CoDMD: Copula-aware Distribution Matching Distillation for Fast Video Generation

2026-06-20

AI Summary

Existing few-step video diffusion distillation methods, such as DMD, suffer from layout instability, oversaturation, and motion distortion at low NFE due to neglecting structural dependencies across samples and frames. This work proposes CoDMD, which introduces a Copula-aware mechanism into video distillation for the first time. By incorporating a lightweight relational regularizer, CoDMD constructs pairwise cross-sample and cross-frame relation matrices from score estimates of both teacher and student models, enabling explicit modeling of spatiotemporal dependency structures through distribution matching to mitigate mode collapse. Notably, CoDMD requires no additional networks, data, or sampling trajectories. It achieves distillation from 50 to 4 steps (approximately 25× acceleration) on Wan-2.1-T2V 1.3B/14B models, attaining VBench scores of 84.46 and 84.87, outperforming current trajectory- and distribution-based distillation approaches.

Abstract

Few-step distillation for video diffusion models has attracted significant attention, driven by the urgent demand for efficient deployment in real-world scenarios. However, Distribution Matching Distillation (DMD), a leading paradigm, tends to degrade under limited NFE budgets, manifesting in video generation as layout instability, oversaturation, and broken motion dynamics. We trace this failure to a structural limitation: standard DMD is an intra-sample distribution-matching objective with coordinate-wise gradients, and thus imposes no explicit constraint on the relational geometry across batch elements or temporal frames, leaving the underlying copula largely unregulated. Combined with the mode-seeking tendency of its reverse-KL objective, this absence of relational guidance makes DMD prone to collapsing into local optima in the few-step regime. Motivated by this insight, we propose Copula-aware DMD (CoDMD), a lightweight relational regularizer that reuses score estimates already produced by the frozen teacher and the online fake model to construct pairwise relation matrices across samples and frames. These are matched through a supplementary distributional objective that requires no additional networks, datasets, or sampling trajectories. On the Wan-2.1-T2V model series at 1.3B & 14B scales, CoDMD distills 50-step teachers into 4-step students, achieving an approximate 25$\times$ speed-up while attaining VBench scores of 84.46 & 84.87, outperforming prior trajectory-based (rCM 82.81 & 84.05) and distribution-based (DMD 83.38 & 83.81) methods.

Problem

• video generation
• distribution matching distillation
• copula
• few-step distillation
• relational geometry

Innovation

• Copula-aware
• Distribution Matching Distillation
• Few-step Video Generation
• Relational Regularization
• Video Diffusion Models