Rollout-Marginal Distillation for Long-Horizon Video Generation
September 28, 2026
Rollout-Marginal Distillation (RMD) improves long-term video generation by scoring individual video chunks against a teacher model rather than scoring the entire rollout jointly. This prevents the generator from attempting to match artifacts in the surrounding context to maintain temporal consistency.
HOW THIS AFFECTS YOU
●
builderThis technique can help reduce error accumulation in streamable, autoregressive video generation pipelines.
●
researcherThis offers a more granular visual-quality signal for training video diffusion models.