_dict: bool = True
@@ -1094,13 +1120,17 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
                 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
+                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.
         """
-        h = self.encoder(x)
-        if self.quant_conv is not None:
-            h = self.quant_conv(h)
+        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)
@@ -1112,8 +1142,9 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             return self.tiled_decode(z, return_dict=return_dict)
 
         frame_batch_size = self.num_latent_frames_batch_size
+        num_batches = num_frames // frame_batch_size
         dec = []
-        for i in range(num_frames // frame_batch_size):
+        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
@@ -1172,6 +1203,77 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             )
         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.
+                num_batches = num_frames // frame_batch_size if num_frames > 1 else 1
+                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 = self.encoder(tile)
+                    if self.quant_conv is not None:
+                        tile = self.quant_conv(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))
+
+        enc = torch.cat(result_rows, dim=3)
+        return enc
+
     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.
@@ -1212,8 +1314,9 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         for i in range(0, height, overlap_height):
             row = []
             for j in range(0, width, overlap_width):
+                num_batches = num_frames // frame_batch_size
                 time = []
-                for k in range(num_frames // frame_batch_size):
+                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

diffusers/models/embeddings.py

@@ -342,15 +342,61 @@ class CogVideoXPatchEmbed(nn.Module):
         embed_dim: int = 1920,
         text_embed_dim: int = 4096,
         bias: bool = True,
+        sample_width: int = 90,
+        sample_height: int = 60,
+        sample_frames: int = 49,
+        temporal_compression_ratio: int = 4,
+        max_text_seq_length: int = 226,
+        spatial_interpolation_scale: float = 1.875,
+        temporal_interpolation_scale: float = 1.0,
+        use_positional_embeddings: bool = True,
+        use_learned_positional_embeddings: bool = True,
     ) -> None:
         super().__init__()
+
         self.patch_size = patch_size
+        self.embed_dim = embed_dim
+        self.sample_height = sample_height
+        self.sample_width = sample_width
+        self.sample_frames = sample_frames
+        self.temporal_compression_ratio = temporal_compression_ratio
+        self.max_text_seq_length = max_text_seq_length
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.use_positional_embeddings = use_positional_embeddings
+        self.use_learned_positional_embeddings = use_learned_positional_embeddings
 
         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)
 
+        if use_positional_embeddings or use_learned_positional_embeddings:
+            persistent = use_learned_positional_embeddings
+            pos_embedding = self._get_positional_embeddings(sample_height, sample_width, sample_frames)
+            self.register_buffer("pos_embedding", pos_embedding, persistent=persistent)
+
+    def _get_positional_embeddings(self, sample_height: int, sample_width: int, sample_frames: int) -> torch.Tensor:
+        post_patch_height = sample_height // self.patch_size
+        post_patch_width = sample_width // self.patch_size
+        post_time_compression_frames = (sample_frames - 1) // self.temporal_compression_ratio + 1
+        num_patches = post_patch_height * post_patch_width * post_time_compression_frames
+
+        pos_embedding = get_3d_sincos_pos_embed(
+            self.embed_dim,
+            (post_patch_width, post_patch_height),
+            post_time_compression_frames,
+            self.spatial_interpolation_scale,
+            self.temporal_interpolation_scale,
+        )
+        pos_embedding = torch.from_numpy(pos_embedding).flatten(0, 1)
+        joint_pos_embedding = torch.zeros(
+            1, self.max_text_seq_length + num_patches, self.embed_dim, requires_grad=False
+        )
+        joint_pos_embedding.data[:, self.max_text_seq_length :].copy_(pos_embedding)
+
+        return joint_pos_embedding
+
     def forward(self, text_embeds: torch.Tensor, image_embeds: torch.Tensor):
         r"""
         Args:
@@ -371,6 +417,28 @@ class CogVideoXPatchEmbed(nn.Module):
         embeds = torch.cat(
             [text_embeds, image_embeds], dim=1
         ).contiguous()  # [batch, seq_length + num_frames x height x width, channels]
+
+        if self.use_positional_embeddings or self.use_learned_positional_embeddings:
+            if self.use_learned_positional_embeddings and (self.sample_width != width or self.sample_height != height):
+                raise ValueError(
+                    "It is currently not possible to generate videos at a different resolution that the defaults. This should only be the case with 'THUDM/CogVideoX-5b-I2V'."
+                    "If you think this is incorrect, please open an issue at https://github.com/huggingface/diffusers/issues."
+                )
+
+            pre_time_compression_frames = (num_frames - 1) * self.temporal_compression_ratio + 1
+
+            if (
+                self.sample_height != height
+                or self.sample_width != width
+                or self.sample_frames != pre_time_compression_frames
+            ):
+                pos_embedding = self._get_positional_embeddings(height, width, pre_time_compression_frames)
+                pos_embedding = pos_embedding.to(embeds.device, dtype=embeds.dtype)
+            else:
+                pos_embedding = self.pos_embedding
+
+            embeds = embeds + pos_embedding
+
         return embeds
 
 

diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -23,7 +23,7 @@ from ...utils import is_torch_version, logging
 from ...utils.torch_utils import maybe_allow_in_graph
 from ..attention import Attention, FeedForward
 from ..attention_processor import AttentionProcessor, CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
-from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps, get_3d_sincos_pos_embed
+from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
 from ..normalization import AdaLayerNorm, CogVideoXLayerNormZero
@@ -235,37 +235,42 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
         spatial_interpolation_scale: float = 1.875,
         temporal_interpolation_scale: float = 1.0,
         use_rotary_positional_embeddings: bool = False,
+        use_learned_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
+        if not use_rotary_positional_embeddings and use_learned_positional_embeddings:
+            raise ValueError(
+                "There are no CogVideoX checkpoints available with disable rotary embeddings and learned positional "
+                "embeddings. If you're using a custom model and/or believe this should be supported, please open an "
+                "issue at https://github.com/huggingface/diffusers/issues."
+            )
 
         # 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,
+        self.patch_embed = CogVideoXPatchEmbed(
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=inner_dim,
+            text_embed_dim=text_embed_dim,
+            bias=True,
+            sample_width=sample_width,
+            sample_height=sample_height,
+            sample_frames=sample_frames,
+            temporal_compression_ratio=temporal_compression_ratio,
+            max_text_seq_length=max_text_seq_length,
+            spatial_interpolation_scale=spatial_interpolation_scale,
+            temporal_interpolation_scale=temporal_interpolation_scale,
+            use_positional_embeddings=not use_rotary_positional_embeddings,
+            use_learned_positional_embeddings=use_learned_positional_embeddings,
         )
-        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)
+        self.embedding_dropout = nn.Dropout(dropout)
 
-        # 3. Time embeddings
+        # 2. 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
+        # 3. Define spatio-temporal transformers blocks
         self.transformer_blocks = nn.ModuleList(
             [
                 CogVideoXBlock(
@@ -284,7 +289,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
         )
         self.norm_final = nn.LayerNorm(inner_dim, norm_eps, norm_elementwise_affine)
 
-        # 5. Output blocks
+        # 4. Output blocks
         self.norm_out = AdaLayerNorm(
             embedding_dim=time_embed_dim,
             output_dim=2 * inner_dim,
@@ -422,20 +427,13 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
 
         # 2. Patch embedding
         hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
+        hidden_states = self.embedding_dropout(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]
-            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
+        # 3. Transformer blocks
         for i, block in enumerate(self.transformer_blocks):
             if self.training and self.gradient_checkpointing:
 
@@ -471,13 +469,16 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
             hidden_states = self.norm_final(hidden_states)
             hidden_states = hidden_states[:, text_seq_length:]
 
-        # 5. Final block
+        # 4. Final block
         hidden_states = self.norm_out(hidden_states, temb=emb)
         hidden_states = self.proj_out(hidden_states)
 
-        # 6. Unpatchify
+        # 5. Unpatchify
+        # Note: we use `-1` instead of `channels`:
+        #   - It is okay to `channels` use for CogVideoX-2b and CogVideoX-5b (number of input channels is equal to output channels)
+        #   - However, for CogVideoX-5b-I2V also takes concatenated input image latents (number of input channels is twice the output channels)
         p = self.config.patch_size
-        output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, channels, p, p)
+        output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
         output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
 
         if not return_dict:

diffusers/pipelines/__init__.py

@@ -132,7 +132,11 @@ else:
         "AudioLDM2UNet2DConditionModel",
     ]
     _import_structure["blip_diffusion"] = ["BlipDiffusionPipeline"]
-    _import_structure["cogvideo"] = ["CogVideoXPipeline"]
+    _import_structure["cogvideo"] = [
+        "CogVideoXPipeline",
+        "CogVideoXImageToVideoPipeline",
+        "CogVideoXVideoToVideoPipeline",
+    ]
     _import_structure["controlnet"].extend(
         [
             "BlipDiffusionControlNetPipeline",
@@ -452,7 +456,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         )
         from .aura_flow import AuraFlowPipeline
         from .blip_diffusion import BlipDiffusionPipeline
-        from .cogvideo import CogVideoXPipeline
+        from .cogvideo import CogVideoXImageToVideoPipeline, CogVideoXPipeline, CogVideoXVideoToVideoPipeline
         from .controlnet import (
             BlipDiffusionControlNetPipeline,
             StableDiffusionControlNetImg2ImgPipeline,

diffusers/pipelines/cogvideo/__init__.py

@@ -23,6 +23,8 @@ except OptionalDependencyNotAvailable:
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
     _import_structure["pipeline_cogvideox"] = ["CogVideoXPipeline"]
+    _import_structure["pipeline_cogvideox_image2video"] = ["CogVideoXImageToVideoPipeline"]
+    _import_structure["pipeline_cogvideox_video2video"] = ["CogVideoXVideoToVideoPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -33,6 +35,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
         from .pipeline_cogvideox import CogVideoXPipeline
+        from .pipeline_cogvideox_image2video import CogVideoXImageToVideoPipeline
+        from .pipeline_cogvideox_video2video import CogVideoXVideoToVideoPipeline
 
 else:
     import sys

diffusers/pipelines/cogvideo/pipeline_cogvideox.py

@@ -15,7 +15,6 @@
 
 import inspect
 import math
-from dataclasses import dataclass
 from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import torch
@@ -26,9 +25,10 @@ from ...models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
 from ...models.embeddings import get_3d_rotary_pos_embed
 from ...pipelines.pipeline_utils import DiffusionPipeline
 from ...schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
-from ...utils import BaseOutput, logging, replace_example_docstring
+from ...utils import logging, replace_example_docstring
 from ...utils.torch_utils import randn_tensor
 from ...video_processor import VideoProcessor
+from .pipeline_output import CogVideoXPipelineOutput
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -136,21 +136,6 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-@dataclass
-class CogVideoXPipelineOutput(BaseOutput):
-    r"""
-    Output class for CogVideo pipelines.
-
-    Args:
-        frames (`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)`.
-    """
-
-    frames: torch.Tensor
-
-
 class CogVideoXPipeline(DiffusionPipeline):
     r"""
     Pipeline for text-to-video generation using CogVideoX.