CogVideoX-5b-I2V support (#9418)

* draft Init

* draft

* vae encode image

* make style

* image latents preparation

* remove image encoder from conversion script

* fix minor bugs

* make pipeline work

* make style

* remove debug prints

* fix imports

* update example

* make fix-copies

* add fast tests

* fix import

* update vae

* update docs

* update image link

* apply suggestions from review

* apply suggestions from review

* add slow test

* make use of learned positional embeddings

* apply suggestions from review

* doc change

* Update convert_cogvideox_to_diffusers.py

* make style

* final changes

* make style

* fix tests

---------

Co-authored-by: Aryan <aryan@huggingface.co>

docs/source/en/api/loaders/single_file.md

@@ -23,6 +23,8 @@ The [`~loaders.FromSingleFileMixin.from_single_file`] method allows you to load:
 ## Supported pipelines
 
 - [`CogVideoXPipeline`]
+- [`CogVideoXImageToVideoPipeline`]
+- [`CogVideoXVideoToVideoPipeline`]
 - [`StableDiffusionPipeline`]
 - [`StableDiffusionImg2ImgPipeline`]
 - [`StableDiffusionInpaintPipeline`]

docs/source/en/api/pipelines/cogvideox.md

@@ -29,9 +29,12 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.m
 
 This pipeline was contributed by [zRzRzRzRzRzRzR](https://github.com/zRzRzRzRzRzRzR). The original codebase can be found [here](https://huggingface.co/THUDM). The original weights can be found under [hf.co/THUDM](https://huggingface.co/THUDM).
 
-There are two models available that can be used with the CogVideoX pipeline:
-- [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b)
-- [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b)
+There are two models available that can be used with the text-to-video and video-to-video CogVideoX pipelines:
+- [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b): The recommended dtype for running this model is `fp16`.
+- [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b): The recommended dtype for running this model is `bf16`.
+
+There is one model available that can be used with the image-to-video CogVideoX pipeline:
+- [`THUDM/CogVideoX-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-5b-I2V): The recommended dtype for running this model is `bf16`.
 
 ## Inference
 
@@ -41,10 +44,15 @@ First, load the pipeline:
 
 ```python
 import torch
-from diffusers import CogVideoXPipeline
-from diffusers.utils import export_to_video
+from diffusers import CogVideoXPipeline, CogVideoXImageToVideoPipeline
+from diffusers.utils import export_to_video,load_image
+pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b").to("cuda") # or "THUDM/CogVideoX-2b" 
+```
 
-pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b").to("cuda")
+If you are using the image-to-video pipeline, load it as follows:
+
+```python
+pipe = CogVideoXImageToVideoPipeline.from_pretrained("THUDM/CogVideoX-5b-I2V").to("cuda")
 ```
 
 Then change the memory layout of the pipelines `transformer` component to `torch.channels_last`:
@@ -53,7 +61,7 @@ Then change the memory layout of the pipelines `transformer` component to `torch
 pipe.transformer.to(memory_format=torch.channels_last)
 ```
 
-Finally, compile the components and run inference:
+Compile the components and run inference:
 
 ```python
 pipe.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=True)
@@ -63,7 +71,7 @@ prompt = "A panda, dressed in a small, red jacket and a tiny hat, sits on a wood
 video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
 ```
 
-The [benchmark](https://gist.github.com/a-r-r-o-w/5183d75e452a368fd17448fcc810bd3f) results on an 80GB A100 machine are:
+The [T2V benchmark](https://gist.github.com/a-r-r-o-w/5183d75e452a368fd17448fcc810bd3f) results on an 80GB A100 machine are:
 
 ```
 Without torch.compile(): Average inference time: 96.89 seconds.
@@ -98,6 +106,12 @@ It is also worth noting that torchao quantization is fully compatible with [torc
   - all
   - __call__
 
+## CogVideoXImageToVideoPipeline
+
+[[autodoc]] CogVideoXImageToVideoPipeline
+  - all
+  - __call__
+
 ## CogVideoXVideoToVideoPipeline
 
 [[autodoc]] CogVideoXVideoToVideoPipeline

scripts/convert_cogvideox_to_diffusers.py

@@ -4,7 +4,13 @@ from typing import Any, Dict
 import torch
 from transformers import T5EncoderModel, T5Tokenizer
 
-from diffusers import AutoencoderKLCogVideoX, CogVideoXDDIMScheduler, CogVideoXPipeline, CogVideoXTransformer3DModel
+from diffusers import (
+    AutoencoderKLCogVideoX,
+    CogVideoXDDIMScheduler,
+    CogVideoXImageToVideoPipeline,
+    CogVideoXPipeline,
+    CogVideoXTransformer3DModel,
+)
 
 
 def reassign_query_key_value_inplace(key: str, state_dict: Dict[str, Any]):
@@ -78,6 +84,7 @@ TRANSFORMER_KEYS_RENAME_DICT = {
     "mixins.final_layer.norm_final": "norm_out.norm",
     "mixins.final_layer.linear": "proj_out",
     "mixins.final_layer.adaLN_modulation.1": "norm_out.linear",
+    "mixins.pos_embed.pos_embedding": "patch_embed.pos_embedding",  # Specific to CogVideoX-5b-I2V
 }
 
 TRANSFORMER_SPECIAL_KEYS_REMAP = {
@@ -131,15 +138,18 @@ def convert_transformer(
     num_layers: int,
     num_attention_heads: int,
     use_rotary_positional_embeddings: bool,
+    i2v: bool,
     dtype: torch.dtype,
 ):
     PREFIX_KEY = "model.diffusion_model."
 
     original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
     transformer = CogVideoXTransformer3DModel(
+        in_channels=32 if i2v else 16,
         num_layers=num_layers,
         num_attention_heads=num_attention_heads,
         use_rotary_positional_embeddings=use_rotary_positional_embeddings,
+        use_learned_positional_embeddings=i2v,
     ).to(dtype=dtype)
 
     for key in list(original_state_dict.keys()):
@@ -153,7 +163,6 @@ def convert_transformer(
             if special_key not in key:
                 continue
             handler_fn_inplace(key, original_state_dict)
-
     transformer.load_state_dict(original_state_dict, strict=True)
     return transformer
 
@@ -205,6 +214,7 @@ def get_args():
     parser.add_argument("--scaling_factor", type=float, default=1.15258426, help="Scaling factor in the VAE")
     # For CogVideoX-2B, snr_shift_scale is 3.0. For 5B, it is 1.0
     parser.add_argument("--snr_shift_scale", type=float, default=3.0, help="Scaling factor in the VAE")
+    parser.add_argument("--i2v", action="store_true", default=False, help="Whether to save the model weights in fp16")
     return parser.parse_args()
 
 
@@ -225,6 +235,7 @@ if __name__ == "__main__":
             args.num_layers,
             args.num_attention_heads,
             args.use_rotary_positional_embeddings,
+            args.i2v,
             dtype,
         )
     if args.vae_ckpt_path is not None:
@@ -234,7 +245,7 @@ if __name__ == "__main__":
     tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
     text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
 
-    # Apparently, the conversion does not work any more without this :shrug:
+    # Apparently, the conversion does not work anymore without this :shrug:
     for param in text_encoder.parameters():
         param.data = param.data.contiguous()
 
@@ -252,9 +263,17 @@ if __name__ == "__main__":
             "timestep_spacing": "trailing",
         }
     )
+    if args.i2v:
+        pipeline_cls = CogVideoXImageToVideoPipeline
+    else:
+        pipeline_cls = CogVideoXPipeline
 
-    pipe = CogVideoXPipeline(
-        tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+    pipe = pipeline_cls(
+        tokenizer=tokenizer,
+        text_encoder=text_encoder,
+        vae=vae,
+        transformer=transformer,
+        scheduler=scheduler,
     )
 
     if args.fp16:
@@ -265,4 +284,7 @@ if __name__ == "__main__":
     # We don't use variant here because the model must be run in fp16 (2B) or bf16 (5B). It would be weird
     # for users to specify variant when the default is not fp32 and they want to run with the correct default (which
     # is either fp16/bf16 here).
-    pipe.save_pretrained(args.output_path, safe_serialization=True, push_to_hub=args.push_to_hub)
+
+    # This is necessary This is necessary for users with insufficient memory,
+    # such as those using Colab and notebooks, as it can save some memory used for model loading.
+    pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB", push_to_hub=args.push_to_hub)

src/diffusers/__init__.py

@@ -255,6 +255,7 @@ else:
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
             "CLIPImageProjection",
+            "CogVideoXImageToVideoPipeline",
             "CogVideoXPipeline",
             "CogVideoXVideoToVideoPipeline",
             "CycleDiffusionPipeline",
@@ -703,6 +704,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDMPipeline,
             AuraFlowPipeline,
             CLIPImageProjection,
+            CogVideoXImageToVideoPipeline,
             CogVideoXPipeline,
             CogVideoXVideoToVideoPipeline,
             CycleDiffusionPipeline,

src/diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -1089,8 +1089,10 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             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_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
@@ -1140,8 +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
@@ -1233,8 +1236,10 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         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_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
@@ -1309,8 +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

src/diffusers/models/embeddings.py

@@ -350,6 +350,7 @@ class CogVideoXPatchEmbed(nn.Module):
         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__()
 
@@ -363,15 +364,17 @@ class CogVideoXPatchEmbed(nn.Module):
         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:
+        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=False)
+            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
@@ -415,8 +418,15 @@ class CogVideoXPatchEmbed(nn.Module):
             [text_embeds, image_embeds], dim=1
         ).contiguous()  # [batch, seq_length + num_frames x height x width, channels]
 
-        if self.use_positional_embeddings:
+        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

src/diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -235,10 +235,18 @@ 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
 
+        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=patch_size,
@@ -254,6 +262,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
             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,
         )
         self.embedding_dropout = nn.Dropout(dropout)
 
@@ -465,8 +474,11 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
         hidden_states = self.proj_out(hidden_states)
 
         # 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:

src/diffusers/pipelines/__init__.py

@@ -138,7 +138,11 @@ else:
         "AudioLDM2UNet2DConditionModel",
     ]
     _import_structure["blip_diffusion"] = ["BlipDiffusionPipeline"]
-    _import_structure["cogvideo"] = ["CogVideoXPipeline", "CogVideoXVideoToVideoPipeline"]
+    _import_structure["cogvideo"] = [
+        "CogVideoXPipeline",
+        "CogVideoXImageToVideoPipeline",
+        "CogVideoXVideoToVideoPipeline",
+    ]
     _import_structure["controlnet"].extend(
         [
             "BlipDiffusionControlNetPipeline",
@@ -461,7 +465,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         )
         from .aura_flow import AuraFlowPipeline
         from .blip_diffusion import BlipDiffusionPipeline
-        from .cogvideo import CogVideoXPipeline, CogVideoXVideoToVideoPipeline
+        from .cogvideo import CogVideoXImageToVideoPipeline, CogVideoXPipeline, CogVideoXVideoToVideoPipeline
         from .controlnet import (
             BlipDiffusionControlNetPipeline,
             StableDiffusionControlNetImg2ImgPipeline,

src/diffusers/pipelines/cogvideo/__init__.py

@@ -23,6 +23,7 @@ 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:
@@ -34,6 +35,7 @@ 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:

src/diffusers/pipelines/cogvideo/pipeline_cogvideox_image2video.py

@@ -0,0 +1,827 @@
+# 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 typing import Callable, Dict, List, Optional, Tuple, Union
+
+import PIL
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PipelineImageInput
+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 (
+    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
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import CogVideoXImageToVideoPipeline
+        >>> from diffusers.utils import export_to_video, load_image
+
+        >>> pipe = CogVideoXImageToVideoPipeline.from_pretrained("THUDM/CogVideoX-5b-I2V", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+        ... )
+        >>> video = pipe(image, prompt, use_dynamic_cfg=True)
+        >>> export_to_video(video.frames[0], "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
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class CogVideoXImageToVideoPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for image-to-video generation using CogVideoX.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. CogVideoX uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
+            [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
+        tokenizer (`T5Tokenizer`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`CogVideoXTransformer3DModel`]):
+            A text conditioned `CogVideoXTransformer3DModel` to denoise the encoded video latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
+    """
+
+    _optional_components = []
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+    ]
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKLCogVideoX,
+        transformer: CogVideoXTransformer3DModel,
+        scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            vae=vae,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor_spatial = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+        self.vae_scale_factor_temporal = (
+            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    # Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline._get_t5_prompt_embeds
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(
+        self,
+        image: torch.Tensor,
+        batch_size: int = 1,
+        num_channels_latents: int = 16,
+        num_frames: int = 13,
+        height: int = 60,
+        width: int = 90,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.Tensor] = None,
+    ):
+        num_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        shape = (
+            batch_size,
+            num_frames,
+            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."
+            )
+
+        image = image.unsqueeze(2)  # [B, C, F, H, W]
+
+        if isinstance(generator, list):
+            image_latents = [
+                retrieve_latents(self.vae.encode(image[i].unsqueeze(0)), generator[i]) for i in range(batch_size)
+            ]
+        else:
+            image_latents = [retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator) for img in image]
+
+        image_latents = torch.cat(image_latents, dim=0).to(dtype).permute(0, 2, 1, 3, 4)  # [B, F, C, H, W]
+        image_latents = self.vae.config.scaling_factor * image_latents
+
+        padding_shape = (
+            batch_size,
+            num_frames - 1,
+            num_channels_latents,
+            height // self.vae_scale_factor_spatial,
+            width // self.vae_scale_factor_spatial,
+        )
+        latent_padding = torch.zeros(padding_shape, device=device, dtype=dtype)
+        image_latents = torch.cat([image_latents, latent_padding], dim=1)
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents, image_latents
+
+    # Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.decode_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
+        return frames
+
+    # Copied from diffusers.pipelines.animatediff.pipeline_animatediff_video2video.AnimateDiffVideoToVideoPipeline.get_timesteps
+    def get_timesteps(self, num_inference_steps, timesteps, 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 = timesteps[t_start * self.scheduler.order :]
+
+        return timesteps, num_inference_steps - t_start
+
+    # 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
+
+    def check_inputs(
+        self,
+        image,
+        prompt,
+        height,
+        width,
+        negative_prompt,
+        callback_on_step_end_tensor_inputs,
+        video=None,
+        latents=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        if (
+            not isinstance(image, torch.Tensor)
+            and not isinstance(image, PIL.Image.Image)
+            and not isinstance(image, list)
+        ):
+            raise ValueError(
+                "`image` has to be of type `torch.Tensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+                f" {type(image)}"
+            )
+
+        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}."
+                )
+
+        if video is not None and latents is not None:
+            raise ValueError("Only one of `video` or `latents` should be provided")
+
+    # Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.fuse_qkv_projections
+    def fuse_qkv_projections(self) -> None:
+        r"""Enables fused QKV projections."""
+        self.fusing_transformer = True
+        self.transformer.fuse_qkv_projections()
+
+    # Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.unfuse_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
+
+    # Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline._prepare_rotary_positional_embeddings
+    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,
+        )
+
+        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
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: PipelineImageInput,
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 480,
+        width: int = 720,
+        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 = "pil",
+        return_dict: bool = True,
+        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,
+    ) -> Union[CogVideoXPipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            image (`PipelineImageInput`):
+                The input video to condition the generation on. Must be an image, a list of images or a `torch.Tensor`.
+            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 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_output.CogVideoXPipelineOutput`] or `tuple`:
+            [`~pipelines.cogvideo.pipeline_output.CogVideoXPipelineOutput`] 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(
+            image,
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_on_step_end_tensor_inputs,
+            prompt_embeds,
+            negative_prompt_embeds,
+        )
+        self._guidance_scale = guidance_scale
+        self._interrupt = False
+
+        # 2. Default call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        self._num_timesteps = len(timesteps)
+
+        # 5. Prepare latents
+        image = self.video_processor.preprocess(image, height=height, width=width).to(
+            device, dtype=prompt_embeds.dtype
+        )
+
+        latent_channels = self.transformer.config.in_channels // 2
+        latents, image_latents = self.prepare_latents(
+            image,
+            batch_size * num_videos_per_prompt,
+            latent_channels,
+            num_frames,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 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)
+
+                latent_image_input = torch.cat([image_latents] * 2) if do_classifier_free_guidance else image_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])
+
+                # 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,
+                )[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 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:
+            return (video,)
+
+        return CogVideoXPipelineOutput(frames=video)

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -272,6 +272,21 @@ class CLIPImageProjection(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class CogVideoXImageToVideoPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class CogVideoXPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/cogvideo/test_cogvideox_image2video.py

@@ -0,0 +1,387 @@
+# Copyright 2024 The HuggingFace Team.
+#
+# 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 gc
+import inspect
+import unittest
+
+import numpy as np
+import torch
+from PIL import Image
+from transformers import AutoTokenizer, T5EncoderModel
+
+from diffusers import AutoencoderKLCogVideoX, CogVideoXImageToVideoPipeline, CogVideoXTransformer3DModel, DDIMScheduler
+from diffusers.utils import load_image
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    numpy_cosine_similarity_distance,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import (
+    PipelineTesterMixin,
+    check_qkv_fusion_matches_attn_procs_length,
+    check_qkv_fusion_processors_exist,
+    to_np,
+)
+
+
+enable_full_determinism()
+
+
+class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = CogVideoXImageToVideoPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS.union({"image"})
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    test_xformers_attention = False
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = CogVideoXTransformer3DModel(
+            # Product of num_attention_heads * attention_head_dim must be divisible by 16 for 3D positional embeddings
+            # But, since we are using tiny-random-t5 here, we need the internal dim of CogVideoXTransformer3DModel
+            # to be 32. The internal dim is product of num_attention_heads and attention_head_dim
+            # Note: The num_attention_heads and attention_head_dim is different from the T2V and I2V tests because
+            # attention_head_dim must be divisible by 16 for RoPE to work. We also need to maintain a product of 32 as
+            # detailed above.
+            num_attention_heads=2,
+            attention_head_dim=16,
+            in_channels=8,
+            out_channels=4,
+            time_embed_dim=2,
+            text_embed_dim=32,  # Must match with tiny-random-t5
+            num_layers=1,
+            sample_width=2,  # latent width: 2 -> final width: 16
+            sample_height=2,  # latent height: 2 -> final height: 16
+            sample_frames=9,  # latent frames: (9 - 1) / 4 + 1 = 3 -> final frames: 9
+            patch_size=2,
+            temporal_compression_ratio=4,
+            max_text_seq_length=16,
+            use_rotary_positional_embeddings=True,
+            use_learned_positional_embeddings=True,
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderKLCogVideoX(
+            in_channels=3,
+            out_channels=3,
+            down_block_types=(
+                "CogVideoXDownBlock3D",
+                "CogVideoXDownBlock3D",
+                "CogVideoXDownBlock3D",
+                "CogVideoXDownBlock3D",
+            ),
+            up_block_types=(
+                "CogVideoXUpBlock3D",
+                "CogVideoXUpBlock3D",
+                "CogVideoXUpBlock3D",
+                "CogVideoXUpBlock3D",
+            ),
+            block_out_channels=(8, 8, 8, 8),
+            latent_channels=4,
+            layers_per_block=1,
+            norm_num_groups=2,
+            temporal_compression_ratio=4,
+        )
+
+        torch.manual_seed(0)
+        scheduler = DDIMScheduler()
+        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        # Cannot reduce below 16 because convolution kernel becomes bigger than sample
+        # Cannot reduce below 32 because 3D RoPE errors out
+        image_height = 16
+        image_width = 16
+        image = Image.new("RGB", (image_width, image_height))
+        inputs = {
+            "image": image,
+            "prompt": "dance monkey",
+            "negative_prompt": "",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "height": image_height,
+            "width": image_width,
+            "num_frames": 8,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+        }
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+
+        self.assertEqual(generated_video.shape, (8, 3, 16, 16))
+        expected_video = torch.randn(8, 3, 16, 16)
+        max_diff = np.abs(generated_video - expected_video).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-3)
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    def test_vae_tiling(self, expected_diff_max: float = 0.3):
+        # Note(aryan): Investigate why this needs a bit higher tolerance
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        # The reason to modify it this way is because I2V Transformer limits the generation to resolutions.
+        # See the if-statement on "self.use_learned_positional_embeddings"
+        components["transformer"] = CogVideoXTransformer3DModel.from_config(
+            components["transformer"].config,
+            sample_height=16,
+            sample_width=16,
+        )
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_overlap_factor_height=1 / 12,
+            tile_overlap_factor_width=1 / 12,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )
+
+    def test_fused_qkv_projections(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        frames = pipe(**inputs).frames  # [B, F, C, H, W]
+        original_image_slice = frames[0, -2:, -1, -3:, -3:]
+
+        pipe.fuse_qkv_projections()
+        assert check_qkv_fusion_processors_exist(
+            pipe.transformer
+        ), "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
+        assert check_qkv_fusion_matches_attn_procs_length(
+            pipe.transformer, pipe.transformer.original_attn_processors
+        ), "Something wrong with the attention processors concerning the fused QKV projections."
+
+        inputs = self.get_dummy_inputs(device)
+        frames = pipe(**inputs).frames
+        image_slice_fused = frames[0, -2:, -1, -3:, -3:]
+
+        pipe.transformer.unfuse_qkv_projections()
+        inputs = self.get_dummy_inputs(device)
+        frames = pipe(**inputs).frames
+        image_slice_disabled = frames[0, -2:, -1, -3:, -3:]
+
+        assert np.allclose(
+            original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3
+        ), "Fusion of QKV projections shouldn't affect the outputs."
+        assert np.allclose(
+            image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3
+        ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
+        assert np.allclose(
+            original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
+        ), "Original outputs should match when fused QKV projections are disabled."
+
+
+@unittest.skip("The model 'THUDM/CogVideoX-5b-I2V' is not public yet.")
+@slow
+@require_torch_gpu
+class CogVideoXImageToVideoPipelineIntegrationTests(unittest.TestCase):
+    prompt = "A painting of a squirrel eating a burger."
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_cogvideox(self):
+        generator = torch.Generator("cpu").manual_seed(0)
+
+        pipe = CogVideoXImageToVideoPipeline.from_pretrained("THUDM/CogVideoX-5b-I2V", torch_dtype=torch.bfloat16)
+        pipe.enable_model_cpu_offload()
+
+        prompt = self.prompt
+        image = load_image(
+            "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
+        )
+
+        videos = pipe(
+            image=image,
+            prompt=prompt,
+            height=480,
+            width=720,
+            num_frames=16,
+            generator=generator,
+            num_inference_steps=2,
+            output_type="pt",
+        ).frames
+
+        video = videos[0]
+        expected_video = torch.randn(1, 16, 480, 720, 3).numpy()
+
+        max_diff = numpy_cosine_similarity_distance(video, expected_video)
+        assert max_diff < 1e-3, f"Max diff is too high. got {video}"