README.md

---
title: SATA Demo
sdk: docker
app_port: 7860
suggested_hardware: cpu-basic
pinned: false
license: apache-2.0
---

# SATA: Semantic-Aware Motion Encoding for Topology-Agnostic Character Animation

[![Project Page](https://img.shields.io/badge/Project-Page-blue)](https://zzysteve.github.io/project_pages/sata.html)
[![arXiv](https://img.shields.io/badge/arXiv-b31b1b)](https://arxiv.org/abs/2605.27055)
[![Code](https://img.shields.io/badge/Code-GitHub-black)](https://github.com/zzysteve/SATA)

Official Hugging Face Space demo for **Semantic-Aware Motion Encoding for Topology-Agnostic Character Animation**.

SATA is a semantic-aware, topology-agnostic motion representation framework for heterogeneous character animation. It learns a unified latent motion manifold across diverse skeletal topologies and supports motion reconstruction, text-to-motion generation, and zero-shot cross-species retargeting.

## Hugging Face Space

This repository packages the Gradio demo for Hugging Face Spaces using the Docker SDK. The hosted Space is configured to run on CPU by default.

This Space copy keeps deployment-specific runtime adjustments:

- the app listens on the Hugging Face `PORT` environment variable, defaulting to `7860`;
- CPU is used by default through `SATA_DEVICE=cpu` and `SATA_GEN_DEVICE=cpu`;
- demo assets and checkpoints are included in the Space repository so the container can start without a separate artifact download step;
- MDM text-to-latent generation avoids SMPL-backed xyz conversion because the demo only saves latent `z` outputs.

If the Space is moved to GPU hardware later, set `SATA_DEVICE=cuda:0` and `SATA_GEN_DEVICE=cuda:0` in the Space environment.

Run command inside the container:

```shell
python app.py
```

## Main Repository

The source release is maintained at:

```text
https://github.com/zzysteve/SATA
```

## Citation

```bibtex
@inproceedings{zhang2026sata,
  title     = {Semantic-Aware Motion Encoding for Topology-Agnostic Character Animation},
  author    = {Zongye Zhang and Yuzhuo Cui and Qingjie Liu and Yunhong Wang},
  booktitle = {Proceedings of the International Conference on Machine Learning},
  year      = {2026},
  note      = {To appear. arXiv:2605.27055},
  url       = {https://arxiv.org/abs/2605.27055},
}
```

## Contact

For questions about the project, please contact the authors or open an issue after the initial code release.

## License

The original SATA code in this repository is released under the Apache License 2.0. See [LICENSE](LICENSE) and [NOTICE](NOTICE).

The following bundled external components remain governed by their respective upstream licenses:

```text
src/fairmotion       https://github.com/zzysteve/fairmotion.git
src/mdm              https://github.com/zzysteve/SATA-motion-diffusion-model.git
src/momask-preenc    https://github.com/zzysteve/SATA-momask-codes.git
```

Please refer to the license files in each component or upstream repository before using or redistributing those components. Files or assets that carry their own copyright or license notices remain governed by those notices. Model checkpoints, datasets, demo assets, body models, and other downloaded artifacts may have separate terms and are not covered by the top-level Apache License unless explicitly stated by their providers.