Reproduce “AvatarForcing: One-Step Streaming Talking Avatars via Local-Future Sliding-Window Denoising” 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
Liyuan Cui, Wentao Hu, Wenyuan Zhang, Zesong Yang, Fan Shi, Xiaoqiang Liu
A streaming diffusion framework for real-time talking avatar generation, named AvatarForcing, was developed by researchers from Zhejiang University and Kling Team (Kuaishou Technology). The system achieves minute-level temporal stability, identity preservation, and accurate lip synchronization with a latency of 34 ms per frame, outperforming existing methods on long-form video benchmarks.
Key insights
- Bounded look-ahead within a fixed sliding window provides more effective long-horizon stability than increasing denoising passes within a shorter context, offering implicit refinement while maintaining constant latency.
- Maintaining persistent identity and appearance over extended durations requires explicit anchoring mechanisms, with RoPE re-indexing being critical for a style anchor to remain relevant as absolute time progresses.
- Distilling a complex bidirectional teacher through a two-stage process, which includes distribution matching adapted to the student's specific rolling-window inference dynamics, enables high-quality real-time performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.