DistillAlign: Coordinating Mode Covering and Mode Seeking in Autoregressive Video Distillation

Jiaxing Li, Kai Zou, Cindy Zhou, Kaichen Huang, Junyao Gao, Zile Wang, Yang Liu, Bin Liu, Bo An, Yangguang Li

arXiv:2607.26811 · 2026-07-31 공개 · arXiv · PDF

distillation autoregressive-video distribution-matching consistency-distillation reverse-kl video-distillation mode-seeking mode-covering

Abstract

Existing autoregressive video distillation methods commonly adopt a Distribution Matching Distillation (DMD)-based multi-stage pipeline. However, they typically decouple the initialization and DMD stages -- which then pursue different target distributions -- and judge the intermediate student mainly by visual scores such as VBench. In this paper, we revisit this design from a distributional perspective. Given the mode-seeking nature of the distribution matching loss, a good initialization should match the mode coverage of the target DMD teacher, rather than merely pursuing high quality. To analyze this, we introduce a distributional evaluation protocol that measures precision and coverage between student and teacher distributions in a shared latent space. It exposes differences hidden by visual scores: some initializations reach high precision but low coverage, leading to suboptimal refinement, while mode-covering ones preserve broader support. Furthermore, even when the target distributions are aligned, DMD's reverse-KL objective can still drive the student toward high-probability teacher regions in late training, reducing coverage and diversity. To address this, we propose joint distillation, which combines DMD's mode-seeking objective with a Consistency Distillation-based mode-covering constraint. Experiments show that our method improves generation quality, coverage, and diversity; notably, even with a Wan-1.3B DMD teacher, it outperforms baselines refined with Wan-14B, underscoring the importance of distributional alignment in autoregressive video distillation.

한국어 요약

한 줄 요약

DistillAlign은 DMD와 CD를 결합하여 초기화와 최종 정제 단계의 분포 일치를 동시에 달성하는 비디오 디스틸레이션 방법이다.

핵심 기여도

핵심 아이디어

기존 autoregressive video distillation 방법은 DMD 기반 다단계 파이프라인을 사용하지만, 초기화와 DMD 단계가 서로 다른 목표 분포를 추구하며, 초기화 평가는 주로 VBench 점수에 의존한다. 이는 분포적 관점에서 문제가 된다. DMD는 mode-seeking 성질을 가지므로, 초기화 단계에서 teacher의 mode coverage를 반영해야 DMD 정제가 효과적이다. 본 연구는 이 관찰을 바탕으로, 초기화 단계의 평가 기준을 mode coverage로 전환하고, DMD와 CD를 결합한 joint distillation을 제안한다. 이는 DMD의 reverse-KL 손실이 late stage에서 coverage를 줄이는 문제를 해결하며, mode-seeking과 mode-covering을 동시에 달성한다.

기술적 접근법

주요 결과

의의 및 한계

실용적 활용