What it is
An action-conditioned video world model for real-time, long-horizon interactive rollout, optimized to run locally on one desktop GPU with streaming 720P output. Instead of offline clip generation or heavy bidirectional models, it distills a bidirectional action-conditioned teacher into a causal streaming world model, then uses LongForcing to align long self-rollouts and reduce autoregressive drift.
Key results
- 720P streaming at up to 16 FPS on a single RTX 5090
- 1.2 s action-to-first-frame latency end to end
- ~19 GiB peak VRAM in optimized low-bit deployment
Shows that long-horizon interactive world rollout can be made locally deployable by combining rollout-aligned distillation with systems-level inference optimization.