Odyssey-2 Streams Interactive AI Video At 20fps, Shaped By Typing
- Odyssey-2 streams interactive AI video in real time, generating a new frame every 50 milliseconds, or 20 frames per second.
- Odyssey says most bidirectional video models take 1 to 2 minutes to generate only 5 seconds of footage, while Odyssey-2 starts streaming immediately after a text prompt.
- The model became causal and autoregressive, generating each frame only from prior frames and user actions, via a novel multi-stage training pipeline that transitions it away from bidirectional generation.
- Users reshape the running video stream in real time with natural text prompts, producing continuous multi-minute video, with audio/open-ended input and navigation conditioning named as further research.
- Oliver Cameron published the model on October 27th, 2025 through Odyssey, calling it a world model that approximates physical dynamics like wave motion learned from decades of video data.