PRIMAL: Physically Reactive and Interactive Motor Model for Avatar Learning

This repo is developed based on pytorch-lightning, hydra, huggingface libs, and others.

project page

Documentation

The documentation has been moved to the docs/ directory to keep the project organized.

• Installation Guide: Setting up the environment (conda/poetry) and data.
• Unreal Engine Demo: Instructions for the UE demo backend (see Installation).
• Experiment Guide: Detailed instructions for running training, inference, and custom experiments.
• SLURM Cluster Guide: Instructions for running jobs on SLURM clusters.
• Acceleration Model Guide: Specifics about the acceleration-based model variant.

Quick Start

1. Inference with Gradio

Run the gradio demo for the base model:

poetry run python demos/ARDiffusion_gradio.py logs/motion_diffuser_ar/runs/silu_anchor

For the action generation model:

poetry run python demos/ARDiffusionAction_gradio.py logs/motion_diffuser_ar_action/runs/ours

2. Training

To pretrain the base model:

python scripts/train.py --config-name=train_diffusion task_name=motion_diffuser_ar

For detailed training instructions, including adaptation and custom datasets, please refer to the Experiment Guide.

License and Citation

See license file for more details. Please cite the following work if it helps.

@inproceedings{primal:iccv:2025,
  author = {Zhang, Yan and Feng, Yao and Cseke, Alpár and Saini, Nitin and Bajandas, Nathan and Heron, Nicolas and Black, Michael J.},  
   title = {{PRIMAL:} Physically Reactive and Interactive Motor Model for Avatar Learning},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month = oct,
  year = {2025}
}