SolarWM opens long-horizon video worlds

SolarWM’s biggest contribution is reproducibility: it packages the messy data, training, and inference stack needed to make long-horizon world-model research practical. The long-rollout claims are promising, but sustained visual fidelity and error accumulation remain the real tests.

• –A unified pipeline standardizes observations, camera geometry, captions, quality metadata, and provenance across heterogeneous video sources
• –One framework supports four 5B–33B models built on Wan2.2, LTX-2.5, and MiniMax-H3
• –The three-stage recipe combines bidirectional adaptation, teacher forcing, and self-gradient-forcing distillation
• –Open weights and data let developers experiment locally and compare backbones without rebuilding the entire stack
• –The project still needs stronger quantitative benchmarks for minute- and hour-scale coherence, controllability, and physical consistency