What it is
An omnimodal world model that turns video generation into a navigable world, jointly producing 720p video, environmental audio, music, and speech from trajectory control. Unlike prior world models that are mostly visual or use domain-specific controls, it uses one camera-intent interface: discrete commands and poses map to a shared metric 6-DoF trajectory for both first- and third-person audio-visual generation.
Key results
- #1 on WBench Navigation
- Top-tier visual quality on SANA-WM-Bench simple and hard splits
- Supports 720p long-horizon generation with synchronized sound and speech
A single camera-centric 6-DoF control interface can span first- and third-person interactive world generation without separate view-specific controllers.