Reproduce “InfinityHuman: Towards Long-Term Audio-Driven Human” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaodi Li, Pan Xie, Yi Ren, Qijun Gan, Chen Zhang, Fangyuan Kong, Xiang Yin, Bingyue Peng and 1 more
Researchers from ByteDance and Zhejiang University developed InfinityHuman, a system that generates high-resolution, long-duration audio-driven human animations while maintaining visual consistency and natural hand motions. It achieves state-of-the-art results on multiple benchmarks by addressing identity drift and hand distortions that typically plague longer generated videos.
Key insights
- Utilizing pose information as a stable, structural condition within a refiner effectively guides high-resolution generation and combats appearance-related temporal degradation.
- Training the refiner with a
- degraded
- low-resolution latent input allows it to learn to recover details and correct accumulated errors, making it robust to temporal inconsistencies.
- Applying preference learning with a specialized reward model for hands enables the generation of anatomically correct and temporally consistent hand gestures without requiring explicit hand annotations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.