/champ

Champ: Controllable and Consistent Human Image Animation with 3D Parametric Guidance

Primary LanguagePythonMIT LicenseMIT

Champ: Controllable and Consistent Human Image Animation with 3D Parametric Guidance

Shenhao Zhu*1Junming Leo Chen*2Zuozhuo Dai3Yinghui Xu2Xun Cao1Yao Yao1Hao Zhu+1Siyu Zhu+2
1Nanjing University 2Fudan University 3Alibaba Group
ECCV 2024
head.mp4

Framework

framework

News

  • 2024/05/05: 🎉🎉🎉Sample training data on HuggingFace released.

  • 2024/05/02: 🌟🌟🌟Training source code released #99.

  • 2024/04/28: 👏👏👏Smooth SMPLs in Blender method released #96.

  • 2024/04/26: 🚁Great Blender Adds-on CEB Studios for various SMPL process!

  • 2024/04/12: ✨✨✨SMPL & Rendering scripts released! Champ your dance videos now💃🤸‍♂️🕺. See docs.

  • 2024/03/30: 🚀🚀🚀Amazing ComfyUI Wrapper by community. Here is the video tutorial. Thanks to @kijai🥳

  • 2024/03/27: Cool Demo on replicate🌟. Thanks to @camenduru👏

  • 2024/03/27: Visit our roadmap🕒 to preview the future of Champ.

Installation

  • System requirement: Ubuntu20.04/Windows 11, Cuda 12.1
  • Tested GPUs: A100, RTX3090

Create conda environment:

  conda create -n champ python=3.10
  conda activate champ

Install packages with pip

  pip install -r requirements.txt

Install packages with poetry

If you want to run this project on a Windows device, we strongly recommend to use poetry.

poetry install --no-root

Inference

The inference entrypoint script is ${PROJECT_ROOT}/inference.py. Before testing your cases, there are two preparations need to be completed:

  1. Download all required pretrained models.
  2. Prepare your guidance motions.
  3. Run inference.

Download pretrained models

You can easily get all pretrained models required by inference from our HuggingFace repo.

Clone the the pretrained models into ${PROJECT_ROOT}/pretrained_models directory by cmd below:

git lfs install
git clone https://huggingface.co/fudan-generative-ai/champ pretrained_models

Or you can download them separately from their source repo:

  • Champ ckpts: Consist of denoising UNet, guidance encoders, Reference UNet, and motion module.
  • StableDiffusion V1.5: Initialized and fine-tuned from Stable-Diffusion-v1-2. (Thanks to runwayml)
  • sd-vae-ft-mse: Weights are intended to be used with the diffusers library. (Thanks to stablilityai)
  • image_encoder: Fine-tuned from CompVis/stable-diffusion-v1-4-original to accept CLIP image embedding rather than text embeddings. (Thanks to lambdalabs)

Finally, these pretrained models should be organized as follows:

./pretrained_models/
|-- champ
|   |-- denoising_unet.pth
|   |-- guidance_encoder_depth.pth
|   |-- guidance_encoder_dwpose.pth
|   |-- guidance_encoder_normal.pth
|   |-- guidance_encoder_semantic_map.pth
|   |-- reference_unet.pth
|   `-- motion_module.pth
|-- image_encoder
|   |-- config.json
|   `-- pytorch_model.bin
|-- sd-vae-ft-mse
|   |-- config.json
|   |-- diffusion_pytorch_model.bin
|   `-- diffusion_pytorch_model.safetensors
`-- stable-diffusion-v1-5
    |-- feature_extractor
    |   `-- preprocessor_config.json
    |-- model_index.json
    |-- unet
    |   |-- config.json
    |   `-- diffusion_pytorch_model.bin
    `-- v1-inference.yaml

Prepare your guidance motions

Guidance motion data which is produced via SMPL & Rendering is necessary when performing inference.

You can download our pre-rendered samples on our HuggingFace repo and place into ${PROJECT_ROOT}/example_data directory:

git lfs install
git clone https://huggingface.co/datasets/fudan-generative-ai/champ_motions_example example_data

Or you can follow the SMPL & Rendering doc to produce your own motion datas.

Finally, the ${PROJECT_ROOT}/example_data will be like this:

./example_data/
|-- motions/  # Directory includes motions per subfolder
|   |-- motion-01/  # A motion sample
|   |   |-- depth/  # Depth frame sequance
|   |   |-- dwpose/ # Dwpose frame sequance
|   |   |-- mask/   # Mask frame sequance
|   |   |-- normal/ # Normal map frame sequance
|   |   `-- semantic_map/ # Semanic map frame sequance
|   |-- motion-02/
|   |   |-- ...
|   |   `-- ...
|   `-- motion-N/
|       |-- ...
|       `-- ...
`-- ref_images/ # Reference image samples(Optional)
    |-- ref-01.png
    |-- ...
    `-- ref-N.png

Run inference

Now we have all prepared models and motions in ${PROJECT_ROOT}/pretrained_models and ${PROJECT_ROOT}/example_data separately.

Here is the command for inference:

  python inference.py --config configs/inference/inference.yaml

If using poetry, command is

poetry run python inference.py --config configs/inference/inference.yaml

Animation results will be saved in ${PROJECT_ROOT}/results folder. You can change the reference image or the guidance motion by modifying inference.yaml.

The default motion-02 in inference.yaml has about 250 frames, requires ~20GB VRAM.

Note: If your VRAM is insufficient, you can switch to a shorter motion sequence or cut out a segment from a long sequence. We provide a frame range selector in inference.yaml, which you can replace with a list of [min_frame_index, max_frame_index] to conveniently cut out a segment from the sequence.

Train the Model

The training process consists of two distinct stages. For more information, refer to the Training Section in the paper on arXiv.

Prepare Datasets

Prepare your own training videos with human motion (or use our sample training data on HuggingFace) and modify data.video_folder value in training config yaml.

All training videos need to be processed into SMPL & DWPose format. Refer to the Data Process doc.

The directory structure will be like this:

/training_data/
|-- video01/          # A video data frame
|   |-- depth/        # Depth frame sequance
|   |-- dwpose/       # Dwpose frame sequance
|   |-- mask/         # Mask frame sequance
|   |-- normal/       # Normal map frame sequance
|   `-- semantic_map/ # Semanic map frame sequance
|-- video02/
|   |-- ...
|   `-- ...
`-- videoN/
|-- ...
`-- ...

Select another small batch of data as the validation set, and modify the validation.ref_images and validation.guidance_folders roots in training config yaml.

Run Training Scripts

To train the Champ model, use the following command:

# Run training script of stage1
accelerate launch train_s1.py --config configs/train/stage1.yaml

# Modify the `stage1_ckpt_dir` value in yaml and run training script of stage2
accelerate launch train_s2.py --config configs/train/stage2.yaml

Datasets

Type HuggingFace ETA
Inference SMPL motion samples Thu Apr 18 2024
Training Sample datasets for Training Sun May 05 2024

Roadmap

Status Milestone ETA
Inference source code meet everyone on GitHub first time Sun Mar 24 2024
Model and test data on Huggingface Tue Mar 26 2024
Optimize dependencies and go well on Windows Sun Mar 31 2024
Data preprocessing code release Fri Apr 12 2024
Training code release Thu May 02 2024
Sample of training data release on HuggingFace Sun May 05 2024
Smoothing SMPL motion Sun Apr 28 2024
🚀🚀🚀 Gradio demo on HuggingFace TBD

Citation

If you find our work useful for your research, please consider citing the paper:

@inproceedings{zhu2024champ,
      title={Champ: Controllable and Consistent Human Image Animation with 3D Parametric Guidance},
      author={Shenhao Zhu and Junming Leo Chen and Zuozhuo Dai and Yinghui Xu and Xun Cao and Yao Yao and Hao Zhu and Siyu Zhu},
      booktitle={European Conference on Computer Vision (ECCV)},
      year={2024}
}

Opportunities available

Multiple research positions are open at the Generative Vision Lab, Fudan University! Include:

  • Research assistant
  • Postdoctoral researcher
  • PhD candidate
  • Master students

Interested individuals are encouraged to contact us at siyuzhu@fudan.edu.cn for further information.