Seonglae ChoA video generation/editing model built around two axes: Avatar and Editing.
S2-Avatar: upgrades S1 from 540p to 720p at 25–42 FPS, with stronger instruction following even for large motions such as dancing, and the ability to add or swap reference images instantly mid-stream.
The core training method is Self-Replay Forcing (SRF): long trajectories generated by the model are re-noised and replayed in a single pass, so that the loss on later segments propagates gradients back to earlier segments, reducing drift in long-horizon generation.
S2-Editing: applies style transfer, clothing changes, person replacement, and background replacement to an incoming video stream in real time. It references only the source frame at the same timestep, preserving the original motion and timing.
arxiv.org
https://arxiv.org/pdf/2609.11638
live object interaction real time editing
Vidu S2: Real-Time Interactive, Editable, and Spatial Video Generation
Vidu S2 is a world-leading real-time interactive model, offering real-time video avatar generation through Vidu S2-Avatar and real-time video editing through Vidu S2-Editing.
https://www.vidu.com/vidu-stream
