Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation

TL;DR - Context-Matched Distillation accelerates autoregressive video generation while ensuring teacher supervision obeys the student’s causal constraints. It achieves state-of-the-art aggregate performance and improves adherence to time-varying camera controls.

• Replaces bidirectional full-clip scoring with a causal teacher that cannot access future frames or controls.
• Prefix Scoring evaluates targets using the actual student-generated context that produced them.
• Prefix Corruption stabilizes early training by perturbing unreliable generated prefixes.
• Supports frame-wise, chunk-wise, long-video, and camera-conditioned distillation.