Rollout-Marginal Distillation for Long-Horizon Autoregressive Video Generation
RMD separates chunk-quality correction from imperfect rollout context, then adds video-level distillation back for coherence.
The paper targets error buildup in long autoregressive video diffusion rollouts. Its method keeps generated history for prediction, but scores each chunk independently against a chunk teacher so quality fixes are not pulled toward artifacts in neighboring context. It then applies video-level DMD to recover temporal consistency. The authors report stronger long-horizon visual quality than video-level DMD baselines. HF Daily Papers' note
The paper targets error buildup in long autoregressive video diffusion rollouts. Its method keeps generated history for prediction, but scores each chunk independently against a chunk teacher so quality fixes are not pulled toward artifacts in neighboring context. It then applies video-level DMD to recover temporal consistency. The authors report stronger long-horizon visual quality than video-level DMD baselines. HF Daily Papers' note
score 4