Rollout-Marginal Distillation for Long-Horizon Autoregressive Video Generation
Rollout-Marginal Distillation scores video chunks independently so long autoregressive rollouts keep visual quality.
Autoregressive video diffusion enables low-latency generation, but errors accumulate over long rollouts, and joint video-level distribution matching distillation can favor consistency with surrounding artifacts. Rollout-Marginal Distillation keeps generated history for prediction yet scores each chunk independently against a chunk teacher, then applies video-level DMD to restore temporal coherence. Experiments show high visual quality well beyond the training horizon and better results than video-level DMD baselines.
- Joint video DMD can reward matching artifacts in surrounding context.
- RMD scores each chunk independently against a chunk teacher.
- A later video-level DMD pass restores temporal coherence.
- Quality holds beyond the training horizon versus DMD baselines.
From the original article
1 min read at huggingface.co
Autoregressive (AR) video diffusion enables low-latency, streamable video generation, but prediction errors often accumulate over long rollouts. Training the generator on its own rollouts exposes it to these imperfect histories. However, existing video-level distribution matching distillation (DMD) scores the whole rollout jointly.
By Chenjian Gao, Zhihao Hu, Jianqi Ma · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.