Migrated from GitHub

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/demo.gif filter=lfs diff=lfs merge=lfs -text
+assets/examples/speech.wav filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   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.

ORIGINAL_README.md

@@ -0,0 +1,73 @@
+# MEMO
+
+**MEMO: Memory-Guided Diffusion for Expressive Talking Video Generation**
+<br>
+[Longtao Zheng](https://ltzheng.github.io)\*,
+[Yifan Zhang](https://scholar.google.com/citations?user=zuYIUJEAAAAJ)\*,
+[Hanzhong Guo](https://scholar.google.com/citations?user=q3x6KsgAAAAJ)\,
+[Jiachun Pan](https://scholar.google.com/citations?user=nrOvfb4AAAAJ),
+[Zhenxiong Tan](https://scholar.google.com/citations?user=HP9Be6UAAAAJ),
+[Jiahao Lu](https://scholar.google.com/citations?user=h7rbA-sAAAAJ),
+[Chuanxin Tang](https://scholar.google.com/citations?user=3ZC8B7MAAAAJ),
+[Bo An](https://personal.ntu.edu.sg/boan/index.html),
+[Shuicheng Yan](https://scholar.google.com/citations?user=DNuiPHwAAAAJ)
+<br>
+_[Project Page](https://memoavatar.github.io) | [arXiv](https://arxiv.org/abs/2412.04448) | [Model](https://huggingface.co/memoavatar/memo)_
+
+This repository contains the example inference script for the MEMO-preview model. The gif demo below is compressed. See our [project page](https://memoavatar.github.io) for full videos.
+
+<div style="width: 100%; text-align: center;">
+    <img src="assets/demo.gif" alt="Demo GIF" style="width: 100%; height: auto;">
+</div>
+
+## Installation
+
+```bash
+conda create -n memo python=3.10 -y
+conda activate memo
+conda install -c conda-forge ffmpeg -y
+pip install -e .
+```
+
+> Our code will download the checkpoint from Hugging Face automatically, and the models for face analysis and vocal separation will be downloaded to `misc_model_dir` of `configs/inference.yaml`. If you want to download the models manually, please download the checkpoint from [here](https://huggingface.co/memoavatar/memo) and specify the path in `model_name_or_path` of `configs/inference.yaml`.
+
+## Inference
+
+```bash
+python inference.py --config configs/inference.yaml --input_image <IMAGE_PATH> --input_audio <AUDIO_PATH> --output_dir <SAVE_PATH>
+```
+
+For example:
+
+```bash
+python inference.py --config configs/inference.yaml --input_image assets/examples/dicaprio.jpg --input_audio assets/examples/speech.wav --output_dir outputs
+```
+
+> We tested the code on H100 and RTX 4090 GPUs using CUDA 12. Under the default settings (fps=30, inference_steps=20), the inference time is around 1 second per frame on H100 and 2 seconds per frame on RTX 4090. We welcome community contributions to improve the inference speed or interfaces like ComfyUI.
+
+## Acknowledgement
+
+Our work is made possible thanks to high-quality open-source talking video datasets (including [HDTF](https://github.com/MRzzm/HDTF), [VFHQ](https://liangbinxie.github.io/projects/vfhq), [CelebV-HQ](https://celebv-hq.github.io), [MultiTalk](https://multi-talk.github.io), and [MEAD](https://wywu.github.io/projects/MEAD/MEAD.html)) and some pioneering works (such as [EMO](https://humanaigc.github.io/emote-portrait-alive) and [Hallo](https://github.com/fudan-generative-vision/hallo)).
+
+## Ethics Statement
+
+We acknowledge the potential of AI in generating talking videos, with applications spanning education, virtual assistants, and entertainment. However, we are equally aware of the ethical, legal, and societal challenges that misuse of this technology could pose. 
+
+To reduce potential risks, we have only open-sourced a preview model for research purposes. Demos on our website use publicly available materials. We welcome copyright concerns—please contact us if needed, and we will address issues promptly. Users are required to ensure that their actions align with legal regulations, cultural norms, and ethical standards. 
+
+It is strictly prohibited to use the model for creating malicious, misleading, defamatory, or privacy-infringing content, such as deepfake videos for political misinformation, impersonation, harassment, or fraud. We strongly encourage users to review generated content carefully, ensuring it meets ethical guidelines and respects the rights of all parties involved. Users must also ensure that their inputs (e.g., audio and reference images) and outputs are used with proper authorization. Unauthorized use of third-party intellectual property is strictly forbidden. 
+
+While users may claim ownership of content generated by the model, they must ensure compliance with copyright laws, particularly when involving public figures' likeness, voice, or other aspects protected under personality rights.
+
+## Citation
+
+If you find our work useful, please use the following citation:
+
+```bibtex
+@article{zheng2024memo,
+  title={MEMO: Memory-Guided Diffusion for Expressive Talking Video Generation},
+  author={Longtao Zheng and Yifan Zhang and Hanzhong Guo and Jiachun Pan and Zhenxiong Tan and Jiahao Lu and Chuanxin Tang and Bo An and Shuicheng Yan},
+  journal={arXiv preprint arXiv:2412.04448},
+  year={2024}
+}
+```

assets/examples/speech.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16e82c22a2e7104861943b994c40a537271653cb7d0b1b722dda2cda8ab75a7c
+size 2646078

configs/inference.yaml

@@ -0,0 +1,16 @@
+resolution: 512
+num_generated_frames_per_clip: 16
+fps: 30
+num_init_past_frames: 2
+num_past_frames: 16
+inference_steps: 20
+cfg_scale: 3.5
+weight_dtype: bf16
+enable_xformers_memory_efficient_attention: true
+
+model_name_or_path: memoavatar/memo
+# model_name_or_path: checkpoints
+vae: stabilityai/sd-vae-ft-mse
+wav2vec: facebook/wav2vec2-base-960h
+emotion2vec: iic/emotion2vec_plus_large
+misc_model_dir: checkpoints

inference.py

@@ -0,0 +1,251 @@
+import argparse
+import logging
+import os
+
+import torch
+from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler
+from diffusers.utils.import_utils import is_xformers_available
+from omegaconf import OmegaConf
+from packaging import version
+from tqdm import tqdm
+
+from memo.models.audio_proj import AudioProjModel
+from memo.models.image_proj import ImageProjModel
+from memo.models.unet_2d_condition import UNet2DConditionModel
+from memo.models.unet_3d import UNet3DConditionModel
+from memo.pipelines.video_pipeline import VideoPipeline
+from memo.utils.audio_utils import extract_audio_emotion_labels, preprocess_audio, resample_audio
+from memo.utils.vision_utils import preprocess_image, tensor_to_video
+
+
+logger = logging.getLogger("memo")
+logger.setLevel(logging.INFO)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Inference script for MEMO")
+
+    parser.add_argument("--config", type=str, default="configs/inference.yaml")
+    parser.add_argument("--input_image", type=str)
+    parser.add_argument("--input_audio", type=str)
+    parser.add_argument("--output_dir", type=str)
+    parser.add_argument("--seed", type=int, default=42)
+
+    return parser.parse_args()
+
+
+def main():
+    # Parse arguments
+    args = parse_args()
+    input_image_path = args.input_image
+    input_audio_path = args.input_audio
+    if "wav" not in input_audio_path:
+        logger.warning("MEMO might not generate full-length video for non-wav audio file.")
+    output_dir = args.output_dir
+    os.makedirs(output_dir, exist_ok=True)
+    output_video_path = os.path.join(
+        output_dir,
+        f"{os.path.basename(input_image_path).split('.')[0]}_{os.path.basename(input_audio_path).split('.')[0]}.mp4",
+    )
+
+    if os.path.exists(output_video_path):
+        logger.info(f"Output file {output_video_path} already exists. Skipping inference.")
+        return
+
+    generator = torch.manual_seed(args.seed)
+
+    logger.info(f"Loading config from {args.config}")
+    config = OmegaConf.load(args.config)
+
+    # Determine model paths
+    if config.model_name_or_path == "memoavatar/memo":
+        logger.info(
+            f"The MEMO model will be downloaded from Hugging Face to the default cache directory. The models for face analysis and vocal separation will be downloaded to {config.misc_model_dir}."
+        )
+
+        face_analysis = os.path.join(config.misc_model_dir, "misc/face_analysis")
+        os.makedirs(face_analysis, exist_ok=True)
+        for model in [
+            "1k3d68.onnx",
+            "2d106det.onnx",
+            "face_landmarker_v2_with_blendskapes.task",
+            "genderage.onnx",
+            "glintr100.onnx",
+            "scrfd_10g_bnkps.onnx",
+        ]:
+            if not os.path.exists(os.path.join(face_analysis, model)):
+                logger.info(f"Downloading {model} to {face_analysis}")
+                os.system(
+                    f"wget -P {face_analysis} https://huggingface.co/memoavatar/memo/raw/main/misc/face_analysis/models/{model}"
+                )
+        logger.info(f"Use face analysis models from {face_analysis}")
+
+        vocal_separator = os.path.join(config.misc_model_dir, "misc/vocal_separator/Kim_Vocal_2.onnx")
+        if os.path.exists(vocal_separator):
+            logger.info(f"Vocal separator {vocal_separator} already exists. Skipping download.")
+        else:
+            logger.info(f"Downloading vocal separator to {vocal_separator}")
+            os.makedirs(os.path.dirname(vocal_separator), exist_ok=True)
+            os.system(
+                f"wget -P {os.path.dirname(vocal_separator)} https://huggingface.co/memoavatar/memo/raw/main/misc/vocal_separator/Kim_Vocal_2.onnx"
+            )
+    else:
+        logger.info(f"Loading manually specified model path: {config.model_name_or_path}")
+        face_analysis = os.path.join(config.model_name_or_path, "misc/face_analysis")
+        vocal_separator = os.path.join(config.model_name_or_path, "misc/vocal_separator/Kim_Vocal_2.onnx")
+
+    # Set up device and weight dtype
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    if config.weight_dtype == "fp16":
+        weight_dtype = torch.float16
+    elif config.weight_dtype == "bf16":
+        weight_dtype = torch.bfloat16
+    elif config.weight_dtype == "fp32":
+        weight_dtype = torch.float32
+    else:
+        weight_dtype = torch.float32
+    logger.info(f"Inference dtype: {weight_dtype}")
+
+    logger.info(f"Processing image {input_image_path}")
+    img_size = (config.resolution, config.resolution)
+    pixel_values, face_emb = preprocess_image(
+        face_analysis_model=face_analysis,
+        image_path=input_image_path,
+        image_size=config.resolution,
+    )
+
+    logger.info(f"Processing audio {input_audio_path}")
+    cache_dir = os.path.join(output_dir, "audio_preprocess")
+    os.makedirs(cache_dir, exist_ok=True)
+    input_audio_path = resample_audio(
+        input_audio_path,
+        os.path.join(cache_dir, f"{os.path.basename(input_audio_path).split('.')[0]}-16k.wav"),
+    )
+    audio_emb, audio_length = preprocess_audio(
+        wav_path=input_audio_path,
+        num_generated_frames_per_clip=config.num_generated_frames_per_clip,
+        fps=config.fps,
+        wav2vec_model=config.wav2vec,
+        vocal_separator_model=vocal_separator,
+        cache_dir=cache_dir,
+        device=device,
+    )
+
+    logger.info("Processing audio emotion")
+    audio_emotion, num_emotion_classes = extract_audio_emotion_labels(
+        model=config.model_name_or_path,
+        wav_path=input_audio_path,
+        emotion2vec_model=config.emotion2vec,
+        audio_length=audio_length,
+        device=device,
+    )
+
+    logger.info("Loading models")
+    vae = AutoencoderKL.from_pretrained(config.vae).to(device=device, dtype=weight_dtype)
+    reference_net = UNet2DConditionModel.from_pretrained(
+        config.model_name_or_path, subfolder="reference_net", use_safetensors=True
+    )
+    diffusion_net = UNet3DConditionModel.from_pretrained(
+        config.model_name_or_path, subfolder="diffusion_net", use_safetensors=True
+    )
+    image_proj = ImageProjModel.from_pretrained(
+        config.model_name_or_path, subfolder="image_proj", use_safetensors=True
+    )
+    audio_proj = AudioProjModel.from_pretrained(
+        config.model_name_or_path, subfolder="audio_proj", use_safetensors=True
+    )
+
+    vae.requires_grad_(False).eval()
+    reference_net.requires_grad_(False).eval()
+    diffusion_net.requires_grad_(False).eval()
+    image_proj.requires_grad_(False).eval()
+    audio_proj.requires_grad_(False).eval()
+
+    # Enable memory-efficient attention for xFormers
+    if config.enable_xformers_memory_efficient_attention:
+        if is_xformers_available():
+            import xformers
+
+            xformers_version = version.parse(xformers.__version__)
+            if xformers_version == version.parse("0.0.16"):
+                logger.info(
+                    "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
+                )
+            reference_net.enable_xformers_memory_efficient_attention()
+            diffusion_net.enable_xformers_memory_efficient_attention()
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
+
+    # Create inference pipeline
+    noise_scheduler = FlowMatchEulerDiscreteScheduler()
+    pipeline = VideoPipeline(
+        vae=vae,
+        reference_net=reference_net,
+        diffusion_net=diffusion_net,
+        scheduler=noise_scheduler,
+        image_proj=image_proj,
+    )
+    pipeline.to(device=device, dtype=weight_dtype)
+
+    video_frames = []
+    num_clips = audio_emb.shape[0] // config.num_generated_frames_per_clip
+    for t in tqdm(range(num_clips), desc="Generating video clips"):
+        if len(video_frames) == 0:
+            # Initialize the first past frames with reference image
+            past_frames = pixel_values.repeat(config.num_init_past_frames, 1, 1, 1)
+            past_frames = past_frames.to(dtype=pixel_values.dtype, device=pixel_values.device)
+            pixel_values_ref_img = torch.cat([pixel_values, past_frames], dim=0)
+        else:
+            past_frames = video_frames[-1][0]
+            past_frames = past_frames.permute(1, 0, 2, 3)
+            past_frames = past_frames[0 - config.num_past_frames :]
+            past_frames = past_frames * 2.0 - 1.0
+            past_frames = past_frames.to(dtype=pixel_values.dtype, device=pixel_values.device)
+            pixel_values_ref_img = torch.cat([pixel_values, past_frames], dim=0)
+
+        pixel_values_ref_img = pixel_values_ref_img.unsqueeze(0)
+
+        audio_tensor = (
+            audio_emb[
+                t
+                * config.num_generated_frames_per_clip : min(
+                    (t + 1) * config.num_generated_frames_per_clip, audio_emb.shape[0]
+                )
+            ]
+            .unsqueeze(0)
+            .to(device=audio_proj.device, dtype=audio_proj.dtype)
+        )
+        audio_tensor = audio_proj(audio_tensor)
+
+        audio_emotion_tensor = audio_emotion[
+            t
+            * config.num_generated_frames_per_clip : min(
+                (t + 1) * config.num_generated_frames_per_clip, audio_emb.shape[0]
+            )
+        ]
+
+        pipeline_output = pipeline(
+            ref_image=pixel_values_ref_img,
+            audio_tensor=audio_tensor,
+            audio_emotion=audio_emotion_tensor,
+            emotion_class_num=num_emotion_classes,
+            face_emb=face_emb,
+            width=img_size[0],
+            height=img_size[1],
+            video_length=config.num_generated_frames_per_clip,
+            num_inference_steps=config.inference_steps,
+            guidance_scale=config.cfg_scale,
+            generator=generator,
+        )
+
+        video_frames.append(pipeline_output.videos)
+
+    video_frames = torch.cat(video_frames, dim=2)
+    video_frames = video_frames.squeeze(0)
+    video_frames = video_frames[:, :audio_length]
+
+    tensor_to_video(video_frames, output_video_path, input_audio_path, fps=config.fps)
+
+
+if __name__ == "__main__":
+    main()

memo/models/attention.py

@@ -0,0 +1,639 @@
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.models.attention import (
+    AdaLayerNorm,
+    AdaLayerNormZero,
+    Attention,
+    FeedForward,
+)
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from einops import rearrange
+from torch import nn
+
+from memo.models.attention_processor import Attention as CustomAttention
+from memo.models.attention_processor import JointAttnProcessor2_0
+
+
+class GatedSelfAttentionDense(nn.Module):
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+        self.enabled = True
+
+    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+        is_final_block: bool = False,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.is_final_block = is_final_block
+
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        if not is_final_block:
+            self.attn1 = Attention(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+                upcast_attention=upcast_attention,
+            )
+
+            # 2. Cross-Attn
+            if cross_attention_dim is not None or double_self_attention:
+                # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+                # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+                # the second cross attention block.
+                self.norm2 = (
+                    AdaLayerNorm(dim, num_embeds_ada_norm)
+                    if self.use_ada_layer_norm
+                    else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+                )
+                self.attn2 = Attention(
+                    query_dim=dim,
+                    cross_attention_dim=(cross_attention_dim if not double_self_attention else None),
+                    heads=num_attention_heads,
+                    dim_head=attention_head_dim,
+                    dropout=dropout,
+                    bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                )
+            else:
+                self.norm2 = None
+                self.attn2 = None
+
+            # 3. Feed-forward
+            if not self.use_ada_layer_norm_single:
+                self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+            self.ff = FeedForward(
+                dim,
+                dropout=dropout,
+                activation_fn=activation_fn,
+                final_dropout=final_dropout,
+            )
+
+            # 4. Fuser
+            if attention_type in {"gated", "gated-text-image"}:  # Updated line
+                self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+            # 5. Scale-shift for PixArt-Alpha.
+            if self.use_ada_layer_norm_single:
+                self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+            # let chunk size default to None
+            self._chunk_size = None
+            self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        gate_msa = None
+        scale_mlp = None
+        shift_mlp = None
+        gate_mlp = None
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        ref_feature = norm_hidden_states
+        if self.is_final_block:
+            return None, ref_feature
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=(encoder_hidden_states if self.only_cross_attention else None),
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states.repeat(
+                    norm_hidden_states.shape[0] // encoder_hidden_states.shape[0], 1, 1
+                ),
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states, ref_feature
+
+
+class TemporalBasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
+        self.unet_use_temporal_attention = unet_use_temporal_attention
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+        self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+        # Cross-Attn
+        if cross_attention_dim is not None:
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            self.attn2 = None
+
+        if cross_attention_dim is not None:
+            self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+        else:
+            self.norm2 = None
+
+        # Feed-forward
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.norm3 = nn.LayerNorm(dim)
+        self.use_ada_layer_norm_zero = False
+
+        # Temp-Attn
+        if unet_use_temporal_attention is None:
+            unet_use_temporal_attention = False
+        if unet_use_temporal_attention:
+            self.attn_temp = Attention(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+            nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
+            self.norm_temp = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        ref_img_feature: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        video_length=None,
+        uc_mask=None,
+    ):
+        norm_hidden_states = self.norm1(hidden_states)
+
+        # 1. Self-Attention
+        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+        ref_img_feature = ref_img_feature.repeat(video_length, 1, 1)
+        modify_norm_hidden_states = torch.cat((norm_hidden_states, ref_img_feature), dim=1).to(
+            dtype=norm_hidden_states.dtype
+        )
+        hidden_states_uc = (
+            self.attn1(
+                norm_hidden_states,
+                encoder_hidden_states=modify_norm_hidden_states,
+                attention_mask=attention_mask,
+            )
+            + hidden_states
+        )
+        if uc_mask is not None:
+            hidden_states_c = hidden_states_uc.clone()
+            _uc_mask = uc_mask.clone()
+            if hidden_states.shape[0] != _uc_mask.shape[0]:
+                _uc_mask = (
+                    torch.Tensor([1] * (hidden_states.shape[0] // 2) + [0] * (hidden_states.shape[0] // 2))
+                    .to(hidden_states_uc.device)
+                    .bool()
+                )
+            hidden_states_c[_uc_mask] = (
+                self.attn1(
+                    norm_hidden_states[_uc_mask],
+                    encoder_hidden_states=norm_hidden_states[_uc_mask],
+                    attention_mask=attention_mask,
+                )
+                + hidden_states[_uc_mask]
+            )
+            hidden_states = hidden_states_c.clone()
+        else:
+            hidden_states = hidden_states_uc
+
+        if self.attn2 is not None:
+            # Cross-Attention
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+            )
+            hidden_states = (
+                self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                )
+                + hidden_states
+            )
+
+        # Feed-forward
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+        # Temporal-Attention
+        if self.unet_use_temporal_attention:
+            d = hidden_states.shape[1]
+            hidden_states = rearrange(hidden_states, "(b f) d c -> (b d) f c", f=video_length)
+            norm_hidden_states = (
+                self.norm_temp(hidden_states, timestep) if self.use_ada_layer_norm else self.norm_temp(hidden_states)
+            )
+            hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states
+
+
+class LabelEmbedding(nn.Module):
+    def __init__(self, num_classes, hidden_size, dropout_prob):
+        super().__init__()
+        use_cfg_embedding = dropout_prob > 0
+        self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size)
+        self.num_classes = num_classes
+        self.dropout_prob = dropout_prob
+
+    def token_drop(self, labels, force_drop_ids=None):
+        # Drops labels to enable classifier-free guidance.
+        if force_drop_ids is None:
+            drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
+        else:
+            drop_ids = torch.tensor(force_drop_ids == 1)
+        labels = torch.where(drop_ids, self.num_classes, labels)
+
+        return labels
+
+    def forward(self, labels: torch.LongTensor, force_drop_ids=None):
+        use_dropout = self.dropout_prob > 0
+        if (self.training and use_dropout) or (force_drop_ids is not None):
+            labels = self.token_drop(labels, force_drop_ids)
+        embeddings = self.embedding_table(labels)
+
+        return embeddings
+
+
+class EmoAdaLayerNorm(nn.Module):
+    def __init__(
+        self,
+        embedding_dim,
+        num_classes=9,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-5,
+        class_dropout_prob=0.3,
+    ):
+        super().__init__()
+        self.class_embedder = LabelEmbedding(num_classes, embedding_dim, class_dropout_prob)
+        self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(embedding_dim, 2 * embedding_dim, bias=True))
+
+    def forward(self, x, emotion=None):
+        emo_embedding = self.class_embedder(emotion)
+        shift, scale = self.adaLN_modulation(emo_embedding).chunk(2, dim=1)
+        if emotion.shape[0] > 1:
+            repeat = x.shape[0] // emo_embedding.shape[0]
+            scale = scale.unsqueeze(1)
+            scale = torch.repeat_interleave(scale, repeats=repeat, dim=0)
+            shift = shift.unsqueeze(1)
+            shift = torch.repeat_interleave(shift, repeats=repeat, dim=0)
+        else:
+            scale = scale.unsqueeze(1)
+            shift = shift.unsqueeze(1)
+
+        x = self.norm(x) * (1 + scale) + shift
+
+        return x
+
+
+class JointAudioTemporalBasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        depth=0,
+        unet_block_name=None,
+        use_ada_layer_norm=False,
+        emo_drop_rate=0.3,
+        is_final_block=False,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = use_ada_layer_norm
+        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
+        self.unet_use_temporal_attention = unet_use_temporal_attention
+        self.unet_block_name = unet_block_name
+        self.depth = depth
+        self.is_final_block = is_final_block
+
+        self.norm1 = (
+            EmoAdaLayerNorm(dim, num_classes=9, class_dropout_prob=emo_drop_rate)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(dim)
+        )
+        self.attn1 = CustomAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+
+        self.audio_norm1 = (
+            EmoAdaLayerNorm(cross_attention_dim, num_classes=9, class_dropout_prob=emo_drop_rate)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(cross_attention_dim)
+        )
+        self.audio_attn1 = CustomAttention(
+            query_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+
+        self.norm2 = (
+            EmoAdaLayerNorm(dim, num_classes=9, class_dropout_prob=emo_drop_rate)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(dim)
+        )
+        self.audio_norm2 = (
+            EmoAdaLayerNorm(cross_attention_dim, num_classes=9, class_dropout_prob=emo_drop_rate)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(cross_attention_dim)
+        )
+
+        # Joint Attention
+        self.attn2 = CustomAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            cross_attention_dim=dim,
+            added_kv_proj_dim=cross_attention_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+            only_cross_attention=False,
+            out_dim=dim,
+            context_out_dim=cross_attention_dim,
+            context_pre_only=False,
+            processor=JointAttnProcessor2_0(),
+            is_final_block=is_final_block,
+        )
+
+        # Feed-forward
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.norm3 = nn.LayerNorm(dim)
+        if not is_final_block:
+            self.audio_ff = FeedForward(cross_attention_dim, dropout=dropout, activation_fn=activation_fn)
+            self.audio_norm3 = nn.LayerNorm(cross_attention_dim)
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        emotion=None,
+    ):
+        norm_hidden_states = (
+            self.norm1(hidden_states, emotion) if self.use_ada_layer_norm else self.norm1(hidden_states)
+        )
+        norm_encoder_hidden_states = (
+            self.audio_norm1(encoder_hidden_states, emotion)
+            if self.use_ada_layer_norm
+            else self.audio_norm1(encoder_hidden_states)
+        )
+
+        hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask) + hidden_states
+
+        encoder_hidden_states = (
+            self.audio_attn1(norm_encoder_hidden_states, attention_mask=attention_mask) + encoder_hidden_states
+        )
+
+        norm_hidden_states = (
+            self.norm2(hidden_states, emotion) if self.use_ada_layer_norm else self.norm2(hidden_states)
+        )
+        norm_encoder_hidden_states = (
+            self.audio_norm2(encoder_hidden_states, emotion)
+            if self.use_ada_layer_norm
+            else self.audio_norm2(encoder_hidden_states)
+        )
+
+        joint_hidden_states, joint_encoder_hidden_states = self.attn2(
+            norm_hidden_states,
+            norm_encoder_hidden_states,
+        )
+
+        hidden_states = joint_hidden_states + hidden_states
+        if not self.is_final_block:
+            encoder_hidden_states = joint_encoder_hidden_states + encoder_hidden_states
+
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+        if not self.is_final_block:
+            encoder_hidden_states = self.audio_ff(self.audio_norm3(encoder_hidden_states)) + encoder_hidden_states
+        else:
+            encoder_hidden_states = None
+
+        return hidden_states, encoder_hidden_states
+
+
+def zero_module(module):
+    for p in module.parameters():
+        nn.init.zeros_(p)
+
+    return module

memo/models/attention_processor.py

@@ -0,0 +1,2299 @@
+import inspect
+import math
+from typing import Callable, Optional, Union
+
+import torch
+import torch.nn.functional as F
+from diffusers.utils import deprecate, logging
+from diffusers.utils.import_utils import is_torch_npu_available, is_xformers_available
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from torch import nn
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+if is_torch_npu_available():
+    import torch_npu
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+@maybe_allow_in_graph
+class Attention(nn.Module):
+    r"""
+    A cross attention layer.
+
+    Parameters:
+        query_dim (`int`):
+            The number of channels in the query.
+        cross_attention_dim (`int`, *optional*):
+            The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
+        heads (`int`,  *optional*, defaults to 8):
+            The number of heads to use for multi-head attention.
+        dim_head (`int`,  *optional*, defaults to 64):
+            The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability to use.
+        bias (`bool`, *optional*, defaults to False):
+            Set to `True` for the query, key, and value linear layers to contain a bias parameter.
+        upcast_attention (`bool`, *optional*, defaults to False):
+            Set to `True` to upcast the attention computation to `float32`.
+        upcast_softmax (`bool`, *optional*, defaults to False):
+            Set to `True` to upcast the softmax computation to `float32`.
+        cross_attention_norm (`str`, *optional*, defaults to `None`):
+            The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.
+        cross_attention_norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use for the group norm in the cross attention.
+        added_kv_proj_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the added key and value projections. If `None`, no projection is used.
+        norm_num_groups (`int`, *optional*, defaults to `None`):
+            The number of groups to use for the group norm in the attention.
+        spatial_norm_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the spatial normalization.
+        out_bias (`bool`, *optional*, defaults to `True`):
+            Set to `True` to use a bias in the output linear layer.
+        scale_qk (`bool`, *optional*, defaults to `True`):
+            Set to `True` to scale the query and key by `1 / sqrt(dim_head)`.
+        only_cross_attention (`bool`, *optional*, defaults to `False`):
+            Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if
+            `added_kv_proj_dim` is not `None`.
+        eps (`float`, *optional*, defaults to 1e-5):
+            An additional value added to the denominator in group normalization that is used for numerical stability.
+        rescale_output_factor (`float`, *optional*, defaults to 1.0):
+            A factor to rescale the output by dividing it with this value.
+        residual_connection (`bool`, *optional*, defaults to `False`):
+            Set to `True` to add the residual connection to the output.
+        _from_deprecated_attn_block (`bool`, *optional*, defaults to `False`):
+            Set to `True` if the attention block is loaded from a deprecated state dict.
+        processor (`AttnProcessor`, *optional*, defaults to `None`):
+            The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and
+            `AttnProcessor` otherwise.
+    """
+
+    def __init__(
+        self,
+        query_dim: int,
+        cross_attention_dim: Optional[int] = None,
+        heads: int = 8,
+        kv_heads: Optional[int] = None,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias: bool = False,
+        upcast_attention: bool = False,
+        upcast_softmax: bool = False,
+        cross_attention_norm: Optional[str] = None,
+        cross_attention_norm_num_groups: int = 32,
+        qk_norm: Optional[str] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        added_proj_bias: Optional[bool] = True,
+        norm_num_groups: Optional[int] = None,
+        spatial_norm_dim: Optional[int] = None,
+        out_bias: bool = True,
+        scale_qk: bool = True,
+        only_cross_attention: bool = False,
+        eps: float = 1e-5,
+        rescale_output_factor: float = 1.0,
+        residual_connection: bool = False,
+        _from_deprecated_attn_block: bool = False,
+        processor: Optional["AttnProcessor"] = None,
+        out_dim: int = None,
+        context_out_dim: int = None,
+        context_pre_only=None,
+        is_final_block=False,
+    ):
+        super().__init__()
+
+        # To prevent circular import.
+        from memo.models.normalization import FP32LayerNorm
+
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads
+        self.query_dim = query_dim
+        self.use_bias = bias
+        self.is_cross_attention = cross_attention_dim is not None
+        self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
+        self.upcast_attention = upcast_attention
+        self.upcast_softmax = upcast_softmax
+        self.rescale_output_factor = rescale_output_factor
+        self.residual_connection = residual_connection
+        self.dropout = dropout
+        self.fused_projections = False
+        self.out_dim = out_dim if out_dim is not None else query_dim
+        self.context_out_dim = context_out_dim if context_out_dim is not None else self.out_dim
+        self.context_pre_only = context_pre_only
+        self.is_final_block = is_final_block
+
+        # we make use of this private variable to know whether this class is loaded
+        # with an deprecated state dict so that we can convert it on the fly
+        self._from_deprecated_attn_block = _from_deprecated_attn_block
+
+        self.scale_qk = scale_qk
+        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
+
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        # for slice_size > 0 the attention score computation
+        # is split across the batch axis to save memory
+        # You can set slice_size with `set_attention_slice`
+        self.sliceable_head_dim = heads
+
+        self.added_kv_proj_dim = added_kv_proj_dim
+        self.only_cross_attention = only_cross_attention
+
+        if self.added_kv_proj_dim is None and self.only_cross_attention:
+            raise ValueError(
+                "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."
+            )
+
+        if norm_num_groups is not None:
+            self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True)
+        else:
+            self.group_norm = None
+
+        if spatial_norm_dim is not None:
+            self.spatial_norm = SpatialNorm(f_channels=query_dim, zq_channels=spatial_norm_dim)
+        else:
+            self.spatial_norm = None
+
+        if qk_norm is None:
+            self.norm_q = None
+            self.norm_k = None
+        elif qk_norm == "layer_norm":
+            self.norm_q = nn.LayerNorm(dim_head, eps=eps)
+            self.norm_k = nn.LayerNorm(dim_head, eps=eps)
+        elif qk_norm == "fp32_layer_norm":
+            self.norm_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
+            self.norm_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
+        elif qk_norm == "layer_norm_across_heads":
+            # Lumina applys qk norm across all heads
+            self.norm_q = nn.LayerNorm(dim_head * heads, eps=eps)
+            self.norm_k = nn.LayerNorm(dim_head * kv_heads, eps=eps)
+        else:
+            raise ValueError(f"unknown qk_norm: {qk_norm}. Should be None or 'layer_norm'")
+
+        if cross_attention_norm is None:
+            self.norm_cross = None
+        elif cross_attention_norm == "layer_norm":
+            self.norm_cross = nn.LayerNorm(self.cross_attention_dim)
+        elif cross_attention_norm == "group_norm":
+            if self.added_kv_proj_dim is not None:
+                # The given `encoder_hidden_states` are initially of shape
+                # (batch_size, seq_len, added_kv_proj_dim) before being projected
+                # to (batch_size, seq_len, cross_attention_dim). The norm is applied
+                # before the projection, so we need to use `added_kv_proj_dim` as
+                # the number of channels for the group norm.
+                norm_cross_num_channels = added_kv_proj_dim
+            else:
+                norm_cross_num_channels = self.cross_attention_dim
+
+            self.norm_cross = nn.GroupNorm(
+                num_channels=norm_cross_num_channels,
+                num_groups=cross_attention_norm_num_groups,
+                eps=1e-5,
+                affine=True,
+            )
+        else:
+            raise ValueError(
+                f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'"
+            )
+
+        self.to_q = nn.Linear(query_dim, self.inner_dim, bias=bias)
+
+        if not self.only_cross_attention:
+            # only relevant for the `AddedKVProcessor` classes
+            self.to_k = nn.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias)
+            self.to_v = nn.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias)
+        else:
+            self.to_k = None
+            self.to_v = None
+
+        if self.added_kv_proj_dim is not None:
+            self.add_k_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias)
+            self.add_v_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias)
+            if self.context_pre_only is not None:
+                self.add_q_proj = nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
+
+        self.to_out = nn.ModuleList([])
+        self.to_out.append(nn.Linear(self.inner_dim, self.out_dim, bias=out_bias))
+        self.to_out.append(nn.Dropout(dropout))
+
+        if self.context_pre_only is not None and not self.context_pre_only and not is_final_block:
+            self.to_add_out = nn.Linear(self.inner_dim, self.context_out_dim, bias=out_bias)
+
+        if qk_norm is not None and added_kv_proj_dim is not None:
+            if qk_norm == "fp32_layer_norm":
+                self.norm_added_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
+                self.norm_added_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
+        else:
+            self.norm_added_q = None
+            self.norm_added_k = None
+
+        # set attention processor
+        # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
+        # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+        # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+        if processor is None:
+            processor = (
+                AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
+            )
+        self.set_processor(processor)
+
+    def set_use_npu_flash_attention(self, use_npu_flash_attention: bool) -> None:
+        r"""
+        Set whether to use npu flash attention from `torch_npu` or not.
+
+        """
+        if use_npu_flash_attention:
+            processor = AttnProcessorNPU()
+        else:
+            # set attention processor
+            # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
+            # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+            # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+            processor = (
+                AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
+            )
+        self.set_processor(processor)
+
+    def set_use_memory_efficient_attention_xformers(
+        self,
+        use_memory_efficient_attention_xformers: bool,
+        attention_op: Optional[Callable] = None,
+    ) -> None:
+        r"""
+        Set whether to use memory efficient attention from `xformers` or not.
+
+        Args:
+            use_memory_efficient_attention_xformers (`bool`):
+                Whether to use memory efficient attention from `xformers` or not.
+            attention_op (`Callable`, *optional*):
+                The attention operation to use. Defaults to `None` which uses the default attention operation from
+                `xformers`.
+        """
+        is_custom_diffusion = hasattr(self, "processor") and isinstance(
+            self.processor,
+            (
+                CustomDiffusionAttnProcessor,
+                CustomDiffusionXFormersAttnProcessor,
+                CustomDiffusionAttnProcessor2_0,
+            ),
+        )
+
+        is_joint_diffusion = hasattr(self, "processor") and isinstance(
+            self.processor,
+            (JointAttnProcessor2_0),
+        )
+        is_added_kv_processor = hasattr(self, "processor") and isinstance(
+            self.processor,
+            (
+                AttnAddedKVProcessor,
+                AttnAddedKVProcessor2_0,
+                SlicedAttnAddedKVProcessor,
+                XFormersAttnAddedKVProcessor,
+            ),
+        )
+
+        if use_memory_efficient_attention_xformers:
+            if is_added_kv_processor and is_custom_diffusion:
+                raise NotImplementedError(
+                    f"Memory efficient attention is currently not supported for custom diffusion for attention processor type {self.processor}"
+                )
+            if not is_xformers_available():
+                raise ModuleNotFoundError(
+                    (
+                        "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+                        " xformers"
+                    ),
+                    name="xformers",
+                )
+            elif not torch.cuda.is_available():
+                raise ValueError(
+                    "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
+                    " only available for GPU "
+                )
+            else:
+                try:
+                    # Make sure we can run the memory efficient attention
+                    _ = xformers.ops.memory_efficient_attention(
+                        torch.randn((1, 2, 40), device="cuda"),
+                        torch.randn((1, 2, 40), device="cuda"),
+                        torch.randn((1, 2, 40), device="cuda"),
+                    )
+                except Exception as e:
+                    raise e
+
+            if is_custom_diffusion:
+                processor = CustomDiffusionXFormersAttnProcessor(
+                    train_kv=self.processor.train_kv,
+                    train_q_out=self.processor.train_q_out,
+                    hidden_size=self.processor.hidden_size,
+                    cross_attention_dim=self.processor.cross_attention_dim,
+                    attention_op=attention_op,
+                )
+                processor.load_state_dict(self.processor.state_dict())
+                if hasattr(self.processor, "to_k_custom_diffusion"):
+                    processor.to(self.processor.to_k_custom_diffusion.weight.device)
+            elif is_added_kv_processor:
+                # TODO(Patrick, Suraj, William) - currently xformers doesn't work for UnCLIP
+                # which uses this type of cross attention ONLY because the attention mask of format
+                # [0, ..., -10.000, ..., 0, ...,] is not supported
+                # throw warning
+                logger.info(
+                    "Memory efficient attention with `xformers` might currently not work correctly if an attention mask is required for the attention operation."
+                )
+                processor = XFormersAttnAddedKVProcessor(attention_op=attention_op)
+            elif is_joint_diffusion:
+                processor = JointAttnProcessor2_0()
+            else:
+                processor = XFormersAttnProcessor(attention_op=attention_op)
+        else:
+            if is_custom_diffusion:
+                attn_processor_class = (
+                    CustomDiffusionAttnProcessor2_0
+                    if hasattr(F, "scaled_dot_product_attention")
+                    else CustomDiffusionAttnProcessor
+                )
+                processor = attn_processor_class(
+                    train_kv=self.processor.train_kv,
+                    train_q_out=self.processor.train_q_out,
+                    hidden_size=self.processor.hidden_size,
+                    cross_attention_dim=self.processor.cross_attention_dim,
+                )
+                processor.load_state_dict(self.processor.state_dict())
+                if hasattr(self.processor, "to_k_custom_diffusion"):
+                    processor.to(self.processor.to_k_custom_diffusion.weight.device)
+            else:
+                # set attention processor
+                # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
+                # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+                # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+                processor = (
+                    AttnProcessor2_0()
+                    if hasattr(F, "scaled_dot_product_attention") and self.scale_qk
+                    else AttnProcessor()
+                )
+
+        self.set_processor(processor)
+
+    def set_attention_slice(self, slice_size: int) -> None:
+        r"""
+        Set the slice size for attention computation.
+
+        Args:
+            slice_size (`int`):
+                The slice size for attention computation.
+        """
+        if slice_size is not None and slice_size > self.sliceable_head_dim:
+            raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")
+
+        if slice_size is not None and self.added_kv_proj_dim is not None:
+            processor = SlicedAttnAddedKVProcessor(slice_size)
+        elif slice_size is not None:
+            processor = SlicedAttnProcessor(slice_size)
+        elif self.added_kv_proj_dim is not None:
+            processor = AttnAddedKVProcessor()
+        else:
+            # set attention processor
+            # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
+            # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+            # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+            processor = (
+                AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
+            )
+
+        self.set_processor(processor)
+
+    def set_processor(self, processor: "AttnProcessor") -> None:
+        r"""
+        Set the attention processor to use.
+
+        Args:
+            processor (`AttnProcessor`):
+                The attention processor to use.
+        """
+        # if current processor is in `self._modules` and if passed `processor` is not, we need to
+        # pop `processor` from `self._modules`
+        if (
+            hasattr(self, "processor")
+            and isinstance(self.processor, torch.nn.Module)
+            and not isinstance(processor, torch.nn.Module)
+        ):
+            logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}")
+            self._modules.pop("processor")
+        self.processor = processor
+
+    def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor":
+        r"""
+        Get the attention processor in use.
+
+        Args:
+            return_deprecated_lora (`bool`, *optional*, defaults to `False`):
+                Set to `True` to return the deprecated LoRA attention processor.
+
+        Returns:
+            "AttentionProcessor": The attention processor in use.
+        """
+        if not return_deprecated_lora:
+            return self.processor
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **cross_attention_kwargs,
+    ) -> torch.Tensor:
+        r"""
+        The forward method of the `Attention` class.
+
+        Args:
+            hidden_states (`torch.Tensor`):
+                The hidden states of the query.
+            encoder_hidden_states (`torch.Tensor`, *optional*):
+                The hidden states of the encoder.
+            attention_mask (`torch.Tensor`, *optional*):
+                The attention mask to use. If `None`, no mask is applied.
+            **cross_attention_kwargs:
+                Additional keyword arguments to pass along to the cross attention.
+
+        Returns:
+            `torch.Tensor`: The output of the attention layer.
+        """
+        # The `Attention` class can call different attention processors / attention functions
+        # here we simply pass along all tensors to the selected processor class
+        # For standard processors that are defined here, `**cross_attention_kwargs` is empty
+
+        attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
+        quiet_attn_parameters = {"ip_adapter_masks"}
+        unused_kwargs = [
+            k for k, _ in cross_attention_kwargs.items() if k not in attn_parameters and k not in quiet_attn_parameters
+        ]
+        if len(unused_kwargs) > 0:
+            logger.warning(
+                f"cross_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored."
+            )
+        cross_attention_kwargs = {k: w for k, w in cross_attention_kwargs.items() if k in attn_parameters}
+
+        return self.processor(
+            self,
+            hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+    def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor:
+        r"""
+        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`. `heads`
+        is the number of heads initialized while constructing the `Attention` class.
+
+        Args:
+            tensor (`torch.Tensor`): The tensor to reshape.
+
+        Returns:
+            `torch.Tensor`: The reshaped tensor.
+        """
+        head_size = self.heads
+        batch_size, seq_len, dim = tensor.shape
+        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
+        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
+        return tensor
+
+    def head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor:
+        r"""
+        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size, seq_len, heads, dim // heads]` `heads` is
+        the number of heads initialized while constructing the `Attention` class.
+
+        Args:
+            tensor (`torch.Tensor`): The tensor to reshape.
+            out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor. If `3`, the tensor is
+                reshaped to `[batch_size * heads, seq_len, dim // heads]`.
+
+        Returns:
+            `torch.Tensor`: The reshaped tensor.
+        """
+        head_size = self.heads
+        if tensor.ndim == 3:
+            batch_size, seq_len, dim = tensor.shape
+            extra_dim = 1
+        else:
+            batch_size, extra_dim, seq_len, dim = tensor.shape
+        tensor = tensor.reshape(batch_size, seq_len * extra_dim, head_size, dim // head_size)
+        tensor = tensor.permute(0, 2, 1, 3)
+
+        if out_dim == 3:
+            tensor = tensor.reshape(batch_size * head_size, seq_len * extra_dim, dim // head_size)
+
+        return tensor
+
+    def get_attention_scores(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        attention_mask: torch.Tensor = None,
+    ) -> torch.Tensor:
+        r"""
+        Compute the attention scores.
+
+        Args:
+            query (`torch.Tensor`): The query tensor.
+            key (`torch.Tensor`): The key tensor.
+            attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied.
+
+        Returns:
+            `torch.Tensor`: The attention probabilities/scores.
+        """
+        dtype = query.dtype
+        if self.upcast_attention:
+            query = query.float()
+            key = key.float()
+
+        if attention_mask is None:
+            baddbmm_input = torch.empty(
+                query.shape[0],
+                query.shape[1],
+                key.shape[1],
+                dtype=query.dtype,
+                device=query.device,
+            )
+            beta = 0
+        else:
+            baddbmm_input = attention_mask
+            beta = 1
+
+        attention_scores = torch.baddbmm(
+            baddbmm_input,
+            query,
+            key.transpose(-1, -2),
+            beta=beta,
+            alpha=self.scale,
+        )
+        del baddbmm_input
+
+        if self.upcast_softmax:
+            attention_scores = attention_scores.float()
+
+        attention_probs = attention_scores.softmax(dim=-1)
+        del attention_scores
+
+        attention_probs = attention_probs.to(dtype)
+
+        return attention_probs
+
+    def prepare_attention_mask(
+        self,
+        attention_mask: torch.Tensor,
+        target_length: int,
+        batch_size: int,
+        out_dim: int = 3,
+    ) -> torch.Tensor:
+        r"""
+        Prepare the attention mask for the attention computation.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                The attention mask to prepare.
+            target_length (`int`):
+                The target length of the attention mask. This is the length of the attention mask after padding.
+            batch_size (`int`):
+                The batch size, which is used to repeat the attention mask.
+            out_dim (`int`, *optional*, defaults to `3`):
+                The output dimension of the attention mask. Can be either `3` or `4`.
+
+        Returns:
+            `torch.Tensor`: The prepared attention mask.
+        """
+        head_size = self.heads
+        if attention_mask is None:
+            return attention_mask
+
+        current_length: int = attention_mask.shape[-1]
+        if current_length != target_length:
+            if attention_mask.device.type == "mps":
+                # HACK: MPS: Does not support padding by greater than dimension of input tensor.
+                # Instead, we can manually construct the padding tensor.
+                padding_shape = (
+                    attention_mask.shape[0],
+                    attention_mask.shape[1],
+                    target_length,
+                )
+                padding = torch.zeros(
+                    padding_shape,
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device,
+                )
+                attention_mask = torch.cat([attention_mask, padding], dim=2)
+            else:
+                # TODO: for pipelines such as stable-diffusion, padding cross-attn mask:
+                #       we want to instead pad by (0, remaining_length), where remaining_length is:
+                #       remaining_length: int = target_length - current_length
+                # TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding
+                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
+
+        if out_dim == 3:
+            if attention_mask.shape[0] < batch_size * head_size:
+                attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
+        elif out_dim == 4:
+            attention_mask = attention_mask.unsqueeze(1)
+            attention_mask = attention_mask.repeat_interleave(head_size, dim=1)
+
+        return attention_mask
+
+    def norm_encoder_hidden_states(self, encoder_hidden_states: torch.Tensor) -> torch.Tensor:
+        r"""
+        Normalize the encoder hidden states. Requires `self.norm_cross` to be specified when constructing the
+        `Attention` class.
+
+        Args:
+            encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder.
+
+        Returns:
+            `torch.Tensor`: The normalized encoder hidden states.
+        """
+        assert self.norm_cross is not None, "self.norm_cross must be defined to call self.norm_encoder_hidden_states"
+
+        if isinstance(self.norm_cross, nn.LayerNorm):
+            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
+        elif isinstance(self.norm_cross, nn.GroupNorm):
+            # Group norm norms along the channels dimension and expects
+            # input to be in the shape of (N, C, *). In this case, we want
+            # to norm along the hidden dimension, so we need to move
+            # (batch_size, sequence_length, hidden_size) ->
+            # (batch_size, hidden_size, sequence_length)
+            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
+            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
+        else:
+            assert False
+
+        return encoder_hidden_states
+
+    @torch.no_grad()
+    def fuse_projections(self, fuse=True):
+        device = self.to_q.weight.data.device
+        dtype = self.to_q.weight.data.dtype
+
+        if not self.is_cross_attention:
+            # fetch weight matrices.
+            concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data])
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            # create a new single projection layer and copy over the weights.
+            self.to_qkv = nn.Linear(
+                in_features,
+                out_features,
+                bias=self.use_bias,
+                device=device,
+                dtype=dtype,
+            )
+            self.to_qkv.weight.copy_(concatenated_weights)
+            if self.use_bias:
+                concatenated_bias = torch.cat([self.to_q.bias.data, self.to_k.bias.data, self.to_v.bias.data])
+                self.to_qkv.bias.copy_(concatenated_bias)
+
+        else:
+            concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data])
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            self.to_kv = nn.Linear(
+                in_features,
+                out_features,
+                bias=self.use_bias,
+                device=device,
+                dtype=dtype,
+            )
+            self.to_kv.weight.copy_(concatenated_weights)
+            if self.use_bias:
+                concatenated_bias = torch.cat([self.to_k.bias.data, self.to_v.bias.data])
+                self.to_kv.bias.copy_(concatenated_bias)
+
+        self.fused_projections = fuse
+
+
+class AttnProcessor:
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        temb: Optional[torch.Tensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class MemoryLinearAttnProcessor:
+    r"""
+    Processor for performing linear attention-related computations.
+    """
+
+    def __init__(self):
+        self.memory = {"KV": None, "Z": None}
+        self.decay = 0.9
+
+    def reset_memory_state(self):
+        """Reset memory to the initial state."""
+        self.memory = {"KV": None, "Z": None}
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        motion_frames: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        temb: Optional[torch.Tensor] = None,
+        is_new_audio: bool = True,
+        update_past_memory: bool = False,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        # Reset memory if it's a new data segment
+        # Need to modify inference code
+        if is_new_audio:
+            self.reset_memory_state()
+
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+            with torch.no_grad():
+                motion_frames = attn.spatial_norm(motion_frames, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+            with torch.no_grad():
+                (
+                    motion_frames_batch_size,
+                    motion_frames_channel,
+                    motion_frames_height,
+                    motion_frames_width,
+                ) = motion_frames.shape
+                motion_frames = motion_frames.view(
+                    motion_frames_batch_size,
+                    motion_frames_channel,
+                    motion_frames_height * motion_frames_width,
+                ).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+            with torch.no_grad():
+                motion_frames = attn.group_norm(motion_frames.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        with torch.no_grad():
+            motion_frames_query = attn.to_q(motion_frames)
+            motion_frames_key = attn.to_k(motion_frames)
+            motion_frames_value = attn.to_v(motion_frames)
+
+            motion_frames_query = attn.head_to_batch_dim(motion_frames_query)
+            motion_frames_key = attn.head_to_batch_dim(motion_frames_key)
+            motion_frames_value = attn.head_to_batch_dim(motion_frames_value)
+
+        query = torch.softmax(query, dim=-1)
+        key = torch.softmax(key, dim=-2)
+
+        with torch.no_grad():
+            motion_frames_key = torch.softmax(motion_frames_key, dim=-2)
+
+        # Compute linear attention using the new formulation
+        query = query * attn.scale
+
+        # Update Memory
+        if update_past_memory or is_new_audio:
+            with torch.no_grad():
+                # frame-level decay for memory update
+                seq_length = motion_frames_key.size(1)
+                decay_factors = self.decay ** torch.arange(
+                    seq_length - 1,
+                    -1,
+                    -1,
+                    device=motion_frames_key.device,
+                    dtype=motion_frames_key.dtype,
+                )  # [0, 1, ..., seq_length - 1]
+                decay_factors = decay_factors.view(1, seq_length, 1)
+                decayed_motion_frames_key = motion_frames_key * decay_factors
+                decayed_motion_frames_value = motion_frames_value * decay_factors
+
+                batch_size, seq_length, _ = decayed_motion_frames_key.shape
+                keys_unsqueezed = decayed_motion_frames_key.unsqueeze(3)
+                values_unsqueezed = decayed_motion_frames_value.unsqueeze(2)
+
+                KV_t_all = keys_unsqueezed * values_unsqueezed
+                KV_cumsum = KV_t_all.sum(dim=1)
+
+                Z_cumsum = decayed_motion_frames_key.sum(dim=1)
+
+                if self.memory["KV"] is None and self.memory["Z"] is None:
+                    self.memory["KV"] = KV_cumsum
+                    self.memory["Z"] = Z_cumsum.unsqueeze(1)  # [batch_size, 1, d_model]
+                else:
+                    self.memory["KV"] = self.memory["KV"] * (self.decay**seq_length) + KV_cumsum
+                    self.memory["Z"] = self.memory["Z"] * (self.decay**seq_length) + Z_cumsum.unsqueeze(
+                        1
+                    )  # [batch_size, seq_length, d_model]
+
+        KV = self.decay * self.memory["KV"] + torch.einsum("bnd,bne->bde", key, value)
+        Z = self.decay * self.memory["Z"] + key.sum(dim=-2, keepdim=True)
+
+        # Compute Linear Attn
+        query_KV = torch.einsum("bnd,bde->bne", query, KV)
+        query_Z = torch.einsum("bnd,bod->bno", query, Z)
+
+        hidden_states = query_KV / (query_Z.clamp(min=1e-10))
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class CustomDiffusionAttnProcessor(nn.Module):
+    r"""
+    Processor for implementing attention for the Custom Diffusion method.
+
+    Args:
+        train_kv (`bool`, defaults to `True`):
+            Whether to newly train the key and value matrices corresponding to the text features.
+        train_q_out (`bool`, defaults to `True`):
+            Whether to newly train query matrices corresponding to the latent image features.
+        hidden_size (`int`, *optional*, defaults to `None`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`, *optional*, defaults to `None`):
+            The number of channels in the `encoder_hidden_states`.
+        out_bias (`bool`, defaults to `True`):
+            Whether to include the bias parameter in `train_q_out`.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability to use.
+    """
+
+    def __init__(
+        self,
+        train_kv: bool = True,
+        train_q_out: bool = True,
+        hidden_size: Optional[int] = None,
+        cross_attention_dim: Optional[int] = None,
+        out_bias: bool = True,
+        dropout: float = 0.0,
+    ):
+        super().__init__()
+        self.train_kv = train_kv
+        self.train_q_out = train_q_out
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+
+        # `_custom_diffusion` id for easy serialization and loading.
+        if self.train_kv:
+            self.to_k_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+            self.to_v_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        if self.train_q_out:
+            self.to_q_custom_diffusion = nn.Linear(hidden_size, hidden_size, bias=False)
+            self.to_out_custom_diffusion = nn.ModuleList([])
+            self.to_out_custom_diffusion.append(nn.Linear(hidden_size, hidden_size, bias=out_bias))
+            self.to_out_custom_diffusion.append(nn.Dropout(dropout))
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, sequence_length, _ = hidden_states.shape
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+        if self.train_q_out:
+            query = self.to_q_custom_diffusion(hidden_states).to(attn.to_q.weight.dtype)
+        else:
+            query = attn.to_q(hidden_states.to(attn.to_q.weight.dtype))
+
+        if encoder_hidden_states is None:
+            crossattn = False
+            encoder_hidden_states = hidden_states
+        else:
+            crossattn = True
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        if self.train_kv:
+            key = self.to_k_custom_diffusion(encoder_hidden_states.to(self.to_k_custom_diffusion.weight.dtype))
+            value = self.to_v_custom_diffusion(encoder_hidden_states.to(self.to_v_custom_diffusion.weight.dtype))
+            key = key.to(attn.to_q.weight.dtype)
+            value = value.to(attn.to_q.weight.dtype)
+        else:
+            key = attn.to_k(encoder_hidden_states)
+            value = attn.to_v(encoder_hidden_states)
+
+        if crossattn:
+            detach = torch.ones_like(key)
+            detach[:, :1, :] = detach[:, :1, :] * 0.0
+            key = detach * key + (1 - detach) * key.detach()
+            value = detach * value + (1 - detach) * value.detach()
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        if self.train_q_out:
+            # linear proj
+            hidden_states = self.to_out_custom_diffusion[0](hidden_states)
+            # dropout
+            hidden_states = self.to_out_custom_diffusion[1](hidden_states)
+        else:
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            hidden_states = attn.to_out[1](hidden_states)
+
+        return hidden_states
+
+
+class AttnAddedKVProcessor:
+    r"""
+    Processor for performing attention-related computations with extra learnable key and value matrices for the text
+    encoder.
+    """
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        residual = hidden_states
+
+        hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+        query = attn.head_to_batch_dim(query)
+
+        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+        encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj)
+        encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj)
+
+        if not attn.only_cross_attention:
+            key = attn.to_k(hidden_states)
+            value = attn.to_v(hidden_states)
+            key = attn.head_to_batch_dim(key)
+            value = attn.head_to_batch_dim(value)
+            key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)
+            value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)
+        else:
+            key = encoder_hidden_states_key_proj
+            value = encoder_hidden_states_value_proj
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
+        hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
+class AttnAddedKVProcessor2_0:
+    r"""
+    Processor for performing scaled dot-product attention (enabled by default if you're using PyTorch 2.0), with extra
+    learnable key and value matrices for the text encoder.
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "AttnAddedKVProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        residual = hidden_states
+
+        hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size, out_dim=4)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+        query = attn.head_to_batch_dim(query, out_dim=4)
+
+        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+        encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj, out_dim=4)
+        encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj, out_dim=4)
+
+        if not attn.only_cross_attention:
+            key = attn.to_k(hidden_states)
+            value = attn.to_v(hidden_states)
+            key = attn.head_to_batch_dim(key, out_dim=4)
+            value = attn.head_to_batch_dim(value, out_dim=4)
+            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
+            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
+        else:
+            key = encoder_hidden_states_key_proj
+            value = encoder_hidden_states_value_proj
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, residual.shape[1])
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
+        hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
+class JointAttnProcessor2_0:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        input_ndim = hidden_states.ndim
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+        context_input_ndim = encoder_hidden_states.ndim
+        if context_input_ndim == 4:
+            batch_size, channel, height, width = encoder_hidden_states.shape
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size = encoder_hidden_states.shape[0]
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
+        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+        query = torch.cat([query, encoder_hidden_states_query_proj], dim=1)
+        key = torch.cat([key, encoder_hidden_states_key_proj], dim=1)
+        value = torch.cat([value, encoder_hidden_states_value_proj], dim=1)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # Split the attention outputs.
+        hidden_states, encoder_hidden_states = (
+            hidden_states[:, : residual.shape[1]],
+            hidden_states[:, residual.shape[1] :],
+        )
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        if not attn.context_pre_only and not attn.is_final_block:
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+        else:
+            encoder_hidden_states = None
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        return hidden_states, encoder_hidden_states
+
+
+class FusedJointAttnProcessor2_0:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        input_ndim = hidden_states.ndim
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+        context_input_ndim = encoder_hidden_states.ndim
+        if context_input_ndim == 4:
+            batch_size, channel, height, width = encoder_hidden_states.shape
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size = encoder_hidden_states.shape[0]
+
+        # `sample` projections.
+        qkv = attn.to_qkv(hidden_states)
+        split_size = qkv.shape[-1] // 3
+        query, key, value = torch.split(qkv, split_size, dim=-1)
+
+        # `context` projections.
+        encoder_qkv = attn.to_added_qkv(encoder_hidden_states)
+        split_size = encoder_qkv.shape[-1] // 3
+        (
+            encoder_hidden_states_query_proj,
+            encoder_hidden_states_key_proj,
+            encoder_hidden_states_value_proj,
+        ) = torch.split(encoder_qkv, split_size, dim=-1)
+
+        # attention
+        query = torch.cat([query, encoder_hidden_states_query_proj], dim=1)
+        key = torch.cat([key, encoder_hidden_states_key_proj], dim=1)
+        value = torch.cat([value, encoder_hidden_states_value_proj], dim=1)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # Split the attention outputs.
+        hidden_states, encoder_hidden_states = (
+            hidden_states[:, : residual.shape[1]],
+            hidden_states[:, residual.shape[1] :],
+        )
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+        if not attn.context_pre_only:
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+        if context_input_ndim == 4:
+            encoder_hidden_states = encoder_hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        return hidden_states, encoder_hidden_states
+
+
+class XFormersAttnAddedKVProcessor:
+    r"""
+    Processor for implementing memory efficient attention using xFormers.
+
+    Args:
+        attention_op (`Callable`, *optional*, defaults to `None`):
+            The base
+            [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
+            use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
+            operator.
+    """
+
+    def __init__(self, attention_op: Optional[Callable] = None):
+        self.attention_op = attention_op
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+        query = attn.head_to_batch_dim(query)
+
+        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+        encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj)
+        encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj)
+
+        if not attn.only_cross_attention:
+            key = attn.to_k(hidden_states)
+            value = attn.to_v(hidden_states)
+            key = attn.head_to_batch_dim(key)
+            value = attn.head_to_batch_dim(value)
+            key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)
+            value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)
+        else:
+            key = encoder_hidden_states_key_proj
+            value = encoder_hidden_states_value_proj
+
+        hidden_states = xformers.ops.memory_efficient_attention(
+            query,
+            key,
+            value,
+            attn_bias=attention_mask,
+            op=self.attention_op,
+            scale=attn.scale,
+        )
+        hidden_states = hidden_states.to(query.dtype)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
+        hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
+class XFormersAttnProcessor:
+    r"""
+    Processor for implementing memory efficient attention using xFormers.
+
+    Args:
+        attention_op (`Callable`, *optional*, defaults to `None`):
+            The base
+            [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
+            use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
+            operator.
+    """
+
+    def __init__(self, attention_op: Optional[Callable] = None):
+        self.attention_op = attention_op
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        temb: Optional[torch.Tensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, key_tokens, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        attention_mask = attn.prepare_attention_mask(attention_mask, key_tokens, batch_size)
+        if attention_mask is not None:
+            # expand our mask's singleton query_tokens dimension:
+            #   [batch*heads,            1, key_tokens] ->
+            #   [batch*heads, query_tokens, key_tokens]
+            # so that it can be added as a bias onto the attention scores that xformers computes:
+            #   [batch*heads, query_tokens, key_tokens]
+            # we do this explicitly because xformers doesn't broadcast the singleton dimension for us.
+            _, query_tokens, _ = hidden_states.shape
+            attention_mask = attention_mask.expand(-1, query_tokens, -1)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query).contiguous()
+        key = attn.head_to_batch_dim(key).contiguous()
+        value = attn.head_to_batch_dim(value).contiguous()
+
+        hidden_states = xformers.ops.memory_efficient_attention(
+            query,
+            key,
+            value,
+            attn_bias=attention_mask,
+            op=self.attention_op,
+            scale=attn.scale,
+        )
+        hidden_states = hidden_states.to(query.dtype)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessorNPU:
+    r"""
+    Processor for implementing flash attention using torch_npu. Torch_npu supports only fp16 and bf16 data types. If
+    fp32 is used, F.scaled_dot_product_attention will be used for computation, but the acceleration effect on NPU is
+    not significant.
+
+    """
+
+    def __init__(self):
+        if not is_torch_npu_available():
+            raise ImportError("AttnProcessorNPU requires torch_npu extensions and is supported only on npu devices.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        temb: Optional[torch.Tensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        residual = hidden_states
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        if query.dtype in (torch.float16, torch.bfloat16):
+            hidden_states = torch_npu.npu_fusion_attention(
+                query,
+                key,
+                value,
+                attn.heads,
+                input_layout="BNSD",
+                pse=None,
+                atten_mask=attention_mask,
+                scale=1.0 / math.sqrt(query.shape[-1]),
+                pre_tockens=65536,
+                next_tockens=65536,
+                keep_prob=1.0,
+                sync=False,
+                inner_precise=0,
+            )[0]
+        else:
+            # TODO: add support for attn.scale when we move to Torch 2.1
+            hidden_states = F.scaled_dot_product_attention(
+                query,
+                key,
+                value,
+                attn_mask=attention_mask,
+                dropout_p=0.0,
+                is_causal=False,
+            )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        temb: Optional[torch.Tensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        residual = hidden_states
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class FusedAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). It uses
+    fused projection layers. For self-attention modules, all projection matrices (i.e., query, key, value) are fused.
+    For cross-attention modules, key and value projection matrices are fused.
+
+    <Tip warning={true}>
+
+    This API is currently 🧪 experimental in nature and can change in future.
+
+    </Tip>
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "FusedAttnProcessor2_0 requires at least PyTorch 2.0, to use it. Please upgrade PyTorch to > 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        temb: Optional[torch.Tensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        residual = hidden_states
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        if encoder_hidden_states is None:
+            qkv = attn.to_qkv(hidden_states)
+            split_size = qkv.shape[-1] // 3
+            query, key, value = torch.split(qkv, split_size, dim=-1)
+        else:
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+            query = attn.to_q(hidden_states)
+
+            kv = attn.to_kv(encoder_hidden_states)
+            split_size = kv.shape[-1] // 2
+            key, value = torch.split(kv, split_size, dim=-1)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class CustomDiffusionXFormersAttnProcessor(nn.Module):
+    r"""
+    Processor for implementing memory efficient attention using xFormers for the Custom Diffusion method.
+
+    Args:
+    train_kv (`bool`, defaults to `True`):
+        Whether to newly train the key and value matrices corresponding to the text features.
+    train_q_out (`bool`, defaults to `True`):
+        Whether to newly train query matrices corresponding to the latent image features.
+    hidden_size (`int`, *optional*, defaults to `None`):
+        The hidden size of the attention layer.
+    cross_attention_dim (`int`, *optional*, defaults to `None`):
+        The number of channels in the `encoder_hidden_states`.
+    out_bias (`bool`, defaults to `True`):
+        Whether to include the bias parameter in `train_q_out`.
+    dropout (`float`, *optional*, defaults to 0.0):
+        The dropout probability to use.
+    attention_op (`Callable`, *optional*, defaults to `None`):
+        The base
+        [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to use
+        as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best operator.
+    """
+
+    def __init__(
+        self,
+        train_kv: bool = True,
+        train_q_out: bool = False,
+        hidden_size: Optional[int] = None,
+        cross_attention_dim: Optional[int] = None,
+        out_bias: bool = True,
+        dropout: float = 0.0,
+        attention_op: Optional[Callable] = None,
+    ):
+        super().__init__()
+        self.train_kv = train_kv
+        self.train_q_out = train_q_out
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.attention_op = attention_op
+
+        # `_custom_diffusion` id for easy serialization and loading.
+        if self.train_kv:
+            self.to_k_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+            self.to_v_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        if self.train_q_out:
+            self.to_q_custom_diffusion = nn.Linear(hidden_size, hidden_size, bias=False)
+            self.to_out_custom_diffusion = nn.ModuleList([])
+            self.to_out_custom_diffusion.append(nn.Linear(hidden_size, hidden_size, bias=out_bias))
+            self.to_out_custom_diffusion.append(nn.Dropout(dropout))
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if self.train_q_out:
+            query = self.to_q_custom_diffusion(hidden_states).to(attn.to_q.weight.dtype)
+        else:
+            query = attn.to_q(hidden_states.to(attn.to_q.weight.dtype))
+
+        if encoder_hidden_states is None:
+            crossattn = False
+            encoder_hidden_states = hidden_states
+        else:
+            crossattn = True
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        if self.train_kv:
+            key = self.to_k_custom_diffusion(encoder_hidden_states.to(self.to_k_custom_diffusion.weight.dtype))
+            value = self.to_v_custom_diffusion(encoder_hidden_states.to(self.to_v_custom_diffusion.weight.dtype))
+            key = key.to(attn.to_q.weight.dtype)
+            value = value.to(attn.to_q.weight.dtype)
+        else:
+            key = attn.to_k(encoder_hidden_states)
+            value = attn.to_v(encoder_hidden_states)
+
+        if crossattn:
+            detach = torch.ones_like(key)
+            detach[:, :1, :] = detach[:, :1, :] * 0.0
+            key = detach * key + (1 - detach) * key.detach()
+            value = detach * value + (1 - detach) * value.detach()
+
+        query = attn.head_to_batch_dim(query).contiguous()
+        key = attn.head_to_batch_dim(key).contiguous()
+        value = attn.head_to_batch_dim(value).contiguous()
+
+        hidden_states = xformers.ops.memory_efficient_attention(
+            query,
+            key,
+            value,
+            attn_bias=attention_mask,
+            op=self.attention_op,
+            scale=attn.scale,
+        )
+        hidden_states = hidden_states.to(query.dtype)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        if self.train_q_out:
+            # linear proj
+            hidden_states = self.to_out_custom_diffusion[0](hidden_states)
+            # dropout
+            hidden_states = self.to_out_custom_diffusion[1](hidden_states)
+        else:
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            hidden_states = attn.to_out[1](hidden_states)
+
+        return hidden_states
+
+
+class CustomDiffusionAttnProcessor2_0(nn.Module):
+    r"""
+    Processor for implementing attention for the Custom Diffusion method using PyTorch 2.0’s memory-efficient scaled
+    dot-product attention.
+
+    Args:
+        train_kv (`bool`, defaults to `True`):
+            Whether to newly train the key and value matrices corresponding to the text features.
+        train_q_out (`bool`, defaults to `True`):
+            Whether to newly train query matrices corresponding to the latent image features.
+        hidden_size (`int`, *optional*, defaults to `None`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`, *optional*, defaults to `None`):
+            The number of channels in the `encoder_hidden_states`.
+        out_bias (`bool`, defaults to `True`):
+            Whether to include the bias parameter in `train_q_out`.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability to use.
+    """
+
+    def __init__(
+        self,
+        train_kv: bool = True,
+        train_q_out: bool = True,
+        hidden_size: Optional[int] = None,
+        cross_attention_dim: Optional[int] = None,
+        out_bias: bool = True,
+        dropout: float = 0.0,
+    ):
+        super().__init__()
+        self.train_kv = train_kv
+        self.train_q_out = train_q_out
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+
+        # `_custom_diffusion` id for easy serialization and loading.
+        if self.train_kv:
+            self.to_k_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+            self.to_v_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        if self.train_q_out:
+            self.to_q_custom_diffusion = nn.Linear(hidden_size, hidden_size, bias=False)
+            self.to_out_custom_diffusion = nn.ModuleList([])
+            self.to_out_custom_diffusion.append(nn.Linear(hidden_size, hidden_size, bias=out_bias))
+            self.to_out_custom_diffusion.append(nn.Dropout(dropout))
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, sequence_length, _ = hidden_states.shape
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+        if self.train_q_out:
+            query = self.to_q_custom_diffusion(hidden_states)
+        else:
+            query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            crossattn = False
+            encoder_hidden_states = hidden_states
+        else:
+            crossattn = True
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        if self.train_kv:
+            key = self.to_k_custom_diffusion(encoder_hidden_states.to(self.to_k_custom_diffusion.weight.dtype))
+            value = self.to_v_custom_diffusion(encoder_hidden_states.to(self.to_v_custom_diffusion.weight.dtype))
+            key = key.to(attn.to_q.weight.dtype)
+            value = value.to(attn.to_q.weight.dtype)
+
+        else:
+            key = attn.to_k(encoder_hidden_states)
+            value = attn.to_v(encoder_hidden_states)
+
+        if crossattn:
+            detach = torch.ones_like(key)
+            detach[:, :1, :] = detach[:, :1, :] * 0.0
+            key = detach * key + (1 - detach) * key.detach()
+            value = detach * value + (1 - detach) * value.detach()
+
+        inner_dim = hidden_states.shape[-1]
+
+        head_dim = inner_dim // attn.heads
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if self.train_q_out:
+            # linear proj
+            hidden_states = self.to_out_custom_diffusion[0](hidden_states)
+            # dropout
+            hidden_states = self.to_out_custom_diffusion[1](hidden_states)
+        else:
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            hidden_states = attn.to_out[1](hidden_states)
+
+        return hidden_states
+
+
+class SlicedAttnProcessor:
+    r"""
+    Processor for implementing sliced attention.
+
+    Args:
+        slice_size (`int`, *optional*):
+            The number of steps to compute attention. Uses as many slices as `attention_head_dim // slice_size`, and
+            `attention_head_dim` must be a multiple of the `slice_size`.
+    """
+
+    def __init__(self, slice_size: int):
+        self.slice_size = slice_size
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        residual = hidden_states
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+        dim = query.shape[-1]
+        query = attn.head_to_batch_dim(query)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        batch_size_attention, query_tokens, _ = query.shape
+        hidden_states = torch.zeros(
+            (batch_size_attention, query_tokens, dim // attn.heads),
+            device=query.device,
+            dtype=query.dtype,
+        )
+
+        for i in range(batch_size_attention // self.slice_size):
+            start_idx = i * self.slice_size
+            end_idx = (i + 1) * self.slice_size
+
+            query_slice = query[start_idx:end_idx]
+            key_slice = key[start_idx:end_idx]
+            attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None
+
+            attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
+
+            attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
+
+            hidden_states[start_idx:end_idx] = attn_slice
+
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class SlicedAttnAddedKVProcessor:
+    r"""
+    Processor for implementing sliced attention with extra learnable key and value matrices for the text encoder.
+
+    Args:
+        slice_size (`int`, *optional*):
+            The number of steps to compute attention. Uses as many slices as `attention_head_dim // slice_size`, and
+            `attention_head_dim` must be a multiple of the `slice_size`.
+    """
+
+    def __init__(self, slice_size):
+        self.slice_size = slice_size
+
+    def __call__(
+        self,
+        attn: "Attention",
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        temb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
+
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+        dim = query.shape[-1]
+        query = attn.head_to_batch_dim(query)
+
+        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+        encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj)
+        encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj)
+
+        if not attn.only_cross_attention:
+            key = attn.to_k(hidden_states)
+            value = attn.to_v(hidden_states)
+            key = attn.head_to_batch_dim(key)
+            value = attn.head_to_batch_dim(value)
+            key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)
+            value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)
+        else:
+            key = encoder_hidden_states_key_proj
+            value = encoder_hidden_states_value_proj
+
+        batch_size_attention, query_tokens, _ = query.shape
+        hidden_states = torch.zeros(
+            (batch_size_attention, query_tokens, dim // attn.heads),
+            device=query.device,
+            dtype=query.dtype,
+        )
+
+        for i in range(batch_size_attention // self.slice_size):
+            start_idx = i * self.slice_size
+            end_idx = (i + 1) * self.slice_size
+
+            query_slice = query[start_idx:end_idx]
+            key_slice = key[start_idx:end_idx]
+            attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None
+
+            attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
+
+            attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
+
+            hidden_states[start_idx:end_idx] = attn_slice
+
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
+        hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
+class SpatialNorm(nn.Module):
+    """
+    Spatially conditioned normalization as defined in https://arxiv.org/abs/2209.09002.
+
+    Args:
+        f_channels (`int`):
+            The number of channels for input to group normalization layer, and output of the spatial norm layer.
+        zq_channels (`int`):
+            The number of channels for the quantized vector as described in the paper.
+    """
+
+    def __init__(
+        self,
+        f_channels: int,
+        zq_channels: int,
+    ):
+        super().__init__()
+        self.norm_layer = nn.GroupNorm(num_channels=f_channels, num_groups=32, eps=1e-6, affine=True)
+        self.conv_y = nn.Conv2d(zq_channels, f_channels, kernel_size=1, stride=1, padding=0)
+        self.conv_b = nn.Conv2d(zq_channels, f_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, f: torch.Tensor, zq: torch.Tensor) -> torch.Tensor:
+        f_size = f.shape[-2:]
+        zq = F.interpolate(zq, size=f_size, mode="nearest")
+        norm_f = self.norm_layer(f)
+        new_f = norm_f * self.conv_y(zq) + self.conv_b(zq)
+        return new_f
+
+
+ADDED_KV_ATTENTION_PROCESSORS = (
+    AttnAddedKVProcessor,
+    SlicedAttnAddedKVProcessor,
+    AttnAddedKVProcessor2_0,
+    XFormersAttnAddedKVProcessor,
+)
+
+CROSS_ATTENTION_PROCESSORS = (
+    AttnProcessor,
+    AttnProcessor2_0,
+    XFormersAttnProcessor,
+    SlicedAttnProcessor,
+)
+
+AttentionProcessor = Union[
+    AttnProcessor,
+    AttnProcessor2_0,
+    FusedAttnProcessor2_0,
+    XFormersAttnProcessor,
+    SlicedAttnProcessor,
+    AttnAddedKVProcessor,
+    SlicedAttnAddedKVProcessor,
+    AttnAddedKVProcessor2_0,
+    XFormersAttnAddedKVProcessor,
+    CustomDiffusionAttnProcessor,
+    CustomDiffusionXFormersAttnProcessor,
+    CustomDiffusionAttnProcessor2_0,
+]

memo/models/audio_proj.py

@@ -0,0 +1,48 @@
+import torch
+from diffusers import ConfigMixin, ModelMixin
+from einops import rearrange
+from torch import nn
+
+
+class AudioProjModel(ModelMixin, ConfigMixin):
+    def __init__(
+        self,
+        seq_len=5,
+        blocks=12,  # add a new parameter blocks
+        channels=768,  # add a new parameter channels
+        intermediate_dim=512,
+        output_dim=768,
+        context_tokens=32,
+    ):
+        super().__init__()
+
+        self.seq_len = seq_len
+        self.blocks = blocks
+        self.channels = channels
+        self.input_dim = seq_len * blocks * channels  # update input_dim to be the product of blocks and channels.
+        self.intermediate_dim = intermediate_dim
+        self.context_tokens = context_tokens
+        self.output_dim = output_dim
+
+        # define multiple linear layers
+        self.proj1 = nn.Linear(self.input_dim, intermediate_dim)
+        self.proj2 = nn.Linear(intermediate_dim, intermediate_dim)
+        self.proj3 = nn.Linear(intermediate_dim, context_tokens * output_dim)
+
+        self.norm = nn.LayerNorm(output_dim)
+
+    def forward(self, audio_embeds):
+        video_length = audio_embeds.shape[1]
+        audio_embeds = rearrange(audio_embeds, "bz f w b c -> (bz f) w b c")
+        batch_size, window_size, blocks, channels = audio_embeds.shape
+        audio_embeds = audio_embeds.view(batch_size, window_size * blocks * channels)
+
+        audio_embeds = torch.relu(self.proj1(audio_embeds))
+        audio_embeds = torch.relu(self.proj2(audio_embeds))
+
+        context_tokens = self.proj3(audio_embeds).reshape(batch_size, self.context_tokens, self.output_dim)
+
+        context_tokens = self.norm(context_tokens)
+        context_tokens = rearrange(context_tokens, "(bz f) m c -> bz f m c", f=video_length)
+
+        return context_tokens

memo/models/emotion_classifier.py

@@ -0,0 +1,31 @@
+from collections import OrderedDict
+
+import torch
+from diffusers import ConfigMixin, ModelMixin
+
+
+class AudioEmotionClassifierModel(ModelMixin, ConfigMixin):
+    num_emotion_classes = 9
+
+    def __init__(self, num_classifier_layers=5, num_classifier_channels=2048):
+        super().__init__()
+
+        if num_classifier_layers == 1:
+            self.layers = torch.nn.Linear(1024, self.num_emotion_classes)
+        else:
+            layer_list = [
+                ("fc1", torch.nn.Linear(1024, num_classifier_channels)),
+                ("relu1", torch.nn.ReLU()),
+            ]
+            for n in range(num_classifier_layers - 2):
+                layer_list.append((f"fc{n+2}", torch.nn.Linear(num_classifier_channels, num_classifier_channels)))
+                layer_list.append((f"relu{n+2}", torch.nn.ReLU()))
+            layer_list.append(
+                (f"fc{num_classifier_layers}", torch.nn.Linear(num_classifier_channels, self.num_emotion_classes))
+            )
+            self.layers = torch.nn.Sequential(OrderedDict(layer_list))
+
+    def forward(self, x):
+        x = self.layers(x)
+        x = torch.softmax(x, dim=-1)
+        return x

memo/models/image_proj.py

@@ -0,0 +1,26 @@
+import torch
+from diffusers import ConfigMixin, ModelMixin
+
+
+class ImageProjModel(ModelMixin, ConfigMixin):
+    def __init__(
+        self,
+        cross_attention_dim=768,
+        clip_embeddings_dim=512,
+        clip_extra_context_tokens=4,
+    ):
+        super().__init__()
+
+        self.generator = None
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = torch.nn.LayerNorm(cross_attention_dim)
+
+    def forward(self, image_embeds):
+        embeds = image_embeds
+        clip_extra_context_tokens = self.proj(embeds).reshape(
+            -1, self.clip_extra_context_tokens, self.cross_attention_dim
+        )
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens

memo/models/motion_module.py

@@ -0,0 +1,386 @@
+import math
+
+import torch
+import xformers
+import xformers.ops
+from diffusers.models.attention import FeedForward
+from diffusers.models.attention_processor import Attention
+from diffusers.utils.import_utils import is_xformers_available
+from einops import rearrange, repeat
+from torch import nn
+
+from memo.models.attention import zero_module
+from memo.models.attention_processor import (
+    MemoryLinearAttnProcessor,
+)
+
+
+class PositionalEncoding(nn.Module):
+    def __init__(self, d_model, dropout=0.0, max_len=24):
+        super().__init__()
+        self.dropout = nn.Dropout(p=dropout)
+        position = torch.arange(max_len).unsqueeze(1)
+        div_term = torch.exp(torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model))
+        pe = torch.zeros(1, max_len, d_model)
+        pe[0, :, 0::2] = torch.sin(position * div_term)
+        pe[0, :, 1::2] = torch.cos(position * div_term)
+        self.register_buffer("pe", pe)
+
+    def forward(self, x, offset=0):
+        x = x + self.pe[:, offset : offset + x.size(1)]
+        return self.dropout(x)
+
+
+class MemoryLinearAttnTemporalModule(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads=8,
+        num_transformer_block=2,
+        attention_block_types=("Temporal_Self", "Temporal_Self"),
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+        temporal_attention_dim_div=1,
+        zero_initialize=True,
+    ):
+        super().__init__()
+
+        self.temporal_transformer = TemporalLinearAttnTransformer(
+            in_channels=in_channels,
+            num_attention_heads=num_attention_heads,
+            attention_head_dim=in_channels // num_attention_heads // temporal_attention_dim_div,
+            num_layers=num_transformer_block,
+            attention_block_types=attention_block_types,
+            temporal_position_encoding=temporal_position_encoding,
+            temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+        )
+
+        if zero_initialize:
+            self.temporal_transformer.proj_out = zero_module(self.temporal_transformer.proj_out)
+
+    def forward(
+        self,
+        hidden_states,
+        motion_frames,
+        encoder_hidden_states,
+        is_new_audio=True,
+        update_past_memory=False,
+    ):
+        hidden_states = self.temporal_transformer(
+            hidden_states,
+            motion_frames,
+            encoder_hidden_states,
+            is_new_audio=is_new_audio,
+            update_past_memory=update_past_memory,
+        )
+
+        output = hidden_states
+        return output
+
+
+class TemporalLinearAttnTransformer(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads,
+        attention_head_dim,
+        num_layers,
+        attention_block_types=(
+            "Temporal_Self",
+            "Temporal_Self",
+        ),
+        dropout=0.0,
+        norm_num_groups=32,
+        cross_attention_dim=768,
+        activation_fn="geglu",
+        attention_bias=False,
+        upcast_attention=False,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+    ):
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+
+        self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+        self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TemporalLinearAttnTransformerBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    attention_block_types=attention_block_types,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    attention_bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+        self.proj_out = nn.Linear(inner_dim, in_channels)
+
+    def forward(
+        self,
+        hidden_states,
+        motion_frames,
+        encoder_hidden_states=None,
+        is_new_audio=True,
+        update_past_memory=False,
+    ):
+        assert hidden_states.dim() == 5, f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        n_motion_frames = motion_frames.shape[2]
+
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+        with torch.no_grad():
+            motion_frames = rearrange(motion_frames, "b c f h w -> (b f) c h w")
+
+        batch, _, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        with torch.no_grad():
+            motion_frames = self.norm(motion_frames)
+
+        inner_dim = hidden_states.shape[1]
+        hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
+        hidden_states = self.proj_in(hidden_states)
+
+        with torch.no_grad():
+            (
+                motion_frames_batch,
+                motion_frames_inner_dim,
+                motion_frames_height,
+                motion_frames_weight,
+            ) = motion_frames.shape
+
+            motion_frames = motion_frames.permute(0, 2, 3, 1).reshape(
+                motion_frames_batch,
+                motion_frames_height * motion_frames_weight,
+                motion_frames_inner_dim,
+            )
+            motion_frames = self.proj_in(motion_frames)
+
+        # Transformer Blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(
+                hidden_states,
+                motion_frames,
+                encoder_hidden_states=encoder_hidden_states,
+                video_length=video_length,
+                n_motion_frames=n_motion_frames,
+                is_new_audio=is_new_audio,
+                update_past_memory=update_past_memory,
+            )
+
+        # output
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+        output = hidden_states + residual
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+
+        return output
+
+
+class TemporalLinearAttnTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_attention_heads,
+        attention_head_dim,
+        attention_block_types=(
+            "Temporal_Self",
+            "Temporal_Self",
+        ),
+        dropout=0.0,
+        cross_attention_dim=768,
+        activation_fn="geglu",
+        attention_bias=False,
+        upcast_attention=False,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+    ):
+        super().__init__()
+
+        attention_blocks = []
+        norms = []
+
+        for block_name in attention_block_types:
+            attention_blocks.append(
+                MemoryLinearAttention(
+                    attention_mode=block_name.split("_", maxsplit=1)[0],
+                    cross_attention_dim=cross_attention_dim if block_name.endswith("_Cross") else None,
+                    query_dim=dim,
+                    heads=num_attention_heads,
+                    dim_head=attention_head_dim,
+                    dropout=dropout,
+                    bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+            )
+            norms.append(nn.LayerNorm(dim))
+
+        self.attention_blocks = nn.ModuleList(attention_blocks)
+        self.norms = nn.ModuleList(norms)
+
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.ff_norm = nn.LayerNorm(dim)
+
+    def forward(
+        self,
+        hidden_states,
+        motion_frames,
+        encoder_hidden_states=None,
+        video_length=None,
+        n_motion_frames=None,
+        is_new_audio=True,
+        update_past_memory=False,
+    ):
+        for attention_block, norm in zip(self.attention_blocks, self.norms):
+            norm_hidden_states = norm(hidden_states)
+            with torch.no_grad():
+                norm_motion_frames = norm(motion_frames)
+            hidden_states = (
+                attention_block(
+                    norm_hidden_states,
+                    norm_motion_frames,
+                    encoder_hidden_states=encoder_hidden_states if attention_block.is_cross_attention else None,
+                    video_length=video_length,
+                    n_motion_frames=n_motion_frames,
+                    is_new_audio=is_new_audio,
+                    update_past_memory=update_past_memory,
+                )
+                + hidden_states
+            )
+
+        hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states
+
+        output = hidden_states
+        return output
+
+
+class MemoryLinearAttention(Attention):
+    def __init__(
+        self,
+        *args,
+        attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        assert attention_mode == "Temporal"
+
+        self.attention_mode = attention_mode
+        self.is_cross_attention = kwargs.get("cross_attention_dim") is not None
+        self.query_dim = kwargs["query_dim"]
+        self.temporal_position_encoding_max_len = temporal_position_encoding_max_len
+        self.pos_encoder = (
+            PositionalEncoding(
+                kwargs["query_dim"],
+                dropout=0.0,
+                max_len=temporal_position_encoding_max_len,
+            )
+            if (temporal_position_encoding and attention_mode == "Temporal")
+            else None
+        )
+
+    def extra_repr(self):
+        return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}"
+
+    def set_use_memory_efficient_attention_xformers(
+        self,
+        use_memory_efficient_attention_xformers: bool,
+        attention_op=None,
+    ):
+        if use_memory_efficient_attention_xformers:
+            if not is_xformers_available():
+                raise ModuleNotFoundError(
+                    (
+                        "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+                        " xformers"
+                    ),
+                    name="xformers",
+                )
+
+            if not torch.cuda.is_available():
+                raise ValueError(
+                    "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
+                    " only available for GPU "
+                )
+
+            try:
+                # Make sure we can run the memory efficient attention
+                _ = xformers.ops.memory_efficient_attention(
+                    torch.randn((1, 2, 40), device="cuda"),
+                    torch.randn((1, 2, 40), device="cuda"),
+                    torch.randn((1, 2, 40), device="cuda"),
+                )
+            except Exception as e:
+                raise e
+            processor = MemoryLinearAttnProcessor()
+        else:
+            processor = MemoryLinearAttnProcessor()
+
+        self.set_processor(processor)
+
+    def forward(
+        self,
+        hidden_states,
+        motion_frames,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        video_length=None,
+        n_motion_frames=None,
+        is_new_audio=True,
+        update_past_memory=False,
+        **cross_attention_kwargs,
+    ):
+        if self.attention_mode == "Temporal":
+            d = hidden_states.shape[1]
+            hidden_states = rearrange(
+                hidden_states,
+                "(b f) d c -> (b d) f c",
+                f=video_length,
+            )
+
+            if self.pos_encoder is not None:
+                hidden_states = self.pos_encoder(hidden_states)
+
+            with torch.no_grad():
+                motion_frames = rearrange(motion_frames, "(b f) d c -> (b d) f c", f=n_motion_frames)
+
+            encoder_hidden_states = (
+                repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d)
+                if encoder_hidden_states is not None
+                else encoder_hidden_states
+            )
+
+        else:
+            raise NotImplementedError
+
+        hidden_states = self.processor(
+            self,
+            hidden_states,
+            motion_frames,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            n_motion_frames=n_motion_frames,
+            is_new_audio=is_new_audio,
+            update_past_memory=update_past_memory,
+            **cross_attention_kwargs,
+        )
+
+        if self.attention_mode == "Temporal":
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

memo/models/normalization.py

@@ -0,0 +1,15 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+class FP32LayerNorm(nn.LayerNorm):
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        origin_dtype = inputs.dtype
+        return F.layer_norm(
+            inputs.float(),
+            self.normalized_shape,
+            self.weight.float() if self.weight is not None else None,
+            self.bias.float() if self.bias is not None else None,
+            self.eps,
+        ).to(origin_dtype)

memo/models/resnet.py

@@ -0,0 +1,218 @@
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+from torch import nn
+
+
+class InflatedConv3d(nn.Conv2d):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class InflatedGroupNorm(nn.GroupNorm):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class Upsample3D(nn.Module):
+    def __init__(
+        self,
+        channels,
+        use_conv=False,
+        use_conv_transpose=False,
+        out_channels=None,
+        name="conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        if use_conv_transpose:
+            raise NotImplementedError
+        if use_conv:
+            self.conv = InflatedConv3d(self.channels, self.out_channels, 3, padding=1)
+
+    def forward(self, hidden_states, output_size=None):
+        assert hidden_states.shape[1] == self.channels
+
+        if self.use_conv_transpose:
+            raise NotImplementedError
+
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
+        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+        if hidden_states.shape[0] >= 64:
+            hidden_states = hidden_states.contiguous()
+
+        # if `output_size` is passed we force the interpolation output
+        # size and do not make use of `scale_factor=2`
+        if output_size is None:
+            hidden_states = F.interpolate(hidden_states, scale_factor=[1.0, 2.0, 2.0], mode="nearest")
+        else:
+            hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")
+
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class Downsample3D(nn.Module):
+    def __init__(self, channels, use_conv=False, out_channels=None, padding=1, name="conv"):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+
+        if use_conv:
+            self.conv = InflatedConv3d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
+        else:
+            raise NotImplementedError
+
+    def forward(self, hidden_states):
+        assert hidden_states.shape[1] == self.channels
+        if self.use_conv and self.padding == 0:
+            raise NotImplementedError
+
+        assert hidden_states.shape[1] == self.channels
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class ResnetBlock3D(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout=0.0,
+        temb_channels=512,
+        groups=32,
+        groups_out=None,
+        pre_norm=True,
+        eps=1e-6,
+        non_linearity="swish",
+        time_embedding_norm="default",
+        output_scale_factor=1.0,
+        use_in_shortcut=None,
+        use_inflated_groupnorm=None,
+    ):
+        super().__init__()
+        self.pre_norm = pre_norm
+        self.pre_norm = True
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.time_embedding_norm = time_embedding_norm
+        self.output_scale_factor = output_scale_factor
+
+        if groups_out is None:
+            groups_out = groups
+
+        assert use_inflated_groupnorm is not None
+        if use_inflated_groupnorm:
+            self.norm1 = InflatedGroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True)
+        else:
+            self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True)
+
+        self.conv1 = InflatedConv3d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+        if temb_channels is not None:
+            if self.time_embedding_norm == "default":
+                time_emb_proj_out_channels = out_channels
+            elif self.time_embedding_norm == "scale_shift":
+                time_emb_proj_out_channels = out_channels * 2
+            else:
+                raise ValueError(f"unknown time_embedding_norm : {self.time_embedding_norm} ")
+
+            self.time_emb_proj = torch.nn.Linear(temb_channels, time_emb_proj_out_channels)
+        else:
+            self.time_emb_proj = None
+
+        if use_inflated_groupnorm:
+            self.norm2 = InflatedGroupNorm(num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True)
+        else:
+            self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = InflatedConv3d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+        if non_linearity == "swish":
+            self.nonlinearity = F.silu()
+        elif non_linearity == "mish":
+            self.nonlinearity = Mish()
+        elif non_linearity == "silu":
+            self.nonlinearity = nn.SiLU()
+
+        self.use_in_shortcut = self.in_channels != self.out_channels if use_in_shortcut is None else use_in_shortcut
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = InflatedConv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, input_tensor, temb):
+        hidden_states = input_tensor
+
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv1(hidden_states)
+
+        if temb is not None:
+            if temb.dim() == 3:
+                temb = self.time_emb_proj(self.nonlinearity(temb))
+                temb = temb.transpose(1, 2).unsqueeze(-1).unsqueeze(-1)
+            else:
+                temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if temb is not None and self.time_embedding_norm == "default":
+            hidden_states = hidden_states + temb
+
+        hidden_states = self.norm2(hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+class Mish(torch.nn.Module):
+    def forward(self, hidden_states):
+        return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))

memo/models/transformer_2d.py

@@ -0,0 +1,280 @@
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
+from torch import nn
+
+from memo.models.attention import BasicTransformerBlock
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+    ref_feature_list: list[torch.FloatTensor]
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        is_final_block: bool = False,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.is_final_block = is_final_block
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+        # 1. Transformer2DModel can process both standard continuous images of
+        # shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of
+        # shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate(
+                "norm_type!=num_embeds_ada_norm",
+                "1.0.0",
+                deprecation_message,
+                standard_warn=False,
+            )
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+
+        if self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+
+        if not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups,
+            num_channels=in_channels,
+            eps=1e-6,
+            affine=True,
+        )
+        if use_linear_projection:
+            self.proj_in = linear_cls(in_channels, inner_dim)
+        else:
+            self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                    is_final_block=(is_final_block and d == num_layers - 1),
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        # TODO: should use out_channels for continuous projections
+        if not is_final_block:
+            if use_linear_projection:
+                self.proj_out = linear_cls(inner_dim, in_channels)
+            else:
+                self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
+
+        self.caption_projection = None
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        if attention_mask is not None and attention_mask.ndim == 2:
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 1. Input
+        batch, _, height, width = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale) if not USE_PEFT_BACKEND else self.proj_in(hidden_states)
+            )
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale) if not USE_PEFT_BACKEND else self.proj_in(hidden_states)
+            )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        ref_feature_list = []
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+
+                        return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states, ref_feature = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, ref_feature = block(
+                    hidden_states,  # shape [5, 4096, 320]
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,  # shape [1,4,768]
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+            ref_feature_list.append(ref_feature)
+
+        # 3. Output
+        output = None
+
+        if self.is_final_block:
+            if not return_dict:
+                return (output, ref_feature_list)
+
+            return Transformer2DModelOutput(sample=output, ref_feature_list=ref_feature_list)
+
+        if self.is_input_continuous:
+            if not self.use_linear_projection:
+                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+            else:
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+            output = hidden_states + residual
+        if not return_dict:
+            return (output, ref_feature_list)
+
+        return Transformer2DModelOutput(sample=output, ref_feature_list=ref_feature_list)

memo/models/transformer_3d.py

@@ -0,0 +1,252 @@
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models import ModelMixin
+from diffusers.utils import BaseOutput
+from einops import rearrange, repeat
+from torch import nn
+
+from memo.models.attention import JointAudioTemporalBasicTransformerBlock, TemporalBasicTransformerBlock
+
+
+def create_custom_forward(module, return_dict=None):
+    def custom_forward(*inputs):
+        if return_dict is not None:
+            return module(*inputs, return_dict=return_dict)
+
+        return module(*inputs)
+
+    return custom_forward
+
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        activation_fn: str = "geglu",
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_audio_module=False,
+        depth=0,
+        unet_block_name=None,
+        emo_drop_rate=0.3,
+        is_final_block=False,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+        self.use_audio_module = use_audio_module
+        # Define input layers
+        self.in_channels = in_channels
+        self.is_final_block = is_final_block
+
+        self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+        if use_linear_projection:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+
+        if use_audio_module:
+            self.transformer_blocks = nn.ModuleList(
+                [
+                    JointAudioTemporalBasicTransformerBlock(
+                        dim=inner_dim,
+                        num_attention_heads=num_attention_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        cross_attention_dim=cross_attention_dim,
+                        activation_fn=activation_fn,
+                        attention_bias=attention_bias,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                        unet_use_temporal_attention=unet_use_temporal_attention,
+                        depth=depth,
+                        unet_block_name=unet_block_name,
+                        use_ada_layer_norm=True,
+                        emo_drop_rate=emo_drop_rate,
+                        is_final_block=(is_final_block and d == num_layers - 1),
+                    )
+                    for d in range(num_layers)
+                ]
+            )
+        else:
+            self.transformer_blocks = nn.ModuleList(
+                [
+                    TemporalBasicTransformerBlock(
+                        inner_dim,
+                        num_attention_heads,
+                        attention_head_dim,
+                        dropout=dropout,
+                        cross_attention_dim=cross_attention_dim,
+                        activation_fn=activation_fn,
+                        attention_bias=attention_bias,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                    )
+                    for _ in range(num_layers)
+                ]
+            )
+
+        # 4. Define output layers
+        if use_linear_projection:
+            self.proj_out = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states,
+        ref_img_feature=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        timestep=None,
+        emotion=None,
+        uc_mask=None,
+        return_dict: bool = True,
+    ):
+        # Input
+        assert hidden_states.dim() == 5, f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+
+        if self.use_audio_module:
+            if encoder_hidden_states.dim() == 4:
+                encoder_hidden_states = rearrange(
+                    encoder_hidden_states,
+                    "bs f margin dim -> (bs f) margin dim",
+                )
+        else:
+            if encoder_hidden_states.shape[0] != hidden_states.shape[0]:
+                encoder_hidden_states = repeat(encoder_hidden_states, "b n c -> (b f) n c", f=video_length)
+
+        batch, _, height, weight = hidden_states.shape
+        residual = hidden_states
+        if self.use_audio_module:
+            residual_audio = encoder_hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = self.proj_in(hidden_states)
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
+            hidden_states = self.proj_in(hidden_states)
+
+        # Blocks
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+                if isinstance(block, TemporalBasicTransformerBlock):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        hidden_states,
+                        ref_img_feature,
+                        None,  # attention_mask
+                        encoder_hidden_states,
+                        timestep,
+                        None,  # cross_attention_kwargs
+                        video_length,
+                        uc_mask,
+                    )
+                elif isinstance(block, JointAudioTemporalBasicTransformerBlock):
+                    (
+                        hidden_states,
+                        encoder_hidden_states,
+                    ) = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        hidden_states,
+                        encoder_hidden_states,
+                        attention_mask,
+                        emotion,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        hidden_states,
+                        encoder_hidden_states,
+                        timestep,
+                        attention_mask,
+                        video_length,
+                    )
+            else:
+                if isinstance(block, TemporalBasicTransformerBlock):
+                    hidden_states = block(
+                        hidden_states=hidden_states,
+                        ref_img_feature=ref_img_feature,
+                        encoder_hidden_states=encoder_hidden_states,
+                        timestep=timestep,
+                        video_length=video_length,
+                        uc_mask=uc_mask,
+                    )
+                elif isinstance(block, JointAudioTemporalBasicTransformerBlock):
+                    hidden_states, encoder_hidden_states = block(
+                        hidden_states,  # shape [2, 4096, 320]
+                        encoder_hidden_states=encoder_hidden_states,  # shape [2, 20, 640]
+                        attention_mask=attention_mask,
+                        emotion=emotion,
+                    )
+                else:
+                    hidden_states = block(
+                        hidden_states,  # shape [2, 4096, 320]
+                        encoder_hidden_states=encoder_hidden_states,  # shape [2, 20, 640]
+                        attention_mask=attention_mask,
+                        timestep=timestep,
+                        video_length=video_length,
+                    )
+
+        # Output
+        if not self.use_linear_projection:
+            hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
+            hidden_states = self.proj_out(hidden_states)
+        else:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+        output = hidden_states + residual
+
+        if self.use_audio_module and not self.is_final_block:
+            audio_output = encoder_hidden_states + residual_audio
+        else:
+            audio_output = None
+
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+        if not return_dict:
+            if self.use_audio_module:
+                return output, audio_output
+            else:
+                return output
+
+        if self.use_audio_module:
+            return output, audio_output
+        else:
+            return output

memo/models/unet_2d_blocks.py

@@ -0,0 +1,935 @@
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import Attention
+from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
+from diffusers.models.transformers.dual_transformer_2d import DualTransformer2DModel
+from diffusers.utils import is_torch_version, logging
+from diffusers.utils.torch_utils import apply_freeu
+from torch import nn
+
+from memo.models.transformer_2d import Transformer2DModel
+
+
+logger = logging.get_logger(__name__)
+
+
+def create_custom_forward(module, return_dict=None):
+    def custom_forward(*inputs):
+        if return_dict is not None:
+            return module(*inputs, return_dict=return_dict)
+
+        return module(*inputs)
+
+    return custom_forward
+
+
+def get_down_block(
+    down_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    temb_channels: int,
+    add_downsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    downsample_padding: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    attention_head_dim: Optional[int] = None,
+    dropout: float = 0.0,
+):
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warning("It is recommended to provide `attention_head_dim` when calling `get_down_block`.")
+        logger.warning(f"Defaulting `attention_head_dim` to {num_attention_heads}.")
+        attention_head_dim = num_attention_heads
+
+    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
+    if down_block_type == "DownBlock2D":
+        return DownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+
+    if down_block_type == "CrossAttnDownBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
+        return CrossAttnDownBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    prev_output_channel: int,
+    temb_channels: int,
+    add_upsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    resolution_idx: Optional[int] = None,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    attention_head_dim: Optional[int] = None,
+    dropout: float = 0.0,
+    is_final_block: bool = False,
+) -> nn.Module:
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warning("It is recommended to provide `attention_head_dim` when calling `get_up_block`.")
+        logger.warning(f"Defaulting `attention_head_dim` to {num_attention_heads}.")
+        attention_head_dim = num_attention_heads
+
+    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    if up_block_type == "UpBlock2D":
+        return UpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    if up_block_type == "CrossAttnUpBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
+        return CrossAttnUpBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+            is_final_block=is_final_block,
+        )
+
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class AutoencoderTinyBlock(nn.Module):
+    def __init__(self, in_channels: int, out_channels: int, act_fn: str):
+        super().__init__()
+        act_fn = get_activation(act_fn)
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+        )
+        self.skip = (
+            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+        self.fuse = nn.ReLU()
+
+    def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        return self.fuse(self.conv(x) + self.skip(x))
+
+
+class UNetMidBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",  # default, spatial
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        attn_groups: Optional[int] = None,
+        resnet_pre_norm: bool = True,
+        add_attention: bool = True,
+        attention_head_dim: int = 1,
+        output_scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        self.add_attention = add_attention
+
+        if attn_groups is None:
+            attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        if attention_head_dim is None:
+            logger.warning(
+                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+            )
+            attention_head_dim = in_channels
+
+        for _ in range(num_layers):
+            if self.add_attention:
+                attentions.append(
+                    Attention(
+                        in_channels,
+                        heads=in_channels // attention_head_dim,
+                        dim_head=attention_head_dim,
+                        rescale_output_factor=output_scale_factor,
+                        eps=resnet_eps,
+                        norm_num_groups=attn_groups,
+                        spatial_norm_dim=(temb_channels if resnet_time_scale_shift == "spatial" else None),
+                        residual_connection=True,
+                        bias=True,
+                        upcast_softmax=True,
+                        _from_deprecated_attn_block=True,
+                    )
+                )
+            else:
+                attentions.append(None)
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                hidden_states = attn(hidden_states, temb=temb)
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        output_scale_factor: float = 1.0,
+        cross_attention_dim: int = 1280,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+        # support for variable transformer layers per block
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for i in range(num_layers):
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+        hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+        ref_feature_list = []
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            hidden_states, ref_feature = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                cross_attention_kwargs=cross_attention_kwargs,
+                attention_mask=attention_mask,
+                encoder_attention_mask=encoder_attention_mask,
+                return_dict=False,
+            )
+            if self.training and self.gradient_checkpointing:
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+            ref_feature_list.append(ref_feature)
+
+        return hidden_states, ref_feature_list
+
+
+class CrossAttnDownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        downsample_padding: int = 1,
+        add_downsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        additional_residuals: Optional[torch.FloatTensor] = None,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        blocks = list(zip(self.resnets, self.attentions))
+        ref_feature_list = []
+        for i, (resnet, attn) in enumerate(blocks):
+            if self.training and self.gradient_checkpointing:
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+            hidden_states, ref_feature = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                cross_attention_kwargs=cross_attention_kwargs,
+                attention_mask=attention_mask,
+                encoder_attention_mask=encoder_attention_mask,
+                return_dict=False,
+            )
+            ref_feature_list.append(ref_feature)
+
+            # apply additional residuals to the output of the last pair of resnet and attention blocks
+            if i == len(blocks) - 1 and additional_residuals is not None:
+                hidden_states = hidden_states + additional_residuals
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states, ref_feature_list
+
+
+class DownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_downsample: bool = True,
+        downsample_padding: int = 1,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        scale: float = 1.0,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        ref_feature_list = []
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+            ref_feature_list.append(hidden_states)
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states, ref_feature_list
+
+
+class CrossAttnUpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+        is_final_block: bool = False,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        self.is_final_block = is_final_block
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                        is_final_block=(is_final_block and i == num_layers - 1),
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        upsample_size: Optional[int] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        ref_feature_list = []
+        for i, (resnet, attn) in enumerate(zip(self.resnets, self.attentions)):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+            hidden_states, ref_feature = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                cross_attention_kwargs=cross_attention_kwargs,
+                attention_mask=attention_mask,
+                encoder_attention_mask=encoder_attention_mask,
+                return_dict=False,
+            )
+            ref_feature_list.append(ref_feature)
+
+        if self.is_final_block:
+            assert hidden_states is None
+        else:
+            if self.upsamplers is not None:
+                for upsampler in self.upsamplers:
+                    hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale)
+
+        return hidden_states, ref_feature_list
+
+
+class UpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        upsample_size: Optional[int] = None,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        ref_feature_list = []
+        for resnet in self.resnets:
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+            ref_feature_list.append(hidden_states)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
+
+        return hidden_states, ref_feature_list

memo/models/unet_2d_condition.py

@@ -0,0 +1,1185 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    GLIGENTextBoundingboxProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    deprecate,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from torch import nn
+
+from memo.models.unet_2d_blocks import (
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+
+logger = logging.get_logger(__name__)
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    ref_features: list[torch.FloatTensor] = None
+
+
+class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to
+        `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+       reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+        *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+        *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        _out_channels: int = 4,
+        _center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = (
+            "UpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        addition_embed_type_num_heads=64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads`"
+                "because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131."
+                "Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in
+        # https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                "Must provide the same number of `down_block_types` as `up_block_types`."
+                f"`down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                "Must provide the same number of `block_out_channels` as `down_block_types`."
+                f"`block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                "Must provide the same number of `only_cross_attention` as `down_block_types`."
+                f"`only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                "Must provide the same number of `num_attention_heads` as `down_block_types`."
+                f"`num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+            raise ValueError(
+                "Must provide the same number of `attention_head_dim` as `down_block_types`."
+                f"`attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                "Must provide the same number of `cross_attention_dim` as `down_block_types`."
+                f"`cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                "Must provide the same number of `layers_per_block` as `down_block_types`."
+                f"`layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=conv_in_kernel,
+            padding=conv_in_padding,
+        )
+
+        # time
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2,
+                set_W_to_weight=False,
+                log=False,
+                flip_sin_to_cos=flip_sin_to_cos,
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim,
+                time_embed_dim,
+                num_heads=addition_embed_type_num_heads,
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim,
+                image_embed_dim=cross_attention_dim,
+                time_embed_dim=time_embed_dim,
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                attention_head_dim=(attention_head_dim[i] if attention_head_dim[i] is not None else output_channel),
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                attention_type=attention_type,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            raise NotImplementedError(f"Unsupport mid_block_type: {mid_block_type}")
+        elif mid_block_type == "UNetMidBlock2D":
+            self.mid_block = UNetMidBlock2D(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                num_layers=0,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                add_attention=False,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                attention_head_dim=(attention_head_dim[i] if attention_head_dim[i] is not None else output_channel),
+                dropout=dropout,
+                is_final_block=is_final_block,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+        self.conv_norm_out = None
+
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, (tuple, list)):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = GLIGENTextBoundingboxProjection(
+                positive_len=positive_len,
+                out_dim=cross_attention_dim,
+                feature_type=feature_type,
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(
+        self,
+        processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]],
+        _remove_lora=False,
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i, size in enumerate(slice_size):
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1, s2, b1, b2):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for _, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for _, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+                    setattr(upsample_block, k, None)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        cond_tensor: torch.FloatTensor = None,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        post_process: bool = False,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor]
+                (https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+                example from ControlNet side model(s)
+            mid_block_additional_residual (`torch.Tensor`, *optional*):
+                additional residual to be added to UNet mid block output, for example from ControlNet side model
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        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=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image'"
+                    "which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time'"
+                    "which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time'"
+                    "which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image'"
+                    "which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint'"
+                    "which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb, hint = self.add_embedding(image_embs, hint)
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj'"
+                    "which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj'"
+                    "which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj'"
+                    "which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds).to(encoder_hidden_states.dtype)
+            encoder_hidden_states = torch.cat([encoder_hidden_states, image_embeds], dim=1)
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+        if cond_tensor is not None:
+            sample = sample + cond_tensor
+
+        # 2.5 GLIGEN position net
+        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+        # 3. down
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        ref_features = {"down": [], "mid": [], "up": []}
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples, ref_feature_list = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples, ref_feature_list = downsample_block(
+                    hidden_states=sample, temb=emb, scale=lora_scale
+                )
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+            ref_features["down"].append(ref_feature_list)
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+                sample, ref_feature_list = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+                ref_features["mid"].append(ref_feature_list)
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample, ref_feature_list = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample, ref_feature_list = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    scale=lora_scale,
+                )
+            ref_features["up"].append(ref_feature_list)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return ref_features
+
+        return UNet2DConditionOutput(ref_features=ref_features)

memo/models/unet_3d.py

@@ -0,0 +1,583 @@
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.attention_processor import AttentionProcessor
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import BaseOutput, logging
+
+from memo.models.resnet import InflatedConv3d, InflatedGroupNorm
+from memo.models.unet_3d_blocks import (
+    UNetMidBlock3DCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+
+logger = logging.get_logger(__name__)
+
+
+@dataclass
+class UNet3DConditionOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+class UNet3DConditionModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 8,
+        out_channels: int = 8,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ),
+        mid_block_type: str = "UNetMidBlock3DCrossAttn",
+        up_block_types: Tuple[str] = (
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: int = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: int = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: int = 1280,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        use_inflated_groupnorm=False,
+        # Additional
+        motion_module_resolutions=(1, 2, 4, 8),
+        motion_module_kwargs=None,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        # audio
+        audio_attention_dim=768,
+        emo_drop_rate=0.3,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        time_embed_dim = block_out_channels[0] * 4
+
+        # input
+        self.conv_in = InflatedConv3d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1))
+
+        # time
+        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+        timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        self.down_blocks = nn.ModuleList([])
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            res = 2**i
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=res in motion_module_resolutions,
+                motion_module_kwargs=motion_module_kwargs,
+                audio_attention_dim=audio_attention_dim,
+                depth=i,
+                emo_drop_rate=emo_drop_rate,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock3DCrossAttn":
+            self.mid_block = UNetMidBlock3DCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                motion_module_kwargs=motion_module_kwargs,
+                audio_attention_dim=audio_attention_dim,
+                depth=3,
+                emo_drop_rate=emo_drop_rate,
+            )
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the videos
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_attention_head_dim = list(reversed(attention_head_dim))
+        only_cross_attention = list(reversed(only_cross_attention))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            res = 2 ** (3 - i)
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=layers_per_block + 1,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=reversed_attention_head_dim[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=res in motion_module_resolutions,
+                motion_module_kwargs=motion_module_kwargs,
+                audio_attention_dim=audio_attention_dim,
+                depth=3 - i,
+                emo_drop_rate=emo_drop_rate,
+                is_final_block=is_final_block,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if use_inflated_groupnorm:
+            self.conv_norm_out = InflatedGroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+        else:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+        self.conv_act = nn.SiLU()
+        self.conv_out = InflatedConv3d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
+
+    @property
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "set_processor"):
+                processors[f"{name}.processor"] = module.processor
+
+            for sub_name, child in module.named_children():
+                if "temporal_transformer" not in sub_name:
+                    fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            if "temporal_transformer" not in name:
+                fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_slicable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_slicable_dims(module)
+
+        num_slicable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_slicable_layers * [1]
+
+        slice_size = num_slicable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i, size in enumerate(slice_size):
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                if "temporal_transformer" not in sub_name:
+                    fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            if "temporal_transformer" not in name:
+                fn_recursive_attn_processor(name, module, processor)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        ref_features: dict,
+        timestep: Union[torch.Tensor, float, int, list],
+        encoder_hidden_states: torch.Tensor,
+        audio_embedding: Optional[torch.Tensor] = None,
+        audio_emotion: Optional[torch.Tensor] = None,
+        class_labels: Optional[torch.Tensor] = None,
+        mask_cond_fea: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        uc_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        is_new_audio=True,
+        update_past_memory=False,
+    ) -> Union[UNet3DConditionOutput, Tuple]:
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # time
+        timesteps = timestep
+        if isinstance(timesteps, list):
+            t_emb_list = []
+            for timesteps in timestep:
+                if not torch.is_tensor(timesteps):
+                    # This would be a good case for the `match` statement (Python 3.10+)
+                    is_mps = sample.device.type == "mps"
+                    if isinstance(timestep, float):
+                        dtype = torch.float32 if is_mps else torch.float64
+                    else:
+                        dtype = torch.int32 if is_mps else torch.int64
+                    timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+                elif len(timesteps.shape) == 0:
+                    timesteps = timesteps[None].to(sample.device)
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timesteps = timesteps.expand(sample.shape[0])
+                t_emb = self.time_proj(timesteps)
+                t_emb_list.append(t_emb)
+
+            t_emb = torch.stack(t_emb_list, dim=1)
+        else:
+            if not torch.is_tensor(timesteps):
+                # This would be a good case for the `match` statement (Python 3.10+)
+                is_mps = sample.device.type == "mps"
+                if isinstance(timestep, float):
+                    dtype = torch.float32 if is_mps else torch.float64
+                else:
+                    dtype = torch.int32 if is_mps else torch.int64
+                timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+            elif len(timesteps.shape) == 0:
+                timesteps = timesteps[None].to(sample.device)
+            # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+            timesteps = timesteps.expand(sample.shape[0])
+            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=self.dtype)
+        emb = self.time_embedding(t_emb)
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+
+        # pre-process
+        sample = self.conv_in(sample)
+        if mask_cond_fea is not None:
+            sample = sample + mask_cond_fea
+
+        # down
+        down_block_res_samples = (sample,)
+        for i, downsample_block in enumerate(self.down_blocks):
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                sample, res_samples, audio_embedding = downsample_block(
+                    hidden_states=sample,
+                    ref_feature_list=ref_features["down"][i],
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    audio_embedding=audio_embedding,
+                    emotion=audio_emotion,
+                    uc_mask=uc_mask,
+                    is_new_audio=is_new_audio,
+                    update_past_memory=update_past_memory,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    ref_feature_list=ref_features["down"][i],
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    is_new_audio=is_new_audio,
+                    update_past_memory=update_past_memory,
+                )
+
+            down_block_res_samples += res_samples
+
+        if down_block_additional_residuals is not None:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples += (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # mid
+        sample, audio_embedding = self.mid_block(
+            sample,
+            ref_feature_list=ref_features["mid"][0],
+            temb=emb,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            audio_embedding=audio_embedding,
+            emotion=audio_emotion,
+            uc_mask=uc_mask,
+            is_new_audio=is_new_audio,
+            update_past_memory=update_past_memory,
+        )
+
+        if mid_block_additional_residual is not None:
+            sample = sample + mid_block_additional_residual
+
+        # up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample, audio_embedding = upsample_block(
+                    hidden_states=sample,
+                    ref_feature_list=ref_features["up"][i],
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    audio_embedding=audio_embedding,
+                    emotion=audio_emotion,
+                    uc_mask=uc_mask,
+                    is_new_audio=is_new_audio,
+                    update_past_memory=update_past_memory,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    ref_feature_list=ref_features["up"][i],
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    encoder_hidden_states=encoder_hidden_states,
+                    is_new_audio=is_new_audio,
+                    update_past_memory=update_past_memory,
+                )
+
+        # post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet3DConditionOutput(sample=sample)

memo/models/unet_3d_blocks.py

@@ -0,0 +1,1024 @@
+from typing import Any, Dict
+
+import torch
+from diffusers.utils import is_torch_version
+from einops import rearrange
+from torch import nn
+
+from memo.models.motion_module import MemoryLinearAttnTemporalModule
+from memo.models.resnet import Downsample3D, ResnetBlock3D, Upsample3D
+from memo.models.transformer_3d import Transformer3DModel
+
+
+def create_custom_forward(module, return_dict=None):
+    def custom_forward(*inputs):
+        if return_dict is not None:
+            return module(*inputs, return_dict=return_dict)
+
+        return module(*inputs)
+
+    return custom_forward
+
+
+def get_down_block(
+    down_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    temb_channels,
+    add_downsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    audio_attention_dim=None,
+    downsample_padding=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    unet_use_cross_frame_attention=None,
+    unet_use_temporal_attention=None,
+    use_inflated_groupnorm=None,
+    use_motion_module=None,
+    motion_module_kwargs=None,
+    depth=0,
+    emo_drop_rate=0.3,
+):
+    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
+    if down_block_type == "DownBlock3D":
+        return DownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+
+    if down_block_type == "CrossAttnDownBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock3D")
+        return CrossAttnDownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            audio_attention_dim=audio_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+            unet_use_temporal_attention=unet_use_temporal_attention,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_kwargs=motion_module_kwargs,
+            depth=depth,
+            emo_drop_rate=emo_drop_rate,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    prev_output_channel,
+    temb_channels,
+    add_upsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    audio_attention_dim=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    unet_use_cross_frame_attention=None,
+    unet_use_temporal_attention=None,
+    use_inflated_groupnorm=None,
+    use_motion_module=None,
+    motion_module_kwargs=None,
+    depth=0,
+    emo_drop_rate=0.3,
+    is_final_block=False,
+):
+    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    if up_block_type == "UpBlock3D":
+        return UpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+
+    if up_block_type == "CrossAttnUpBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock3D")
+        return CrossAttnUpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            audio_attention_dim=audio_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+            unet_use_temporal_attention=unet_use_temporal_attention,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_kwargs=motion_module_kwargs,
+            depth=depth,
+            emo_drop_rate=emo_drop_rate,
+            is_final_block=is_final_block,
+        )
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class UNetMidBlock3DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        output_scale_factor=1.0,
+        cross_attention_dim=1280,
+        audio_attention_dim=1024,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_inflated_groupnorm=None,
+        motion_module_kwargs=None,
+        depth=0,
+        emo_drop_rate=0.3,
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock3D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+            )
+        ]
+        attentions = []
+        motion_modules = []
+        audio_modules = []
+
+        for _ in range(num_layers):
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    in_channels // attn_num_head_channels,
+                    in_channels=in_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+            )
+            audio_modules.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    in_channels // attn_num_head_channels,
+                    in_channels=in_channels,
+                    num_layers=1,
+                    cross_attention_dim=audio_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=False,
+                    upcast_attention=upcast_attention,
+                    use_audio_module=True,
+                    depth=depth,
+                    unet_block_name="mid",
+                    emo_drop_rate=emo_drop_rate,
+                )
+            )
+
+            motion_modules.append(
+                MemoryLinearAttnTemporalModule(
+                    in_channels=in_channels,
+                    **motion_module_kwargs,
+                )
+            )
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.audio_modules = nn.ModuleList(audio_modules)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        ref_feature_list,
+        temb=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        audio_embedding=None,
+        emotion=None,
+        uc_mask=None,
+        is_new_audio=True,
+        update_past_memory=False,
+    ):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for i, (attn, resnet, audio_module, motion_module) in enumerate(
+            zip(
+                self.attentions,
+                self.resnets[1:],
+                self.audio_modules,
+                self.motion_modules,
+            )
+        ):
+            ref_feature = ref_feature_list[i]
+            ref_feature = ref_feature[0]
+            ref_feature = rearrange(
+                ref_feature,
+                "(b f) (h w) c -> b c f h w",
+                b=hidden_states.shape[0],
+                w=hidden_states.shape[-1],
+            )
+            ref_img_feature = ref_feature[:, :, :1, :, :]
+            ref_img_feature = rearrange(
+                ref_img_feature,
+                "b c f h w -> (b f) (h w) c",
+            )
+            motion_frames = ref_feature[:, :, 1:, :, :]
+
+            hidden_states = attn(
+                hidden_states,
+                ref_img_feature,
+                encoder_hidden_states=encoder_hidden_states,
+                uc_mask=uc_mask,
+                return_dict=False,
+            )
+            if audio_module is not None:
+                hidden_states, audio_embedding = audio_module(
+                    hidden_states,
+                    ref_img_feature=None,
+                    encoder_hidden_states=audio_embedding,
+                    attention_mask=attention_mask,
+                    return_dict=False,
+                    emotion=emotion,
+                )
+            if motion_module is not None:
+                motion_frames = motion_frames.to(device=hidden_states.device, dtype=hidden_states.dtype)
+                hidden_states = motion_module(
+                    hidden_states=hidden_states,
+                    motion_frames=motion_frames,
+                    encoder_hidden_states=encoder_hidden_states,
+                    is_new_audio=is_new_audio,
+                    update_past_memory=update_past_memory,
+                )
+
+            if self.training and self.gradient_checkpointing:
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+        if audio_module is not None:
+            return hidden_states, audio_embedding
+        else:
+            return hidden_states
+
+
+class CrossAttnDownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        audio_attention_dim=1024,
+        output_scale_factor=1.0,
+        downsample_padding=1,
+        add_downsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_kwargs=None,
+        depth=0,
+        emo_drop_rate=0.3,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+        audio_modules = []
+        motion_modules = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+            )
+            audio_modules.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    in_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=audio_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    use_audio_module=True,
+                    depth=depth,
+                    unet_block_name="down",
+                    emo_drop_rate=emo_drop_rate,
+                )
+            )
+            motion_modules.append(
+                MemoryLinearAttnTemporalModule(
+                    in_channels=out_channels,
+                    **motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.audio_modules = nn.ModuleList(audio_modules)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        ref_feature_list,
+        temb=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        audio_embedding=None,
+        emotion=None,
+        uc_mask=None,
+        is_new_audio=True,
+        update_past_memory=False,
+    ):
+        output_states = ()
+
+        for i, (resnet, attn, audio_module, motion_module) in enumerate(
+            zip(self.resnets, self.attentions, self.audio_modules, self.motion_modules)
+        ):
+            ref_feature = ref_feature_list[i]
+            ref_feature = ref_feature[0]
+            ref_feature = rearrange(
+                ref_feature,
+                "(b f) (h w) c -> b c f h w",
+                b=hidden_states.shape[0],
+                w=hidden_states.shape[-1],
+            )
+            ref_img_feature = ref_feature[:, :, :1, :, :]
+            ref_img_feature = rearrange(
+                ref_img_feature,
+                "b c f h w -> (b f) (h w) c",
+            )
+            motion_frames = ref_feature[:, :, 1:, :, :]
+
+            if self.training and self.gradient_checkpointing:
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+            hidden_states = attn(
+                hidden_states,
+                ref_img_feature,
+                encoder_hidden_states=encoder_hidden_states,
+                uc_mask=uc_mask,
+                return_dict=False,
+            )
+
+            if audio_module is not None:
+                hidden_states, audio_embedding = audio_module(
+                    hidden_states,
+                    ref_img_feature=None,
+                    encoder_hidden_states=audio_embedding,
+                    attention_mask=attention_mask,
+                    return_dict=False,
+                    emotion=emotion,
+                )
+
+            # add motion module
+            if motion_module is not None:
+                motion_frames = motion_frames.to(device=hidden_states.device, dtype=hidden_states.dtype)
+                hidden_states = motion_module(
+                    hidden_states=hidden_states,
+                    motion_frames=motion_frames,
+                    encoder_hidden_states=encoder_hidden_states,
+                    is_new_audio=is_new_audio,
+                    update_past_memory=update_past_memory,
+                )
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        if audio_module is not None:
+            return hidden_states, output_states, audio_embedding
+        else:
+            return hidden_states, output_states
+
+
+class DownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_downsample=True,
+        downsample_padding=1,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        motion_modules = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            motion_modules.append(
+                MemoryLinearAttnTemporalModule(
+                    in_channels=out_channels,
+                    **motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        ref_feature_list,
+        temb=None,
+        encoder_hidden_states=None,
+        is_new_audio=True,
+        update_past_memory=False,
+    ):
+        output_states = ()
+
+        for i, (resnet, motion_module) in enumerate(zip(self.resnets, self.motion_modules)):
+            ref_feature = ref_feature_list[i]
+            ref_feature = rearrange(
+                ref_feature,
+                "(b f) c h w -> b c f h w",
+                b=hidden_states.shape[0],
+                w=hidden_states.shape[-1],
+            )
+            motion_frames = ref_feature[:, :, 1:, :, :]
+
+            if self.training and self.gradient_checkpointing:
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+            if motion_module is not None:
+                hidden_states = motion_module(
+                    hidden_states=hidden_states,
+                    motion_frames=motion_frames,
+                    encoder_hidden_states=encoder_hidden_states,
+                    is_new_audio=is_new_audio,
+                    update_past_memory=update_past_memory,
+                )
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        audio_attention_dim=1024,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_motion_module=None,
+        use_inflated_groupnorm=None,
+        motion_module_kwargs=None,
+        depth=0,
+        emo_drop_rate=0.3,
+        is_final_block=False,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+        audio_modules = []
+        motion_modules = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        self.is_final_block = is_final_block
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+            )
+            audio_modules.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    in_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=audio_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    use_audio_module=True,
+                    depth=depth,
+                    unet_block_name="up",
+                    emo_drop_rate=emo_drop_rate,
+                    is_final_block=(is_final_block and i == num_layers - 1),
+                )
+            )
+            motion_modules.append(
+                MemoryLinearAttnTemporalModule(
+                    in_channels=out_channels,
+                    **motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.audio_modules = nn.ModuleList(audio_modules)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample3D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        ref_feature_list,
+        res_hidden_states_tuple,
+        temb=None,
+        encoder_hidden_states=None,
+        upsample_size=None,
+        attention_mask=None,
+        audio_embedding=None,
+        emotion=None,
+        uc_mask=None,
+        is_new_audio=True,
+        update_past_memory=False,
+    ):
+        for i, (resnet, attn, audio_module, motion_module) in enumerate(
+            zip(self.resnets, self.attentions, self.audio_modules, self.motion_modules)
+        ):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            ref_feature = ref_feature_list[i]
+            ref_feature = ref_feature[0]
+            ref_feature = rearrange(
+                ref_feature,
+                "(b f) (h w) c -> b c f h w",
+                b=hidden_states.shape[0],
+                w=hidden_states.shape[-1],
+            )
+            ref_img_feature = ref_feature[:, :, :1, :, :]
+            ref_img_feature = rearrange(
+                ref_img_feature,
+                "b c f h w -> (b f) (h w) c",
+            )
+            motion_frames = ref_feature[:, :, 1:, :, :]
+
+            if self.training and self.gradient_checkpointing:
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+            hidden_states = attn(
+                hidden_states,
+                ref_img_feature,
+                encoder_hidden_states=encoder_hidden_states,
+                uc_mask=uc_mask,
+                return_dict=False,
+            )
+
+            if audio_module is not None:
+                hidden_states, audio_embedding = audio_module(
+                    hidden_states,
+                    ref_img_feature=None,
+                    encoder_hidden_states=audio_embedding,
+                    attention_mask=attention_mask,
+                    return_dict=False,
+                    emotion=emotion,
+                )
+
+            # add motion module
+            if motion_module is not None:
+                motion_frames = motion_frames.to(device=hidden_states.device, dtype=hidden_states.dtype)
+                hidden_states = motion_module(
+                    hidden_states,
+                    motion_frames,
+                    encoder_hidden_states,
+                    is_new_audio=is_new_audio,
+                    update_past_memory=update_past_memory,
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        if audio_module is not None:
+            return hidden_states, audio_embedding
+        else:
+            return hidden_states
+
+
+class UpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        motion_modules = []
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            motion_modules.append(
+                MemoryLinearAttnTemporalModule(
+                    in_channels=out_channels,
+                    **motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample3D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        ref_feature_list,
+        res_hidden_states_tuple,
+        temb=None,
+        upsample_size=None,
+        encoder_hidden_states=None,
+        is_new_audio=True,
+        update_past_memory=False,
+    ):
+        for i, (resnet, motion_module) in enumerate(zip(self.resnets, self.motion_modules)):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            ref_feature = ref_feature_list[i]
+            ref_feature = rearrange(
+                ref_feature,
+                "(b f) c h w -> b c f h w",
+                b=hidden_states.shape[0],
+                w=hidden_states.shape[-1],
+            )
+            motion_frames = ref_feature[:, :, 1:, :, :]
+
+            if self.training and self.gradient_checkpointing:
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+            if motion_module is not None:
+                hidden_states = motion_module(
+                    hidden_states=hidden_states,
+                    motion_frames=motion_frames,
+                    encoder_hidden_states=encoder_hidden_states,
+                    is_new_audio=is_new_audio,
+                    update_past_memory=update_past_memory,
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states

memo/models/wav2vec.py

@@ -0,0 +1,127 @@
+import torch.nn.functional as F
+from transformers import Wav2Vec2Model
+from transformers.modeling_outputs import BaseModelOutput
+
+
+class Wav2VecModel(Wav2Vec2Model):
+    def forward(
+        self,
+        input_values,
+        seq_len,
+        attention_mask=None,
+        mask_time_indices=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        self.config.output_attentions = True
+
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        extract_features = self.feature_extractor(input_values)
+        extract_features = extract_features.transpose(1, 2)
+        extract_features = linear_interpolation(extract_features, seq_len=seq_len)
+
+        if attention_mask is not None:
+            # compute reduced attention_mask corresponding to feature vectors
+            attention_mask = self._get_feature_vector_attention_mask(
+                extract_features.shape[1], attention_mask, add_adapter=False
+            )
+
+        hidden_states, extract_features = self.feature_projection(extract_features)
+        hidden_states = self._mask_hidden_states(
+            hidden_states,
+            mask_time_indices=mask_time_indices,
+            attention_mask=attention_mask,
+        )
+
+        encoder_outputs = self.encoder(
+            hidden_states,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = encoder_outputs[0]
+
+        if self.adapter is not None:
+            hidden_states = self.adapter(hidden_states)
+
+        if not return_dict:
+            return (hidden_states,) + encoder_outputs[1:]
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def feature_extract(
+        self,
+        input_values,
+        seq_len,
+    ):
+        extract_features = self.feature_extractor(input_values)
+        extract_features = extract_features.transpose(1, 2)
+        extract_features = linear_interpolation(extract_features, seq_len=seq_len)
+
+        return extract_features
+
+    def encode(
+        self,
+        extract_features,
+        attention_mask=None,
+        mask_time_indices=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        self.config.output_attentions = True
+
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if attention_mask is not None:
+            # compute reduced attention_mask corresponding to feature vectors
+            attention_mask = self._get_feature_vector_attention_mask(
+                extract_features.shape[1], attention_mask, add_adapter=False
+            )
+
+        hidden_states, extract_features = self.feature_projection(extract_features)
+        hidden_states = self._mask_hidden_states(
+            hidden_states,
+            mask_time_indices=mask_time_indices,
+            attention_mask=attention_mask,
+        )
+
+        encoder_outputs = self.encoder(
+            hidden_states,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = encoder_outputs[0]
+
+        if self.adapter is not None:
+            hidden_states = self.adapter(hidden_states)
+
+        if not return_dict:
+            return (hidden_states,) + encoder_outputs[1:]
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+def linear_interpolation(features, seq_len):
+    features = features.transpose(1, 2)
+    output_features = F.interpolate(features, size=seq_len, align_corners=True, mode="linear")
+    return output_features.transpose(1, 2)

memo/pipelines/video_pipeline.py

@@ -0,0 +1,295 @@
+import inspect
+from dataclasses import dataclass
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+from diffusers import (
+    DDIMScheduler,
+    DiffusionPipeline,
+    DPMSolverMultistepScheduler,
+    EulerAncestralDiscreteScheduler,
+    EulerDiscreteScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+)
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.utils import BaseOutput
+from diffusers.utils.torch_utils import randn_tensor
+from einops import rearrange
+
+
+@dataclass
+class VideoPipelineOutput(BaseOutput):
+    videos: Union[torch.Tensor, np.ndarray]
+
+
+class VideoPipeline(DiffusionPipeline):
+    def __init__(
+        self,
+        vae,
+        reference_net,
+        diffusion_net,
+        image_proj,
+        scheduler: Union[
+            DDIMScheduler,
+            PNDMScheduler,
+            LMSDiscreteScheduler,
+            EulerDiscreteScheduler,
+            EulerAncestralDiscreteScheduler,
+            DPMSolverMultistepScheduler,
+        ],
+    ) -> None:
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            reference_net=reference_net,
+            diffusion_net=diffusion_net,
+            scheduler=scheduler,
+            image_proj=image_proj,
+        )
+
+        self.vae_scale_factor: int = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+        self.ref_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor,
+            do_convert_rgb=True,
+        )
+
+    @property
+    def _execution_device(self):
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def prepare_latents(
+        self,
+        batch_size: int,  # Number of videos to generate in parallel
+        num_channels_latents: int,  # Number of channels in the latents
+        width: int,  # Width of the video frame
+        height: int,  # Height of the video frame
+        video_length: int,  # Length of the video in frames
+        dtype: torch.dtype,  # Data type of the latents
+        device: torch.device,  # Device to store the latents on
+        generator: Optional[torch.Generator] = None,  # Random number generator for reproducibility
+        latents: Optional[torch.Tensor] = None,  # Pre-generated latents (optional)
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            video_length,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        if hasattr(self.scheduler, "init_noise_sigma"):
+            latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def decode_latents(self, latents):
+        video_length = latents.shape[2]
+        latents = 1 / 0.18215 * latents
+        latents = rearrange(latents, "b c f h w -> (b f) c h w")
+        video = []
+        for frame_idx in range(latents.shape[0]):
+            video.append(self.vae.decode(latents[frame_idx : frame_idx + 1]).sample)
+        video = torch.cat(video)
+        video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
+        video = (video / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        video = video.cpu().float().numpy()
+        return video
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        ref_image,
+        face_emb,
+        audio_tensor,
+        width,
+        height,
+        video_length,
+        num_inference_steps,
+        guidance_scale,
+        num_images_per_prompt=1,
+        eta: float = 0.0,
+        audio_emotion=None,
+        emotion_class_num=None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: Optional[str] = "tensor",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+    ):
+        # Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        device = self._execution_device
+
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        batch_size = 1
+
+        # prepare clip image embeddings
+        clip_image_embeds = face_emb
+        clip_image_embeds = clip_image_embeds.to(self.image_proj.device, self.image_proj.dtype)
+
+        encoder_hidden_states = self.image_proj(clip_image_embeds)
+        uncond_encoder_hidden_states = self.image_proj(torch.zeros_like(clip_image_embeds))
+
+        if do_classifier_free_guidance:
+            encoder_hidden_states = torch.cat([uncond_encoder_hidden_states, encoder_hidden_states], dim=0)
+
+        num_channels_latents = self.diffusion_net.in_channels
+
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            width,
+            height,
+            video_length,
+            clip_image_embeds.dtype,
+            device,
+            generator,
+        )
+
+        # Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # Prepare ref image latents
+        ref_image_tensor = rearrange(ref_image, "b f c h w -> (b f) c h w")
+        ref_image_tensor = self.ref_image_processor.preprocess(
+            ref_image_tensor, height=height, width=width
+        )  # (bs, c, width, height)
+        ref_image_tensor = ref_image_tensor.to(dtype=self.vae.dtype, device=self.vae.device)
+        # To save memory on GPUs like RTX 4090, we encode each frame separately
+        # ref_image_latents = self.vae.encode(ref_image_tensor).latent_dist.mean
+        ref_image_latents = []
+        for frame_idx in range(ref_image_tensor.shape[0]):
+            ref_image_latents.append(self.vae.encode(ref_image_tensor[frame_idx : frame_idx + 1]).latent_dist.mean)
+        ref_image_latents = torch.cat(ref_image_latents, dim=0)
+
+        ref_image_latents = ref_image_latents * 0.18215  # (b, 4, h, w)
+
+        if do_classifier_free_guidance:
+            uncond_audio_tensor = torch.zeros_like(audio_tensor)
+            audio_tensor = torch.cat([uncond_audio_tensor, audio_tensor], dim=0)
+            audio_tensor = audio_tensor.to(dtype=self.diffusion_net.dtype, device=self.diffusion_net.device)
+
+        # denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i in range(len(timesteps)):
+                t = timesteps[i]
+                # Forward reference image
+                if i == 0:
+                    ref_features = self.reference_net(
+                        ref_image_latents.repeat((2 if do_classifier_free_guidance else 1), 1, 1, 1),
+                        torch.zeros_like(t),
+                        encoder_hidden_states=encoder_hidden_states,
+                        return_dict=False,
+                    )
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                if hasattr(self.scheduler, "scale_model_input"):
+                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                audio_emotion = torch.tensor(torch.mode(audio_emotion).values.item()).to(
+                    dtype=torch.int, device=self.diffusion_net.device
+                )
+                if do_classifier_free_guidance:
+                    uncond_audio_emotion = torch.full_like(audio_emotion, emotion_class_num)
+                    audio_emotion = torch.cat(
+                        [uncond_audio_emotion.unsqueeze(0), audio_emotion.unsqueeze(0)],
+                        dim=0,
+                    )
+
+                    uc_mask = (
+                        torch.Tensor(
+                            [1] * batch_size * num_images_per_prompt * 16
+                            + [0] * batch_size * num_images_per_prompt * 16
+                        )
+                        .to(device)
+                        .bool()
+                    )
+                else:
+                    uc_mask = None
+
+                noise_pred = self.diffusion_net(
+                    latent_model_input,
+                    ref_features,
+                    t,
+                    encoder_hidden_states=encoder_hidden_states,
+                    audio_embedding=audio_tensor,
+                    audio_emotion=audio_emotion,
+                    uc_mask=uc_mask,
+                ).sample
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        # Post-processing
+        images = self.decode_latents(latents)  # (b, c, f, h, w)
+
+        # Convert to tensor
+        if output_type == "tensor":
+            images = torch.from_numpy(images)
+
+        if not return_dict:
+            return images
+
+        return VideoPipelineOutput(videos=images)

memo/utils/audio_utils.py

@@ -0,0 +1,293 @@
+import logging
+import math
+import os
+import subprocess
+from io import BytesIO
+
+import librosa
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torchaudio
+from audio_separator.separator import Separator
+from einops import rearrange
+from funasr.download.download_from_hub import download_model
+from funasr.models.emotion2vec.model import Emotion2vec
+from transformers import Wav2Vec2FeatureExtractor
+
+from memo.models.emotion_classifier import AudioEmotionClassifierModel
+from memo.models.wav2vec import Wav2VecModel
+
+
+logger = logging.getLogger(__name__)
+
+
+def resample_audio(input_audio_file: str, output_audio_file: str, sample_rate: int = 16000):
+    p = subprocess.Popen(
+        [
+            "ffmpeg",
+            "-y",
+            "-v",
+            "error",
+            "-i",
+            input_audio_file,
+            "-ar",
+            str(sample_rate),
+            output_audio_file,
+        ]
+    )
+    ret = p.wait()
+    assert ret == 0, f"Resample audio failed! Input: {input_audio_file}, Output: {output_audio_file}"
+    return output_audio_file
+
+
+@torch.no_grad()
+def preprocess_audio(
+    wav_path: str,
+    fps: int,
+    wav2vec_model: str,
+    vocal_separator_model: str = None,
+    cache_dir: str = "",
+    device: str = "cuda",
+    sample_rate: int = 16000,
+    num_generated_frames_per_clip: int = -1,
+):
+    """
+    Preprocess the audio file and extract audio embeddings.
+
+    Args:
+        wav_path (str): Path to the input audio file.
+        fps (int): Frames per second for the audio processing.
+        wav2vec_model (str): Path to the pretrained Wav2Vec model.
+        vocal_separator_model (str, optional): Path to the vocal separator model. Defaults to None.
+        cache_dir (str, optional): Directory for cached files. Defaults to "".
+        device (str, optional): Device to use ('cuda' or 'cpu'). Defaults to "cuda".
+        sample_rate (int, optional): Sampling rate for audio processing. Defaults to 16000.
+        num_generated_frames_per_clip (int, optional): Number of generated frames per clip for padding. Defaults to -1.
+
+    Returns:
+        tuple: A tuple containing:
+            - audio_emb (torch.Tensor): The processed audio embeddings.
+            - audio_length (int): The length of the audio in frames.
+    """
+    # Initialize Wav2Vec model
+    audio_encoder = Wav2VecModel.from_pretrained(wav2vec_model).to(device=device)
+    audio_encoder.feature_extractor._freeze_parameters()
+
+    # Initialize Wav2Vec feature extractor
+    wav2vec_feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(wav2vec_model)
+
+    # Initialize vocal separator if provided
+    vocal_separator = None
+    if vocal_separator_model is not None:
+        os.makedirs(cache_dir, exist_ok=True)
+        vocal_separator = Separator(
+            output_dir=cache_dir,
+            output_single_stem="vocals",
+            model_file_dir=os.path.dirname(vocal_separator_model),
+        )
+        vocal_separator.load_model(os.path.basename(vocal_separator_model))
+        assert vocal_separator.model_instance is not None, "Failed to load audio separation model."
+
+    # Perform vocal separation if applicable
+    if vocal_separator is not None:
+        outputs = vocal_separator.separate(wav_path)
+        assert len(outputs) > 0, "Audio separation failed."
+        vocal_audio_file = outputs[0]
+        vocal_audio_name, _ = os.path.splitext(vocal_audio_file)
+        vocal_audio_file = os.path.join(vocal_separator.output_dir, vocal_audio_file)
+        vocal_audio_file = resample_audio(
+            vocal_audio_file,
+            os.path.join(vocal_separator.output_dir, f"{vocal_audio_name}-16k.wav"),
+            sample_rate,
+        )
+    else:
+        vocal_audio_file = wav_path
+
+    # Load audio and extract Wav2Vec features
+    speech_array, sampling_rate = librosa.load(vocal_audio_file, sr=sample_rate)
+    audio_feature = np.squeeze(wav2vec_feature_extractor(speech_array, sampling_rate=sampling_rate).input_values)
+    audio_length = math.ceil(len(audio_feature) / sample_rate * fps)
+    audio_feature = torch.from_numpy(audio_feature).float().to(device=device)
+
+    # Pad audio features to match the required length
+    if num_generated_frames_per_clip > 0 and audio_length % num_generated_frames_per_clip != 0:
+        audio_feature = torch.nn.functional.pad(
+            audio_feature,
+            (
+                0,
+                (num_generated_frames_per_clip - audio_length % num_generated_frames_per_clip) * (sample_rate // fps),
+            ),
+            "constant",
+            0.0,
+        )
+        audio_length += num_generated_frames_per_clip - audio_length % num_generated_frames_per_clip
+    audio_feature = audio_feature.unsqueeze(0)
+
+    # Extract audio embeddings
+    with torch.no_grad():
+        embeddings = audio_encoder(audio_feature, seq_len=audio_length, output_hidden_states=True)
+    assert len(embeddings) > 0, "Failed to extract audio embeddings."
+    audio_emb = torch.stack(embeddings.hidden_states[1:], dim=1).squeeze(0)
+    audio_emb = rearrange(audio_emb, "b s d -> s b d")
+
+    # Concatenate embeddings with surrounding frames
+    audio_emb = audio_emb.cpu().detach()
+    concatenated_tensors = []
+    for i in range(audio_emb.shape[0]):
+        vectors_to_concat = [audio_emb[max(min(i + j, audio_emb.shape[0] - 1), 0)] for j in range(-2, 3)]
+        concatenated_tensors.append(torch.stack(vectors_to_concat, dim=0))
+    audio_emb = torch.stack(concatenated_tensors, dim=0)
+
+    if vocal_separator is not None:
+        del vocal_separator
+    del audio_encoder
+
+    return audio_emb, audio_length
+
+
+@torch.no_grad()
+def extract_audio_emotion_labels(
+    model: str,
+    wav_path: str,
+    emotion2vec_model: str,
+    audio_length: int,
+    sample_rate: int = 16000,
+    device: str = "cuda",
+):
+    """
+    Extract audio emotion labels from an audio file.
+
+    Args:
+        model (str): Path to the MEMO model.
+        wav_path (str): Path to the input audio file.
+        emotion2vec_model (str): Path to the Emotion2vec model.
+        audio_length (int): Target length for interpolated emotion labels.
+        sample_rate (int, optional): Sample rate of the input audio. Default is 16000.
+        device (str, optional): Device to use ('cuda' or 'cpu'). Default is "cuda".
+
+    Returns:
+        torch.Tensor: Processed emotion labels with shape matching the target audio length.
+    """
+    # Load models
+    logger.info("Downloading emotion2vec models from modelscope")
+    kwargs = download_model(model=emotion2vec_model)
+    kwargs["tokenizer"] = None
+    kwargs["input_size"] = None
+    kwargs["frontend"] = None
+    emotion_model = Emotion2vec(**kwargs, vocab_size=-1).to(device)
+    init_param = kwargs.get("init_param", None)
+    load_emotion2vec_model(
+        model=emotion_model,
+        path=init_param,
+        ignore_init_mismatch=kwargs.get("ignore_init_mismatch", True),
+        oss_bucket=kwargs.get("oss_bucket", None),
+        scope_map=kwargs.get("scope_map", []),
+    )
+    emotion_model.eval()
+
+    classifier = AudioEmotionClassifierModel.from_pretrained(
+        model,
+        subfolder="misc/audio_emotion_classifier",
+        use_safetensors=True,
+    ).to(device=device)
+    classifier.eval()
+
+    # Load audio
+    wav, sr = torchaudio.load(wav_path)
+    if sr != sample_rate:
+        wav = torchaudio.functional.resample(wav, sr, sample_rate)
+    wav = wav.view(-1) if wav.dim() == 1 else wav[0].view(-1)
+
+    emotion_labels = torch.full_like(wav, -1, dtype=torch.int32)
+
+    def extract_emotion(x):
+        """
+        Extract emotion for a given audio segment.
+        """
+        x = x.to(device=device)
+        x = F.layer_norm(x, x.shape).view(1, -1)
+        feats = emotion_model.extract_features(x)
+        x = feats["x"].mean(dim=1)  # average across frames
+        x = classifier(x)
+        x = torch.softmax(x, dim=-1)
+        return torch.argmax(x, dim=-1)
+
+    # Process start, middle, and end segments
+    start_label = extract_emotion(wav[: sample_rate * 2]).item()
+    emotion_labels[:sample_rate] = start_label
+
+    for i in range(sample_rate, len(wav) - sample_rate, sample_rate):
+        mid_wav = wav[i - sample_rate : i - sample_rate + sample_rate * 3]
+        mid_label = extract_emotion(mid_wav).item()
+        emotion_labels[i : i + sample_rate] = mid_label
+
+    end_label = extract_emotion(wav[-sample_rate * 2 :]).item()
+    emotion_labels[-sample_rate:] = end_label
+
+    # Interpolate to match the target audio length
+    emotion_labels = emotion_labels.unsqueeze(0).unsqueeze(0).float()
+    emotion_labels = F.interpolate(emotion_labels, size=audio_length, mode="nearest").squeeze(0).squeeze(0).int()
+    num_emotion_classes = classifier.num_emotion_classes
+
+    del emotion_model
+    del classifier
+
+    return emotion_labels, num_emotion_classes
+
+
+def load_emotion2vec_model(
+    path: str,
+    model: torch.nn.Module,
+    ignore_init_mismatch: bool = True,
+    map_location: str = "cpu",
+    oss_bucket=None,
+    scope_map=[],
+):
+    obj = model
+    dst_state = obj.state_dict()
+    logger.debug(f"Emotion2vec checkpoint: {path}")
+    if oss_bucket is None:
+        src_state = torch.load(path, map_location=map_location)
+    else:
+        buffer = BytesIO(oss_bucket.get_object(path).read())
+        src_state = torch.load(buffer, map_location=map_location)
+
+    src_state = src_state["state_dict"] if "state_dict" in src_state else src_state
+    src_state = src_state["model_state_dict"] if "model_state_dict" in src_state else src_state
+    src_state = src_state["model"] if "model" in src_state else src_state
+
+    if isinstance(scope_map, str):
+        scope_map = scope_map.split(",")
+    scope_map += ["module.", "None"]
+
+    for k in dst_state.keys():
+        k_src = k
+        if scope_map is not None:
+            src_prefix = ""
+            dst_prefix = ""
+            for i in range(0, len(scope_map), 2):
+                src_prefix = scope_map[i] if scope_map[i].lower() != "none" else ""
+                dst_prefix = scope_map[i + 1] if scope_map[i + 1].lower() != "none" else ""
+
+                if dst_prefix == "" and (src_prefix + k) in src_state.keys():
+                    k_src = src_prefix + k
+                    if not k_src.startswith("module."):
+                        logger.debug(f"init param, map: {k} from {k_src} in ckpt")
+                elif k.startswith(dst_prefix) and k.replace(dst_prefix, src_prefix, 1) in src_state.keys():
+                    k_src = k.replace(dst_prefix, src_prefix, 1)
+                    if not k_src.startswith("module."):
+                        logger.debug(f"init param, map: {k} from {k_src} in ckpt")
+
+        if k_src in src_state.keys():
+            if ignore_init_mismatch and dst_state[k].shape != src_state[k_src].shape:
+                logger.debug(
+                    f"ignore_init_mismatch:{ignore_init_mismatch}, dst: {k, dst_state[k].shape}, src: {k_src, src_state[k_src].shape}"
+                )
+            else:
+                dst_state[k] = src_state[k_src]
+
+        else:
+            logger.debug(f"Warning, miss key in ckpt: {k}, mapped: {k_src}")
+
+    obj.load_state_dict(dst_state, strict=True)

memo/utils/vision_utils.py

@@ -0,0 +1,100 @@
+import logging
+
+import cv2
+import numpy as np
+import torch
+from insightface.app import FaceAnalysis
+from moviepy.editor import AudioFileClip, VideoClip
+from PIL import Image
+from torchvision import transforms
+
+
+logger = logging.getLogger(__name__)
+
+
+def tensor_to_video(tensor, output_video_path, input_audio_path, fps=30):
+    """
+    Converts a Tensor with shape [c, f, h, w] into a video and adds an audio track from the specified audio file.
+
+    Args:
+        tensor (Tensor): The Tensor to be converted, shaped [c, f, h, w].
+        output_video_path (str): The file path where the output video will be saved.
+        input_audio_path (str): The path to the audio file (WAV file) that contains the audio track to be added.
+        fps (int): The frame rate of the output video. Default is 30 fps.
+    """
+    tensor = tensor.permute(1, 2, 3, 0).cpu().numpy()  # convert to [f, h, w, c]
+    tensor = np.clip(tensor * 255, 0, 255).astype(np.uint8)  # to [0, 255]
+
+    def make_frame(t):
+        frame_index = min(int(t * fps), tensor.shape[0] - 1)
+        return tensor[frame_index]
+
+    video_duration = tensor.shape[0] / fps
+    audio_clip = AudioFileClip(input_audio_path)
+    audio_duration = audio_clip.duration
+    final_duration = min(video_duration, audio_duration)
+    audio_clip = audio_clip.subclip(0, final_duration)
+    new_video_clip = VideoClip(make_frame, duration=final_duration)
+    new_video_clip = new_video_clip.set_audio(audio_clip)
+    new_video_clip.write_videofile(output_video_path, fps=fps, audio_codec="aac")
+
+
+@torch.no_grad()
+def preprocess_image(face_analysis_model: str, image_path: str, image_size: int = 512):
+    """
+    Preprocess the image and extract face embedding.
+
+    Args:
+        face_analysis_model (str): Path to the FaceAnalysis model directory.
+        image_path (str): Path to the image file.
+        image_size (int, optional): Target size for resizing the image. Default is 512.
+
+    Returns:
+        tuple: A tuple containing:
+            - pixel_values (torch.Tensor): Tensor of the preprocessed image.
+            - face_emb (torch.Tensor): Tensor of the face embedding.
+    """
+    # Define the image transformation
+    transform = transforms.Compose(
+        [
+            transforms.Resize((image_size, image_size)),
+            transforms.ToTensor(),
+            transforms.Normalize([0.5], [0.5]),
+        ]
+    )
+
+    # Initialize the FaceAnalysis model
+    face_analysis = FaceAnalysis(
+        name="",
+        root=face_analysis_model,
+        providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+    )
+    face_analysis.prepare(ctx_id=0, det_size=(640, 640))
+
+    # Load and preprocess the image
+    image = Image.open(image_path).convert("RGB")
+    pixel_values = transform(image)
+    pixel_values = pixel_values.unsqueeze(0)
+
+    # Detect faces and extract the face embedding
+    image_bgr = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
+    faces = face_analysis.get(image_bgr)
+    if not faces:
+        logger.warning("No faces detected in the image. Using a zero vector as the face embedding.")
+        face_emb = np.zeros(512)
+    else:
+        # Sort faces by size and select the largest one
+        faces_sorted = sorted(
+            faces,
+            key=lambda x: (x["bbox"][2] - x["bbox"][0]) * (x["bbox"][3] - x["bbox"][1]),
+            reverse=True,
+        )
+        face_emb = faces_sorted[0]["embedding"]
+
+    # Convert face embedding to a PyTorch tensor
+    face_emb = face_emb.reshape(1, -1)
+    face_emb = torch.tensor(face_emb)
+
+    del face_analysis
+
+    return pixel_values, face_emb

pyproject.toml

@@ -0,0 +1,83 @@
+[project]
+name = "memo"
+version = "0.1.0"
+description = "MEMO: Memory-Guided Diffusion for Expressive Talking Video Generation"
+readme = "README.md"
+requires-python = ">=3.10"
+license = {file = "LICENSE"}
+keywords = [
+    'artificial intelligence',
+    'computer vision',
+    'diffusion models',
+    'video generation',
+    'talking head',
+]
+
+dependencies = [
+    'accelerate==1.1.1',
+    'albumentations==1.4.21',
+    'audio-separator==0.24.1',
+    'black==23.12.1',
+    'diffusers==0.31.0',
+    'einops==0.8.0',
+    'ffmpeg-python==0.2.0',
+    'funasr==1.0.27',
+    'huggingface-hub==0.26.2',
+    'imageio==2.36.0',
+    'imageio-ffmpeg==0.5.1',
+    'insightface==0.7.3',
+    'hydra-core==1.3.2',
+    'jax==0.4.35',
+    'mediapipe==0.10.18',
+    'modelscope==1.20.1',
+    'moviepy==1.0.3',
+    'numpy==1.26.4',
+    'omegaconf==2.3.0',
+    'onnxruntime-gpu>=1.20.1',
+    'opencv-python-headless==4.10.0.84',
+    'pillow>=10.4.0',
+    'scikit-learn>=1.5.2',
+    'scipy>=1.14.1',
+    'torch==2.5.1',
+    'torchaudio==2.5.1',
+    'torchvision==0.20.1',
+    'transformers==4.46.3',
+    'tqdm>=4.67.1',
+    'xformers==0.0.28.post3',
+]
+
+[build-system]
+requires = ["setuptools", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools]
+packages = ["memo"]
+
+[tool.ruff]
+line-length = 119
+
+[tool.ruff.lint]
+# Never enforce `E501` (line length violations).
+ignore = ["C901", "E501", "E741", "F402", "F823" ]
+select = ["C", "E", "F", "I", "W"]
+
+# Ignore import violations in all `__init__.py` files.
+[tool.ruff.lint.per-file-ignores]
+"__init__.py" = ["E402", "F401", "F403", "F811"]
+
+[tool.ruff.lint.isort]
+lines-after-imports = 2
+known-first-party = ["vpt_x"]
+
+[tool.ruff.format]
+# Like Black, use double quotes for strings.
+quote-style = "double"
+
+# Like Black, indent with spaces, rather than tabs.
+indent-style = "space"
+
+# Like Black, respect magic trailing commas.
+skip-magic-trailing-comma = false
+
+# Like Black, automatically detect the appropriate line ending.
+line-ending = "auto"