First-hand information for everyone
ScaleWorld is a new causal video model that combines low‑resolution world prediction with high‑resolution refinement in one shared backbone. The system can produce fine‑grained video worlds without a separate refinement network, saving computation and memory.
The model trains a shared backbone that handles two distinct tasks. The prediction branch uses a long low‑resolution history, while the refinement branch works with a short, noisy‑forward high‑resolution history. Training separates specialist optimization from the final consolidation stage. First, prediction and refinement specialists are trained through distribution‑matching distillation. Then a task‑asymmetric on‑policy distillation (TA‑OPD) stage transfers their behaviors into the shared student using local velocity matching.
On the WBench benchmark, the unified ScaleWorld model reaches an overall average score of 81.60, slightly higher than a cascade of separate prediction and refinement specialists (81.58). Across individual metrics—video quality, setting adherence, interaction adherence, consistency, and physical realism—the scores are comparable to existing methods.
The authors note that further weaknesses are discussed in Appendix E, indicating that the model is not without limitations.
Visual evidence: the comparison table from WBench shows the average scores for the shared model and the separate cascade, confirming the model’s competitive performance.