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",
         }
     )
-
-    pipe = CogVideoXPipeline(
-        tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+    if args.i2v:
+        pipeline_cls = CogVideoXImageToVideoPipeline
+    else:
+        pipeline_cls = CogVideoXPipeline
+
+    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/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

+# 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}"