PersonaLive

What it is

A research release (CVPR 2026) of a streamable diffusion framework that animates a single reference portrait using the facial motion of a driving video or webcam feed, producing videos of unbounded length. It ships pretrained weights, offline and online inference scripts, a browser-based streaming interface, optional TensorRT acceleration, and training code. Researchers and hobbyists working on talking-head generation can use it to drive avatars live or render long clips on a 12GB GPU.