What it is
An interactive video world model with external persistent memory and a linear-scaling long-horizon teacher, distilled into a 3-step student for open-ended generation. Instead of keeping all history in the denoiser or KV cache, it externalizes scene state into a camera-indexed memory bank and trains a 3-step student from a linear-scaling long-horizon teacher.
Key results
- SOTA on WBench as a 3-step world model
- Competitive with prior work on VBench-Long and VBench-2.0
- Generates each 1.5s chunk in 2.11s on a single H200 at 384x640
Long-horizon supervision plus external recurrent world state can reduce drift without sacrificing low-latency interactive generation.