Vidu S2, a new research release on Hugging Face, introduces two real-time models: Vidu S2-Avatar, an interactive digital-character system, and Vidu S2-Editing, a real-time video editing model. The work also explores the feasibility of real-time spatial video generation for both systems.
According to the paper, Vidu S2-Avatar supports real-time 720p video generation at 25–42 FPS, dynamic references that can be updated at any moment, and stronger instruction following, including dancing and expressive full-body motion. Users can introduce new reference images mid-conversation to change outfits, interact with objects, or switch scenes.
Vidu S2-Editing transforms an incoming video stream in real time while preserving motion and timing. Supported edits include style rendering or transfer, virtual try-on and clothing replacement, character replacement, and background replacement.
The authors state that Vidu S2 outperforms all baselines and achieves state-of-the-art results across all five public benchmarks in their evaluation, spanning avatar generation and video editing. They also describe bringing generated characters and live video edits into stereoscopic video for immersive VR experiences.
A playable online demo is available at vidu.com/vidu-stream, with API documentation at platform.vidu.com/vidu-stream/doc and the technical report at arxiv.org/pdf/2609.11638.