WorldPlay2 team · Interactive World Model

WorldPlay2

A real-time interactive world model built on Wan2.2 I2V-A14B, steered with frame-aligned actions (camera pitch and yaw, movement, jump) plus structured captions that separate scene, character identity and events such as lighting a stove. A learned history compressor turns past frames into memory tokens about 32x shorter than the full context for long-horizon consistency, and a distillation method called Stable Forcing yields a 4-step autoregressive Fast model alongside 40-step variants. The paper was posted 28 September; weights and inference code arrived 6 October and a browser demo on 7 October. Comparisons against WorldPlay and LingBot-World-V2 are self-reported. Checkpoints are CC BY-NC 4.0, non-commercial.