SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models
SolarWM packages the data, training recipe, model weights, and framework needed to reproduce long-horizon video world model experiments.
The paper says its data engine standardizes 1.43 million clips from 10 datasets into a shared frame-aligned format with camera geometry, captions, quality metadata, decisions, and provenance. It adapts four 5B to 33B video models based on Wan2.2, LTX-2.5, and MiniMax-H3 while preserving their native representations. The authors claim models trained on 5-second sequences can support real-time interactive rollouts lasting minutes to hours. HF Daily Papers' note
The paper says its data engine standardizes 1.43 million clips from 10 datasets into a shared frame-aligned format with camera geometry, captions, quality metadata, decisions, and provenance. It adapts four 5B to 33B video models based on Wan2.2, LTX-2.5, and MiniMax-H3 while preserving their native representations. The authors claim models trained on 5-second sequences can support real-time interactive rollouts lasting minutes to hours. HF Daily Papers' note
score 6