MOGO: Residual Quantized Hierarchical Causal Transformer for High-Quality and Real-Time 3D Human Motion Generation
MOGO: Residual Quantized Hierarchical Causal
Transformer for High-Quality and Real-Time 3D
Human Motion Generation
Dongjie Fu*
Tengjiao Sun*
Pengcheng Fang*
Xiaohao Cai
Hansung Kim
University of Southampton
MOGO
- [2024.06.03] Reorganize github
- [2024.05.21] Submit data process and evaluation algorithms
- [2024.11.23] Fix some bugs.
- [2024.08.04] Release model architecture.
- Code
- App.py
- Inference code of MoSA-VQ and Mogo
- Project Page
- Release pretrained weights train on HumanML3D
- Release pretrained weights train on our own made huge motion dataset
- Code of infinite length continuation and generation
- Controllable motion generation
conda create -n mogo python=3.10 -y
conda activate mogo
pip install -r requirement.txt
By using the following command, you can quickly generate an image with MOGO.
MOGO Generation
CUDA_VISIBLE_DEVICES=0 python gen_t2m.py
MoSA-VQ Generation
CUDA_VISIBLE_DEVICES=0 python gen_t2m.py
We provide a simple GUI for MOGO. You can easily run it by casting
python app.py
Thank you for your interest in this project. We are a startup team, if you are interested in our project, please contact us.
If you use MOGO or MoSA-VQ in your work, please use the following BibTeX entries:
@article{fu2024mogo,
title={Mogo: RQ Hierarchical Causal Transformer for High-Quality 3D Human Motion Generation},
author={Fu, Dongjie},
journal={arXiv preprint arXiv:2412.07797},
year={2024}
}