An explicit-state-driven framework for programmable generative worlds: a state executor maintains a canonical world state and advances it under player actions and world rules; a deterministic compiler projects the state, represented as state-augmented 3D OBBs, into pixel-aligned controls; and a video model renders the observation.
Video world models are powerful interactive renderers, but plausible observations alone are not a world engine. Programmable World Model decouples world-state maintenance and evolution from visual observation generation, making world facts queryable, modifiable, and verifiable — entities persist off-screen, events are irreversible, and the same world state drives any camera and any visual style.
- [2026-09-05]Project page released.
- Project page
- Technical report
- Inference code
- Pretrained weights
Coming soon.