Real-time interactive video world model
Odyssey-2
A live generated video world that responds to text while the stream is running.
01 / Overview
Causal video generation, live text interaction, open-ended visual continuation, low-latency world streaming.
A live generated video world that responds to text while the stream is running.
Judge instruction following, subject persistence, movement continuity, and recovery after a scene change.
Treat showcase clips as examples, not proof of a stable multi-minute world.
02 / Strengths
Where it stands out
- Lets a person influence a generated scene while it is unfolding.
- Makes the difference between fixed AI video and causal interactive video easy to feel.
- Useful for visual improvisation, scene exploration, and live concept experiments.
03 / Boundaries
What not to overclaim
- No claim of stable mesh, GLB, Gaussian-splat, or game-level export.
- Long sessions may reveal identity, geometry, memory, or action-following drift.
- Live availability and usage limits can change independently of the launch article.
Interactive streamed video, not a persistent 3D asset with editable geometry.
04 / Access & output
How the model reaches people
Public Odyssey experience; current access, queues, controls, and plans should be checked on the live site.
- Platforms
- Web experience · Streaming world-model service
- Outputs
- Interactive generated video stream
- Inputs & compatibility
- Text direction during a session · Browser video playback
05 / Practical path
How to evaluate it
- 01
Open the current Odyssey experience and confirm access before planning a session.
- 02
Begin with one clear place, subject, and direction of motion.
- 03
Add short instructions while the stream runs and watch continuity as well as visual quality.
- 04
Use a 3D world tool if the required deliverable is reusable spatial geometry.
Primary evidence