Autoregressive video generate long videos chunk by chunk, but errors accumulate over time. Standard (DMD) scores the whole rollout jointly, which forces the teacher to tolerate artifacts in the generated history to preserve temporal consistency. Rollout-Marginal Distillation (RMD) instead scores each chunk independently against a teacher that has been adapted to judge chunks, providing a clean signal for local visual quality. A subsequent video-level DMD stage restores temporal coherence.
Independent chunk scoring reduces drift in long autoregressive video rollouts
The authors separate per-chunk visual correction from sequence-level temporal refinement, improving quality well beyond the model’s five-second training horizon.
Chinese Tech
Chenjian Gao · Zhihao Hu · Jianqi Ma · Jun Zhang · Weidong Zhang · Tianfan Xue
MMLab · The Chinese University of Hong Kong · Tencent AIPD
Research Digest··2 min read
In one line
Rollout-Marginal Distillation scores each generated chunk independently against a teacher, then refines temporal coherence, preventing error accumulation in long autoregressive video generation.
Before you read this
Three ideas this paper builds ondistribution matching distillationA training method that aligns student and teacher output distributions to compress many-step generative models into efficient few-step versions.29 papers →on-policy distillationA training method where a student model learns from a teacher using its own current policy's outputs, often improving reasoning and instruction following.134 papers →diffusion modelA type of generative model that learns to reverse a gradual noising process to create data.199 papers →
Read it as
Gao et al. address error accumulation in autoregressive video diffusion, where imperfect generated frames become the context for subsequent chunks.
The authors study streamable video that generate successive chunks while conditioning on a sliding window of previously generated frames.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.