preprint · arXiv (2026) · Mar 15, 2026
AvatarForcing: One-Step Streaming Talking Avatars via Local-Future Sliding-Window Denoising
Liyuan Cui, Wentao Hu, Wenyuan Zhang, Zesong Yang, Fan Shi, Xiaoqiang Liu
In plain English
A one-step streaming system that turns a portrait and live audio into a talking-avatar video. It denoises a short window of upcoming frames at once and reuses recent clean frames to keep appearance stable over minutes, running at 34 ms per frame with a 1.3-billion-parameter model.
talking head generation
real-time avatars
diffusion
inference speed