SMooDi: Stylized Motion Diffusion Model

Project Page | Paper

SMooDi: Stylized Motion Diffusion Model
Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, Huaizu Jiang

Citation

If you find our code or paper helpful, please consider starring our repository and citing:

@article{zhong2024smoodi,
      title={SMooDi: Stylized Motion Diffusion Model},
      author={Zhong, Lei and Xie, Yiming and Jampani, Varun and Sun, Deqing and Jiang, Huaizu},
      journal={arXiv preprint arXiv:2407.12783},
      year={2024}
}

TODO List

Release retargeted 100STYLE dataset.
Code for Inference and Pretrained model.
Evaluation code and metrics.
Code for training.

Retargeted 100STYLE Dataset

We have released the retargeted 100STYLE dataset, mapped to the SMPL skeleton, available on Google Drive.

Processing Steps for the 100STYLE Dataset:

Retargeting with Rokoko: We used Rokoko to retarget 100STYLE motions to the SMPL skeleton template in BVH format. You can refer to this Video Tutorial for a detailed guide on using Rokoko.
Extracting 3D Joint Positions: After obtaining the retargeted 100STYLE dataset in BVH format, we utilized CharacterAnimationTools to extract 3D joint positions.
Deriving HumanML3D Features: Following the extraction, we used the instructions in the motion_representation.ipynb notebook available in HumanML3D to derive the HumanML3D features.

PRETRAINED_WEIGHTS

Available on Google Drive.

Getting started

This code requires:

Python 3.9
conda3 or miniconda3
CUDA capable GPU (one is enough)

1. Setup environment

Install ffmpeg (if not already installed):

sudo apt update
sudo apt install ffmpeg

For windows use this instead.

Setup conda env:

conda env create -f environment.yml
conda activate omnicontrol
python -m spacy download en_core_web_sm
pip install git+https://github.com/openai/CLIP.git

Download dependencies:

bash prepare/download_smpl_files.sh
bash prepare/download_glove.sh
bash prepare/download_t2m_evaluators.sh

2. Get data

Full data (text + motion capture)

HumanML3D - Follow the instructions in HumanML3D, then copy the result dataset to our repository:

cp -r ../HumanML3D/HumanML3D ./dataset/HumanML3D

100STYLE - Download the dataset from Google Drive, then copy the files in texts, new_joints, and new_joint_vecs into their corresponding directories within ./dataset/HumanML3D. We use indices larger than 030000 to represent data from the 100STYLE dataset.

3. Download the pretrained models

Download the model(s) you wish to use, then unzip and place them in ./save/.
Download the pretrained model from MLD and then copy it to ./save/.

Motion Synthesis

Please add the content text to ./demo/test.txt and the style motion to ./test_motion, then run:

bash demo.sh

Tips: For some motion styles, the default parameter settings may not achieve the desired results. You can modify the guidance_scale_style in config_cmld_humanml3d.yaml to achieve a better balance between content preservation and style reflection.

Train your own SMooDi

You can train your own model via

bash train.sh

Tips:

In config_cmld_humanml3d.yaml, setting is_recon: True means that cycle loss will not be used during training.
In fact, the improvement in performance from cycle loss is quite limited. If you want to quickly train a model, you can set is_recon: True. With this setting, it will take nearly 50 minutes to train 50 epochs on an A5000 GPU and achieve performance nearly equivalent to the second row in Table 3 of our paper.

Evaluate

You can evaluate model via

bash test.sh