Add CogVideoX text-to-video generation model (#9082)

* add CogVideoX

---------
Co-authored-by: Aryan <aryan@huggingface.co>
Co-authored-by: sayakpaul <spsayakpaul@gmail.com>
Co-authored-by: Aryan <contact.aryanvs@gmail.com>
Co-authored-by: yiyixuxu <yixu310@gmail.com>
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

docs/source/en/_toctree.yml

@@ -239,6 +239,8 @@
       title: VQModel
     - local: api/models/autoencoderkl
       title: AutoencoderKL
+    - local: api/models/autoencoderkl_cogvideox
+      title: AutoencoderKLCogVideoX
     - local: api/models/asymmetricautoencoderkl
       title: AsymmetricAutoencoderKL
     - local: api/models/stable_cascade_unet
@@ -263,6 +265,8 @@
       title: FluxTransformer2DModel
     - local: api/models/latte_transformer3d
       title: LatteTransformer3DModel
+    - local: api/models/cogvideox_transformer3d
+      title: CogVideoXTransformer3DModel
     - local: api/models/lumina_nextdit2d
       title: LuminaNextDiT2DModel
     - local: api/models/transformer_temporal
@@ -302,6 +306,8 @@
       title: AutoPipeline
     - local: api/pipelines/blip_diffusion
       title: BLIP-Diffusion
+    - local: api/pipelines/cogvideox
+      title: CogVideoX
     - local: api/pipelines/consistency_models
       title: Consistency Models
     - local: api/pipelines/controlnet

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

@@ -22,6 +22,7 @@ The [`~loaders.FromSingleFileMixin.from_single_file`] method allows you to load:
 
 ## Supported pipelines
 
+- [`CogVideoXPipeline`]
 - [`StableDiffusionPipeline`]
 - [`StableDiffusionImg2ImgPipeline`]
 - [`StableDiffusionInpaintPipeline`]
@@ -49,6 +50,7 @@ The [`~loaders.FromSingleFileMixin.from_single_file`] method allows you to load:
 - [`UNet2DConditionModel`]
 - [`StableCascadeUNet`]
 - [`AutoencoderKL`]
+- [`AutoencoderKLCogVideoX`]
 - [`ControlNetModel`]
 - [`SD3Transformer2DModel`]
 - [`FluxTransformer2DModel`]

docs/source/en/api/models/autoencoderkl_cogvideox.md

+<!--Copyright 2024 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. -->
+
+# AutoencoderKLCogVideoX
+
+The 3D variational autoencoder (VAE) model with KL loss used in [CogVideoX](https://github.com/THUDM/CogVideo) was introduced in [CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://github.com/THUDM/CogVideo/blob/main/resources/CogVideoX.pdf) by Tsinghua University & ZhipuAI.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLCogVideoX
+
+vae = AutoencoderKLCogVideoX.from_pretrained("THUDM/CogVideoX-2b", subfolder="vae", torch_dtype=torch.float16).to("cuda")
+```
+
+## AutoencoderKLCogVideoX
+
+[[autodoc]] AutoencoderKLCogVideoX
+    - decode
+    - encode
+    - all
+
+## AutoencoderKLOutput
+
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput

docs/source/en/api/models/cogvideox_transformer3d.md

+<!--Copyright 2024 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. -->
+
+# CogVideoXTransformer3DModel
+
+A Diffusion Transformer model for 3D data from [CogVideoX](https://github.com/THUDM/CogVideo) was introduced in [CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://github.com/THUDM/CogVideo/blob/main/resources/CogVideoX.pdf) by Tsinghua University & ZhipuAI.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import CogVideoXTransformer3DModel
+
+vae = CogVideoXTransformer3DModel.from_pretrained("THUDM/CogVideoX-2b", subfolder="transformer", torch_dtype=torch.float16).to("cuda")
+```
+
+## CogVideoXTransformer3DModel
+
+[[autodoc]] CogVideoXTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

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

+<!--Copyright 2024 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.
+-->
+
+# CogVideoX
+
+<!-- TODO: update paper with ArXiv link when ready. -->
+
+[CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://github.com/THUDM/CogVideo/blob/main/resources/CogVideoX.pdf) from Tsinghua University & ZhipuAI.
+
+The abstract from the paper is:
+
+*We introduce CogVideoX, a large-scale diffusion transformer model designed for generating videos based on text prompts. To efficently model video data, we propose to levearge a 3D Variational Autoencoder (VAE) to compresses videos along both spatial and temporal dimensions. To improve the text-video alignment, we propose an expert transformer with the expert adaptive LayerNorm to facilitate the deep fusion between the two modalities. By employing a progressive training technique, CogVideoX is adept at producing coherent, long-duration videos characterized by significant motion. In addition, we develop an effectively text-video data processing pipeline that includes various data preprocessing strategies and a video captioning method. It significantly helps enhance the performance of CogVideoX, improving both generation quality and semantic alignment. Results show that CogVideoX demonstrates state-of-the-art performance across both multiple machine metrics and human evaluations. The model weight of CogVideoX-2B is publicly available at https://github.com/THUDM/CogVideo.*
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.md) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading.md#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+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).
+
+## Inference
+
+Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
+
+First, load the pipeline:
+
+```python
+import torch
+from diffusers import CogVideoXPipeline
+from diffusers.utils import export_to_video
+
+pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b").to("cuda")
+prompt = (
+    "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
+    "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other "
+    "pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, "
+    "casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. "
+    "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical "
+    "atmosphere of this unique musical performance."
+)
+video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
+export_to_video(video, "output.mp4", fps=8)
+```
+
+Then change the memory layout of the pipelines `transformer` and `vae` components to `torch.channels-last`:
+
+```python
+pipeline.transformer.to(memory_format=torch.channels_last)
+pipeline.vae.to(memory_format=torch.channels_last)
+```
+
+Finally, compile the components and run inference:
+
+```python
+pipeline.transformer = torch.compile(pipeline.transformer)
+pipeline.vae.decode = torch.compile(pipeline.vae.decode)
+
+# CogVideoX works very well with long and well-described prompts
+prompt = "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance."
+video = pipeline(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
+```
+
+The [benchmark](TODO: link) results on an 80GB A100 machine are:
+
+```
+Without torch.compile(): Average inference time: TODO seconds.
+With torch.compile(): Average inference time: TODO seconds.
+```
+
+## CogVideoXPipeline
+
+[[autodoc]] CogVideoXPipeline
+  - all
+  - __call__
+
+## CogVideoXPipelineOutput
+
+[[autodoc]] pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput

scripts/convert_cogvideox_to_diffusers.py

+import argparse
+from typing import Any, Dict
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers import AutoencoderKLCogVideoX, CogVideoXDDIMScheduler, CogVideoXPipeline, CogVideoXTransformer3DModel
+
+
+def reassign_query_key_value_inplace(key: str, state_dict: Dict[str, Any]):
+    to_q_key = key.replace("query_key_value", "to_q")
+    to_k_key = key.replace("query_key_value", "to_k")
+    to_v_key = key.replace("query_key_value", "to_v")
+    to_q, to_k, to_v = torch.chunk(state_dict[key], chunks=3, dim=0)
+    state_dict[to_q_key] = to_q
+    state_dict[to_k_key] = to_k
+    state_dict[to_v_key] = to_v
+    state_dict.pop(key)
+
+
+def reassign_query_key_layernorm_inplace(key: str, state_dict: Dict[str, Any]):
+    layer_id, weight_or_bias = key.split(".")[-2:]
+
+    if "query" in key:
+        new_key = f"transformer_blocks.{layer_id}.attn1.norm_q.{weight_or_bias}"
+    elif "key" in key:
+        new_key = f"transformer_blocks.{layer_id}.attn1.norm_k.{weight_or_bias}"
+
+    state_dict[new_key] = state_dict.pop(key)
+
+
+def reassign_adaln_norm_inplace(key: str, state_dict: Dict[str, Any]):
+    layer_id, _, weight_or_bias = key.split(".")[-3:]
+
+    weights_or_biases = state_dict[key].chunk(12, dim=0)
+    norm1_weights_or_biases = torch.cat(weights_or_biases[0:3] + weights_or_biases[6:9])
+    norm2_weights_or_biases = torch.cat(weights_or_biases[3:6] + weights_or_biases[9:12])
+
+    norm1_key = f"transformer_blocks.{layer_id}.norm1.linear.{weight_or_bias}"
+    state_dict[norm1_key] = norm1_weights_or_biases
+
+    norm2_key = f"transformer_blocks.{layer_id}.norm2.linear.{weight_or_bias}"
+    state_dict[norm2_key] = norm2_weights_or_biases
+
+    state_dict.pop(key)
+
+
+def remove_keys_inplace(key: str, state_dict: Dict[str, Any]):
+    state_dict.pop(key)
+
+
+def replace_up_keys_inplace(key: str, state_dict: Dict[str, Any]):
+    key_split = key.split(".")
+    layer_index = int(key_split[2])
+    replace_layer_index = 4 - 1 - layer_index
+
+    key_split[1] = "up_blocks"
+    key_split[2] = str(replace_layer_index)
+    new_key = ".".join(key_split)
+
+    state_dict[new_key] = state_dict.pop(key)
+
+
+TRANSFORMER_KEYS_RENAME_DICT = {
+    "transformer.final_layernorm": "norm_final",
+    "transformer": "transformer_blocks",
+    "attention": "attn1",
+    "mlp": "ff.net",
+    "dense_h_to_4h": "0.proj",
+    "dense_4h_to_h": "2",
+    ".layers": "",
+    "dense": "to_out.0",
+    "input_layernorm": "norm1.norm",
+    "post_attn1_layernorm": "norm2.norm",
+    "time_embed.0": "time_embedding.linear_1",
+    "time_embed.2": "time_embedding.linear_2",
+    "mixins.patch_embed": "patch_embed",
+    "mixins.final_layer.norm_final": "norm_out.norm",
+    "mixins.final_layer.linear": "proj_out",
+    "mixins.final_layer.adaLN_modulation.1": "norm_out.linear",
+}
+
+TRANSFORMER_SPECIAL_KEYS_REMAP = {
+    "query_key_value": reassign_query_key_value_inplace,
+    "query_layernorm_list": reassign_query_key_layernorm_inplace,
+    "key_layernorm_list": reassign_query_key_layernorm_inplace,
+    "adaln_layer.adaLN_modulations": reassign_adaln_norm_inplace,
+    "embed_tokens": remove_keys_inplace,
+}
+
+VAE_KEYS_RENAME_DICT = {
+    "block.": "resnets.",
+    "down.": "down_blocks.",
+    "downsample": "downsamplers.0",
+    "upsample": "upsamplers.0",
+    "nin_shortcut": "conv_shortcut",
+    "encoder.mid.block_1": "encoder.mid_block.resnets.0",
+    "encoder.mid.block_2": "encoder.mid_block.resnets.1",
+    "decoder.mid.block_1": "decoder.mid_block.resnets.0",
+    "decoder.mid.block_2": "decoder.mid_block.resnets.1",
+}
+
+VAE_SPECIAL_KEYS_REMAP = {
+    "loss": remove_keys_inplace,
+    "up.": replace_up_keys_inplace,
+}
+
+TOKENIZER_MAX_LENGTH = 226
+
+
+def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
+    state_dict = saved_dict
+    if "model" in saved_dict.keys():
+        state_dict = state_dict["model"]
+    if "module" in saved_dict.keys():
+        state_dict = state_dict["module"]
+    if "state_dict" in saved_dict.keys():
+        state_dict = state_dict["state_dict"]
+    return state_dict
+
+
+def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
+    state_dict[new_key] = state_dict.pop(old_key)
+
+
+def convert_transformer(ckpt_path: str):
+    PREFIX_KEY = "model.diffusion_model."
+
+    original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
+    transformer = CogVideoXTransformer3DModel()
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[len(PREFIX_KEY) :]
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_inplace(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            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
+
+
+def convert_vae(ckpt_path: str):
+    original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
+    vae = AutoencoderKLCogVideoX()
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in VAE_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_inplace(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in VAE_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    vae.load_state_dict(original_state_dict, strict=True)
+    return vae
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--transformer_ckpt_path", type=str, default=None, help="Path to original transformer checkpoint"
+    )
+    parser.add_argument("--vae_ckpt_path", type=str, default=None, help="Path to original vae checkpoint")
+    parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
+    parser.add_argument("--fp16", action="store_true", default=True, help="Whether to save the model weights in fp16")
+    parser.add_argument(
+        "--push_to_hub", action="store_true", default=False, help="Whether to push to HF Hub after saving"
+    )
+    parser.add_argument(
+        "--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    transformer = None
+    vae = None
+
+    if args.transformer_ckpt_path is not None:
+        transformer = convert_transformer(args.transformer_ckpt_path)
+    if args.vae_ckpt_path is not None:
+        vae = convert_vae(args.vae_ckpt_path)
+
+    text_encoder_id = "google/t5-v1_1-xxl"
+    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)
+
+    scheduler = CogVideoXDDIMScheduler.from_config(
+        {
+            "snr_shift_scale": 3.0,
+            "beta_end": 0.012,
+            "beta_schedule": "scaled_linear",
+            "beta_start": 0.00085,
+            "clip_sample": False,
+            "num_train_timesteps": 1000,
+            "prediction_type": "v_prediction",
+            "rescale_betas_zero_snr": True,
+            "set_alpha_to_one": True,
+            "timestep_spacing": "linspace",
+        }
+    )
+
+    pipe = CogVideoXPipeline(
+        tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+    )
+
+    if args.fp16:
+        pipe = pipe.to(dtype=torch.float16)
+
+    pipe.save_pretrained(args.output_path, safe_serialization=True, push_to_hub=args.push_to_hub)

src/diffusers/__init__.py

@@ -79,9 +79,11 @@ else:
             "AsymmetricAutoencoderKL",
             "AuraFlowTransformer2DModel",
             "AutoencoderKL",
+            "AutoencoderKLCogVideoX",
             "AutoencoderKLTemporalDecoder",
             "AutoencoderOobleck",
             "AutoencoderTiny",
+            "CogVideoXTransformer3DModel",
             "ConsistencyDecoderVAE",
             "ControlNetModel",
             "ControlNetXSAdapter",
@@ -155,6 +157,8 @@ else:
         [
             "AmusedScheduler",
             "CMStochasticIterativeScheduler",
+            "CogVideoXDDIMScheduler",
+            "CogVideoXDPMScheduler",
             "DDIMInverseScheduler",
             "DDIMParallelScheduler",
             "DDIMScheduler",
@@ -248,6 +252,7 @@ else:
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
             "CLIPImageProjection",
+            "CogVideoXPipeline",
             "CycleDiffusionPipeline",
             "FluxPipeline",
             "HunyuanDiTControlNetPipeline",
@@ -535,9 +540,11 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AsymmetricAutoencoderKL,
             AuraFlowTransformer2DModel,
             AutoencoderKL,
+            AutoencoderKLCogVideoX,
             AutoencoderKLTemporalDecoder,
             AutoencoderOobleck,
             AutoencoderTiny,
+            CogVideoXTransformer3DModel,
             ConsistencyDecoderVAE,
             ControlNetModel,
             ControlNetXSAdapter,
@@ -608,6 +615,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .schedulers import (
             AmusedScheduler,
             CMStochasticIterativeScheduler,
+            CogVideoXDDIMScheduler,
+            CogVideoXDPMScheduler,
             DDIMInverseScheduler,
             DDIMParallelScheduler,
             DDIMScheduler,
@@ -682,6 +691,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDMPipeline,
             AuraFlowPipeline,
             CLIPImageProjection,
+            CogVideoXPipeline,
             CycleDiffusionPipeline,
             FluxPipeline,
             HunyuanDiTControlNetPipeline,

src/diffusers/models/__init__.py

@@ -28,6 +28,7 @@ if is_torch_available():
     _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
     _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
     _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl_cogvideox"] = ["AutoencoderKLCogVideoX"]
     _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
     _import_structure["autoencoders.autoencoder_oobleck"] = ["AutoencoderOobleck"]
     _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
@@ -41,6 +42,7 @@ if is_torch_available():
     _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
     _import_structure["transformers.auraflow_transformer_2d"] = ["AuraFlowTransformer2DModel"]
+    _import_structure["transformers.cogvideox_transformer_3d"] = ["CogVideoXTransformer3DModel"]
     _import_structure["transformers.dit_transformer_2d"] = ["DiTTransformer2DModel"]
     _import_structure["transformers.dual_transformer_2d"] = ["DualTransformer2DModel"]
     _import_structure["transformers.hunyuan_transformer_2d"] = ["HunyuanDiT2DModel"]
@@ -77,6 +79,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .autoencoders import (
             AsymmetricAutoencoderKL,
             AutoencoderKL,
+            AutoencoderKLCogVideoX,
             AutoencoderKLTemporalDecoder,
             AutoencoderOobleck,
             AutoencoderTiny,
@@ -92,6 +95,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .modeling_utils import ModelMixin
         from .transformers import (
             AuraFlowTransformer2DModel,
+            CogVideoXTransformer3DModel,
             DiTTransformer2DModel,
             DualTransformer2DModel,
             FluxTransformer2DModel,

src/diffusers/models/autoencoders/__init__.py

 from .autoencoder_asym_kl import AsymmetricAutoencoderKL
 from .autoencoder_kl import AutoencoderKL
+from .autoencoder_kl_cogvideox import AutoencoderKLCogVideoX
 from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
 from .autoencoder_oobleck import AutoencoderOobleck
 from .autoencoder_tiny import AutoencoderTiny

src/diffusers/models/downsampling.py

@@ -285,6 +285,74 @@ class KDownsample2D(nn.Module):
         return F.conv2d(inputs, weight, stride=2)
 
 
+class CogVideoXDownsample3D(nn.Module):
+    # Todo: Wait for paper relase.
+    r"""
+    A 3D Downsampling layer using in [CogVideoX]() by Tsinghua University & ZhipuAI
+
+    Args:
+        in_channels (`int`):
+            Number of channels in the input image.
+        out_channels (`int`):
+            Number of channels produced by the convolution.
+        kernel_size (`int`, defaults to `3`):
+            Size of the convolving kernel.
+        stride (`int`, defaults to `2`):
+            Stride of the convolution.
+        padding (`int`, defaults to `0`):
+            Padding added to all four sides of the input.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to compress the time dimension.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int = 3,
+        stride: int = 2,
+        padding: int = 0,
+        compress_time: bool = False,
+    ):
+        super().__init__()
+
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.compress_time:
+            batch_size, channels, frames, height, width = x.shape
+
+            # (batch_size, channels, frames, height, width) -> (batch_size, height, width, channels, frames) -> (batch_size * height * width, channels, frames)
+            x = x.permute(0, 3, 4, 1, 2).reshape(batch_size * height * width, channels, frames)
+
+            if x.shape[-1] % 2 == 1:
+                x_first, x_rest = x[..., 0], x[..., 1:]
+                if x_rest.shape[-1] > 0:
+                    # (batch_size * height * width, channels, frames - 1) -> (batch_size * height * width, channels, (frames - 1) // 2)
+                    x_rest = F.avg_pool1d(x_rest, kernel_size=2, stride=2)
+
+                x = torch.cat([x_first[..., None], x_rest], dim=-1)
+                # (batch_size * height * width, channels, (frames // 2) + 1) -> (batch_size, height, width, channels, (frames // 2) + 1) -> (batch_size, channels, (frames // 2) + 1, height, width)
+                x = x.reshape(batch_size, height, width, channels, x.shape[-1]).permute(0, 3, 4, 1, 2)
+            else:
+                # (batch_size * height * width, channels, frames) -> (batch_size * height * width, channels, frames // 2)
+                x = F.avg_pool1d(x, kernel_size=2, stride=2)
+                # (batch_size * height * width, channels, frames // 2) -> (batch_size, height, width, channels, frames // 2) -> (batch_size, channels, frames // 2, height, width)
+                x = x.reshape(batch_size, height, width, channels, x.shape[-1]).permute(0, 3, 4, 1, 2)
+
+        # Pad the tensor
+        pad = (0, 1, 0, 1)
+        x = F.pad(x, pad, mode="constant", value=0)
+        batch_size, channels, frames, height, width = x.shape
+        # (batch_size, channels, frames, height, width) -> (batch_size, frames, channels, height, width) -> (batch_size * frames, channels, height, width)
+        x = x.permute(0, 2, 1, 3, 4).reshape(batch_size * frames, channels, height, width)
+        x = self.conv(x)
+        # (batch_size * frames, channels, height, width) -> (batch_size, frames, channels, height, width) -> (batch_size, channels, frames, height, width)
+        x = x.reshape(batch_size, frames, x.shape[1], x.shape[2], x.shape[3]).permute(0, 2, 1, 3, 4)
+        return x
+
+
 def downsample_2d(
     hidden_states: torch.Tensor,
     kernel: Optional[torch.Tensor] = None,

src/diffusers/models/embeddings.py

@@ -78,6 +78,53 @@ def get_timestep_embedding(
     return emb
 
 
+def get_3d_sincos_pos_embed(
+    embed_dim: int,
+    spatial_size: Union[int, Tuple[int, int]],
+    temporal_size: int,
+    spatial_interpolation_scale: float = 1.0,
+    temporal_interpolation_scale: float = 1.0,
+) -> np.ndarray:
+    r"""
+    Args:
+        embed_dim (`int`):
+        spatial_size (`int` or `Tuple[int, int]`):
+        temporal_size (`int`):
+        spatial_interpolation_scale (`float`, defaults to 1.0):
+        temporal_interpolation_scale (`float`, defaults to 1.0):
+    """
+    if embed_dim % 4 != 0:
+        raise ValueError("`embed_dim` must be divisible by 4")
+    if isinstance(spatial_size, int):
+        spatial_size = (spatial_size, spatial_size)
+
+    embed_dim_spatial = 3 * embed_dim // 4
+    embed_dim_temporal = embed_dim // 4
+
+    # 1. Spatial
+    grid_h = np.arange(spatial_size[1], dtype=np.float32) / spatial_interpolation_scale
+    grid_w = np.arange(spatial_size[0], dtype=np.float32) / spatial_interpolation_scale
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, spatial_size[1], spatial_size[0]])
+    pos_embed_spatial = get_2d_sincos_pos_embed_from_grid(embed_dim_spatial, grid)
+
+    # 2. Temporal
+    grid_t = np.arange(temporal_size, dtype=np.float32) / temporal_interpolation_scale
+    pos_embed_temporal = get_1d_sincos_pos_embed_from_grid(embed_dim_temporal, grid_t)
+
+    # 3. Concat
+    pos_embed_spatial = pos_embed_spatial[np.newaxis, :, :]
+    pos_embed_spatial = np.repeat(pos_embed_spatial, temporal_size, axis=0)  # [T, H*W, D // 4 * 3]
+
+    pos_embed_temporal = pos_embed_temporal[:, np.newaxis, :]
+    pos_embed_temporal = np.repeat(pos_embed_temporal, spatial_size[0] * spatial_size[1], axis=1)  # [T, H*W, D // 4]
+
+    pos_embed = np.concatenate([pos_embed_temporal, pos_embed_spatial], axis=-1)  # [T, H*W, D]
+    return pos_embed
+
+
 def get_2d_sincos_pos_embed(
     embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16
 ):
@@ -287,6 +334,46 @@ class LuminaPatchEmbed(nn.Module):
         )
 
 
+class CogVideoXPatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        embed_dim: int = 1920,
+        text_embed_dim: int = 4096,
+        bias: bool = True,
+    ) -> None:
+        super().__init__()
+        self.patch_size = patch_size
+
+        self.proj = nn.Conv2d(
+            in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
+        )
+        self.text_proj = nn.Linear(text_embed_dim, embed_dim)
+
+    def forward(self, text_embeds: torch.Tensor, image_embeds: torch.Tensor):
+        r"""
+        Args:
+            text_embeds (`torch.Tensor`):
+                Input text embeddings. Expected shape: (batch_size, seq_length, embedding_dim).
+            image_embeds (`torch.Tensor`):
+                Input image embeddings. Expected shape: (batch_size, num_frames, channels, height, width).
+        """
+        text_embeds = self.text_proj(text_embeds)
+
+        batch, num_frames, channels, height, width = image_embeds.shape
+        image_embeds = image_embeds.reshape(-1, channels, height, width)
+        image_embeds = self.proj(image_embeds)
+        image_embeds = image_embeds.view(batch, num_frames, *image_embeds.shape[1:])
+        image_embeds = image_embeds.flatten(3).transpose(2, 3)  # [batch, num_frames, height x width, channels]
+        image_embeds = image_embeds.flatten(1, 2)  # [batch, num_frames x height x width, channels]
+
+        embeds = torch.cat(
+            [text_embeds, image_embeds], dim=1
+        ).contiguous()  # [batch, seq_length + num_frames x height x width, channels]
+        return embeds
+
+
 def get_2d_rotary_pos_embed(embed_dim, crops_coords, grid_size, use_real=True):
     """
     RoPE for image tokens with 2d structure.

src/diffusers/models/normalization.py

@@ -34,19 +34,53 @@ class AdaLayerNorm(nn.Module):
 
     Parameters:
         embedding_dim (`int`): The size of each embedding vector.
-        num_embeddings (`int`): The size of the embeddings dictionary.
+        num_embeddings (`int`, *optional*): The size of the embeddings dictionary.
+        output_dim (`int`, *optional*):
+        norm_elementwise_affine (`bool`, defaults to `False):
+        norm_eps (`bool`, defaults to `False`):
+        chunk_dim (`int`, defaults to `0`):
     """
 
-    def __init__(self, embedding_dim: int, num_embeddings: int):
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_embeddings: Optional[int] = None,
+        output_dim: Optional[int] = None,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-5,
+        chunk_dim: int = 0,
+    ):
         super().__init__()
-        self.emb = nn.Embedding(num_embeddings, embedding_dim)
+
+        self.chunk_dim = chunk_dim
+        output_dim = output_dim or embedding_dim * 2
+
+        if num_embeddings is not None:
+            self.emb = nn.Embedding(num_embeddings, embedding_dim)
+        else:
+            self.emb = None
+
         self.silu = nn.SiLU()
-        self.linear = nn.Linear(embedding_dim, embedding_dim * 2)
-        self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False)
+        self.linear = nn.Linear(embedding_dim, output_dim)
+        self.norm = nn.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine)
+
+    def forward(
+        self, x: torch.Tensor, timestep: Optional[torch.Tensor] = None, temb: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        if self.emb is not None:
+            temb = self.emb(timestep)
+
+        temb = self.linear(self.silu(temb))
+
+        if self.chunk_dim == 1:
+            # This is a bit weird why we have the order of "shift, scale" here and "scale, shift" in the
+            # other if-branch. This branch is specific to CogVideoX for now.
+            shift, scale = temb.chunk(2, dim=1)
+            shift = shift[:, None, :]
+            scale = scale[:, None, :]
+        else:
+            scale, shift = temb.chunk(2, dim=0)
 
-    def forward(self, x: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor:
-        emb = self.linear(self.silu(self.emb(timestep)))
-        scale, shift = torch.chunk(emb, 2)
         x = self.norm(x) * (1 + scale) + shift
         return x
 
@@ -321,6 +355,30 @@ class LuminaLayerNormContinuous(nn.Module):
         return x
 
 
+class CogVideoXLayerNormZero(nn.Module):
+    def __init__(
+        self,
+        conditioning_dim: int,
+        embedding_dim: int,
+        elementwise_affine: bool = True,
+        eps: float = 1e-5,
+        bias: bool = True,
+    ) -> None:
+        super().__init__()
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(conditioning_dim, 6 * embedding_dim, bias=bias)
+        self.norm = nn.LayerNorm(embedding_dim, eps=eps, elementwise_affine=elementwise_affine)
+
+    def forward(
+        self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, temb: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        shift, scale, gate, enc_shift, enc_scale, enc_gate = self.linear(self.silu(temb)).chunk(6, dim=1)
+        hidden_states = self.norm(hidden_states) * (1 + scale)[:, None, :] + shift[:, None, :]
+        encoder_hidden_states = self.norm(encoder_hidden_states) * (1 + enc_scale)[:, None, :] + enc_shift[:, None, :]
+        return hidden_states, encoder_hidden_states, gate[:, None, :], enc_gate[:, None, :]
+
+
 if is_torch_version(">=", "2.1.0"):
     LayerNorm = nn.LayerNorm
 else:

src/diffusers/models/transformers/__init__.py

@@ -3,6 +3,7 @@ from ...utils import is_torch_available
 
 if is_torch_available():
     from .auraflow_transformer_2d import AuraFlowTransformer2DModel
+    from .cogvideox_transformer_3d import CogVideoXTransformer3DModel
     from .dit_transformer_2d import DiTTransformer2DModel
     from .dual_transformer_2d import DualTransformer2DModel
     from .hunyuan_transformer_2d import HunyuanDiT2DModel

src/diffusers/models/transformers/cogvideox_transformer_3d.py

+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Dict, Optional, Union
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import is_torch_version, logging
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import Attention, FeedForward
+from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps, get_3d_sincos_pos_embed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNorm, CogVideoXLayerNormZero
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@maybe_allow_in_graph
+class CogVideoXBlock(nn.Module):
+    r"""
+    Transformer block used in [CogVideoX](https://github.com/THUDM/CogVideo) model.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        qk_norm (`bool`, defaults to `True`):
+            Whether or not to use normalization after query and key projections in Attention.
+        norm_elementwise_affine (`bool`, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_eps (`float`, defaults to `1e-5`):
+            Epsilon value for normalization layers.
+        final_dropout (`bool` defaults to `False`):
+            Whether to apply a final dropout after the last feed-forward layer.
+        ff_inner_dim (`int`, *optional*, defaults to `None`):
+            Custom hidden dimension of Feed-forward layer. If not provided, `4 * dim` is used.
+        ff_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in Feed-forward layer.
+        attention_out_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in Attention output projection layer.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        time_embed_dim: int,
+        dropout: float = 0.0,
+        activation_fn: str = "gelu-approximate",
+        attention_bias: bool = False,
+        qk_norm: bool = True,
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        final_dropout: bool = True,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+
+        # 1. Self Attention
+        self.norm1 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            qk_norm="layer_norm" if qk_norm else None,
+            eps=1e-6,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+        )
+
+        # 2. Feed Forward
+        self.norm2 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+    ) -> torch.Tensor:
+        norm_hidden_states, norm_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(
+            hidden_states, encoder_hidden_states, temb
+        )
+
+        # attention
+        text_length = norm_encoder_hidden_states.size(1)
+
+        # CogVideoX uses concatenated text + video embeddings with self-attention instead of using
+        # them in cross-attention individually
+        norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
+        attn_output = self.attn1(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=None,
+        )
+
+        hidden_states = hidden_states + gate_msa * attn_output[:, text_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_output[:, :text_length]
+
+        # norm & modulate
+        norm_hidden_states, norm_encoder_hidden_states, gate_ff, enc_gate_ff = self.norm2(
+            hidden_states, encoder_hidden_states, temb
+        )
+
+        # feed-forward
+        norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
+        ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = hidden_states + gate_ff * ff_output[:, text_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_length]
+        return hidden_states, encoder_hidden_states
+
+
+class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
+    """
+    A Transformer model for video-like data in [CogVideoX](https://github.com/THUDM/CogVideo).
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input.
+        out_channels (`int`, *optional*):
+            The number of channels in the output.
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        patch_size (`int`, *optional*):
+            The size of the patches to use in the patch embedding layer.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states. During inference, you can denoise for up to but not more steps than
+            `num_embeds_ada_norm`.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The type of normalization to use. Options are `"layer_norm"` or `"ada_layer_norm"`.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether or not to use elementwise affine in normalization layers.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon value to use in normalization layers.
+        caption_channels (`int`, *optional*):
+            The number of channels in the caption embeddings.
+        video_length (`int`, *optional*):
+            The number of frames in the video-like data.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 30,
+        attention_head_dim: int = 64,
+        in_channels: Optional[int] = 16,
+        out_channels: Optional[int] = 16,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        time_embed_dim: int = 512,
+        text_embed_dim: int = 4096,
+        num_layers: int = 30,
+        dropout: float = 0.0,
+        attention_bias: bool = True,
+        sample_width: int = 90,
+        sample_height: int = 60,
+        sample_frames: int = 49,
+        patch_size: int = 2,
+        temporal_compression_ratio: int = 4,
+        max_text_seq_length: int = 226,
+        activation_fn: str = "gelu-approximate",
+        timestep_activation_fn: str = "silu",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        spatial_interpolation_scale: float = 1.875,
+        temporal_interpolation_scale: float = 1.0,
+    ):
+        super().__init__()
+        inner_dim = num_attention_heads * attention_head_dim
+
+        post_patch_height = sample_height // patch_size
+        post_patch_width = sample_width // patch_size
+        post_time_compression_frames = (sample_frames - 1) // temporal_compression_ratio + 1
+        self.num_patches = post_patch_height * post_patch_width * post_time_compression_frames
+
+        # 1. Patch embedding
+        self.patch_embed = CogVideoXPatchEmbed(patch_size, in_channels, inner_dim, text_embed_dim, bias=True)
+        self.embedding_dropout = nn.Dropout(dropout)
+
+        # 2. 3D positional embeddings
+        spatial_pos_embedding = get_3d_sincos_pos_embed(
+            inner_dim,
+            (post_patch_width, post_patch_height),
+            post_time_compression_frames,
+            spatial_interpolation_scale,
+            temporal_interpolation_scale,
+        )
+        spatial_pos_embedding = torch.from_numpy(spatial_pos_embedding).flatten(0, 1)
+        pos_embedding = torch.zeros(1, max_text_seq_length + self.num_patches, inner_dim, requires_grad=False)
+        pos_embedding.data[:, max_text_seq_length:].copy_(spatial_pos_embedding)
+        self.register_buffer("pos_embedding", pos_embedding, persistent=False)
+
+        # 3. Time embeddings
+        self.time_proj = Timesteps(inner_dim, flip_sin_to_cos, freq_shift)
+        self.time_embedding = TimestepEmbedding(inner_dim, time_embed_dim, timestep_activation_fn)
+
+        # 4. Define spatio-temporal transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                CogVideoXBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    time_embed_dim=time_embed_dim,
+                    dropout=dropout,
+                    activation_fn=activation_fn,
+                    attention_bias=attention_bias,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+        self.norm_final = nn.LayerNorm(inner_dim, norm_eps, norm_elementwise_affine)
+
+        # 5. Output blocks
+        self.norm_out = AdaLayerNorm(
+            embedding_dim=time_embed_dim,
+            output_dim=2 * inner_dim,
+            norm_elementwise_affine=norm_elementwise_affine,
+            norm_eps=norm_eps,
+            chunk_dim=1,
+        )
+        self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        self.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: Union[int, float, torch.LongTensor],
+        timestep_cond: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        batch_size, num_frames, channels, height, width = hidden_states.shape
+
+        # 1. Time embedding
+        timesteps = timestep
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=hidden_states.dtype)
+        emb = self.time_embedding(t_emb, timestep_cond)
+
+        # 2. Patch embedding
+        hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
+
+        # 3. Position embedding
+        seq_length = height * width * num_frames // (self.config.patch_size**2)
+
+        pos_embeds = self.pos_embedding[:, : self.config.max_text_seq_length + seq_length]
+        hidden_states = hidden_states + pos_embeds
+        hidden_states = self.embedding_dropout(hidden_states)
+
+        encoder_hidden_states = hidden_states[:, : self.config.max_text_seq_length]
+        hidden_states = hidden_states[:, self.config.max_text_seq_length :]
+
+        # 5. Transformer blocks
+        for i, block in enumerate(self.transformer_blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states,
+                    emb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    temb=emb,
+                )
+
+        hidden_states = self.norm_final(hidden_states)
+
+        # 6. Final block
+        hidden_states = self.norm_out(hidden_states, temb=emb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # 7. Unpatchify
+        p = self.config.patch_size
+        output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, channels, p, p)
+        output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)

src/diffusers/models/upsampling.py

@@ -348,6 +348,70 @@ class KUpsample2D(nn.Module):
         return F.conv_transpose2d(inputs, weight, stride=2, padding=self.pad * 2 + 1)
 
 
+class CogVideoXUpsample3D(nn.Module):
+    r"""
+    A 3D Upsample layer using in CogVideoX by Tsinghua University & ZhipuAI # Todo: Wait for paper relase.
+
+    Args:
+        in_channels (`int`):
+            Number of channels in the input image.
+        out_channels (`int`):
+            Number of channels produced by the convolution.
+        kernel_size (`int`, defaults to `3`):
+            Size of the convolving kernel.
+        stride (`int`, defaults to `1`):
+            Stride of the convolution.
+        padding (`int`, defaults to `1`):
+            Padding added to all four sides of the input.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to compress the time dimension.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int = 3,
+        stride: int = 1,
+        padding: int = 1,
+        compress_time: bool = False,
+    ) -> None:
+        super().__init__()
+
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        if self.compress_time:
+            if inputs.shape[2] > 1 and inputs.shape[2] % 2 == 1:
+                # split first frame
+                x_first, x_rest = inputs[:, :, 0], inputs[:, :, 1:]
+
+                x_first = F.interpolate(x_first, scale_factor=2.0)
+                x_rest = F.interpolate(x_rest, scale_factor=2.0)
+                x_first = x_first[:, :, None, :, :]
+                inputs = torch.cat([x_first, x_rest], dim=2)
+            elif inputs.shape[2] > 1:
+                inputs = F.interpolate(inputs, scale_factor=2.0)
+            else:
+                inputs = inputs.squeeze(2)
+                inputs = F.interpolate(inputs, scale_factor=2.0)
+                inputs = inputs[:, :, None, :, :]
+        else:
+            # only interpolate 2D
+            b, c, t, h, w = inputs.shape
+            inputs = inputs.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+            inputs = F.interpolate(inputs, scale_factor=2.0)
+            inputs = inputs.reshape(b, t, c, *inputs.shape[2:]).permute(0, 2, 1, 3, 4)
+
+        b, c, t, h, w = inputs.shape
+        inputs = inputs.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        inputs = self.conv(inputs)
+        inputs = inputs.reshape(b, t, *inputs.shape[1:]).permute(0, 2, 1, 3, 4)
+
+        return inputs
+
+
 def upfirdn2d_native(
     tensor: torch.Tensor,
     kernel: torch.Tensor,

src/diffusers/pipelines/__init__.py

@@ -132,6 +132,7 @@ else:
         "AudioLDM2UNet2DConditionModel",
     ]
     _import_structure["blip_diffusion"] = ["BlipDiffusionPipeline"]
+    _import_structure["cogvideo"] = ["CogVideoXPipeline"]
     _import_structure["controlnet"].extend(
         [
             "BlipDiffusionControlNetPipeline",
@@ -451,6 +452,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         )
         from .aura_flow import AuraFlowPipeline
         from .blip_diffusion import BlipDiffusionPipeline
+        from .cogvideo import CogVideoXPipeline
         from .controlnet import (
             BlipDiffusionControlNetPipeline,
             StableDiffusionControlNetImg2ImgPipeline,

src/diffusers/pipelines/cogvideo/__init__.py

+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_cogvideox"] = ["CogVideoXPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_cogvideox import CogVideoXPipeline
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/schedulers/__init__.py

@@ -43,12 +43,14 @@ else:
     _import_structure["scheduling_consistency_decoder"] = ["ConsistencyDecoderScheduler"]
     _import_structure["scheduling_consistency_models"] = ["CMStochasticIterativeScheduler"]
     _import_structure["scheduling_ddim"] = ["DDIMScheduler"]
+    _import_structure["scheduling_ddim_cogvideox"] = ["CogVideoXDDIMScheduler"]
     _import_structure["scheduling_ddim_inverse"] = ["DDIMInverseScheduler"]
     _import_structure["scheduling_ddim_parallel"] = ["DDIMParallelScheduler"]
     _import_structure["scheduling_ddpm"] = ["DDPMScheduler"]
     _import_structure["scheduling_ddpm_parallel"] = ["DDPMParallelScheduler"]
     _import_structure["scheduling_ddpm_wuerstchen"] = ["DDPMWuerstchenScheduler"]
     _import_structure["scheduling_deis_multistep"] = ["DEISMultistepScheduler"]
+    _import_structure["scheduling_dpm_cogvideox"] = ["CogVideoXDPMScheduler"]
     _import_structure["scheduling_dpmsolver_multistep"] = ["DPMSolverMultistepScheduler"]
     _import_structure["scheduling_dpmsolver_multistep_inverse"] = ["DPMSolverMultistepInverseScheduler"]
     _import_structure["scheduling_dpmsolver_singlestep"] = ["DPMSolverSinglestepScheduler"]
@@ -141,12 +143,14 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .scheduling_consistency_decoder import ConsistencyDecoderScheduler
         from .scheduling_consistency_models import CMStochasticIterativeScheduler
         from .scheduling_ddim import DDIMScheduler
+        from .scheduling_ddim_cogvideox import CogVideoXDDIMScheduler
         from .scheduling_ddim_inverse import DDIMInverseScheduler
         from .scheduling_ddim_parallel import DDIMParallelScheduler
         from .scheduling_ddpm import DDPMScheduler
         from .scheduling_ddpm_parallel import DDPMParallelScheduler
         from .scheduling_ddpm_wuerstchen import DDPMWuerstchenScheduler
         from .scheduling_deis_multistep import DEISMultistepScheduler
+        from .scheduling_dpm_cogvideox import CogVideoXDPMScheduler
         from .scheduling_dpmsolver_multistep import DPMSolverMultistepScheduler
         from .scheduling_dpmsolver_multistep_inverse import DPMSolverMultistepInverseScheduler
         from .scheduling_dpmsolver_singlestep import DPMSolverSinglestepScheduler