Paper 2607.26811
DistillAlign: Coordinating Mode Covering and Mode Seeking in Autoregressive Video Distillation
- Published
- Jul 2026
- Research lab
- Independent
- Citations
- 0
- GitHub
- 97 stars
01 In brief
Summary
The paper DistillAlign revisits autoregressive video distillation from a distributional perspective.
It argues that existing multi-stage pipelines, which separate initialization (e.g., ODE or consistency distillation) from DMD refinement, often have misaligned target distributions.
Since DMD is mode-seeking, a good initialization must match the mode coverage of the DMD teacher, not just achieve high visual scores.
The authors introduce a teacher-normalized evaluation protocol measuring precision and coverage in a shared latent space, revealing that some initializations have high precision but low coverage, leading to suboptimal refinement.
They also show that even with aligned targets, pure DMD causes late-stage distribution drift, reducing diversity.
To address this, they propose joint distillation, combining DMD's mode-seeking objective with a consistency-distillation-based mode-covering constraint.
Experiments show that their method improves quality, coverage, and diversity; notably, with a Wan-1.3B DMD teacher, it outperforms baselines refined with Wan-14B, highlighting the importance of distributional alignment.
02 From the paper
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.