BondiPAN/liveact-human-animation

LiveAct: Audio-Driven Human Animation

This Space provides a short-form, interactive demo of Soul-AILab/LiveAct. It accepts a portrait, a brief voice clip, and an optional prompt describing natural motion or expression, then returns a synthesized MP4 with the source audio muxed into the output. The implementation is adapted from the official SoulX-LiveAct inference repository for a single GPU request path.

Responsible use. Upload only images, recordings, and voices you own or are expressly authorized to use. Do not use this demonstration to impersonate real people, mislead audiences, generate non-consensual intimate material, or create harmful content. Treat all output as synthetic media and label it accordingly when shared.

How to use it

The Space runs one request at a time. Generation latency depends on the shared GPU queue and the length of the submitted audio. This demo is not a guarantee of real-time performance.

Implementation notes

LiveAct is a multimodal human-animation system that combines an image condition with audio and text controls. The upstream project reports real-time configurations on multi-GPU hardware and provides a consumer-GPU configuration using FP8 KV cache and model offloading. This Space uses a single-GPU, short-clip configuration designed for interactive evaluation rather than hour-scale streaming.

The app vendors only the upstream inference modules needed for LiveAct’s Wan backbone, VAE, audio encoder, and preprocessing. Model weights are downloaded at startup from Soul-AILab/LiveAct and TencentGameMate/chinese-wav2vec2-base; they are not stored in this repository.

License and attribution

The model card for Soul-AILab/LiveAct identifies the model as Apache-2.0 licensed. The application integrates and adapts code from the following projects; their respective licenses and notices continue to apply.

If you use LiveAct in academic work, please cite the original paper:

@misc{zhen2026soulxliveacthourscalerealtimehuman,
  title={SoulX-LiveAct: Towards Hour-Scale Real-Time Human Animation with Neighbor Forcing and ConvKV Memory},
  author={Dingcheng Zhen and Xu Zheng and Ruixin Zhang and Zhiqi Jiang and Yichao Yan and Ming Tao and Shunshun Yin},
  year={2026},
  eprint={2603.11746},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.11746}
}