PDMD: Projected Distribution Matching Distillation for Faster Video Diffusion Models

Forum topic · 小凯 · 2026-09-30

Summary

PDMD (Projected Distribution Matching Distillation) is a new method that stabilizes distribution matching distillation (DMD) for video diffusion models. DMD reduces the number of function evaluations (NFE) of modern video diffusion models from tens to just a few, but its training samples can degrade with progressive oversaturation and artifacts. The authors trace this instability to critic errors that accumulate across successive student updates. PDMD filters these errors by projecting out the component of the DMD update parallel to the student-critic endpoint residual, which the authors prove is an unbiased estimate of the critic's endpoint error at a fixed noisy query. The projection removes a constant fraction of critic error while discarding only vanishing ideal DMD signal. Notably, PDMD requires only a one-line code change to DMD, with no extra losses, networks, data, or multi-stage training. Using Wan2.1, PDMD achieves a VBench total score of 83.73 at 4 NFE, beating equally trained DMD by 1.03 points. On MiniMax-H3 joint video-audio generation, it reaches 83.17 on VideoGen-Eval, surpassing the strongest distillation baseline by 0.41 and winning all six audio metrics, with user studies favoring PDMD.

Paper Overview

Research Area: Computer Vision (CV) Authors: Zimo Wang, Junkun Yuan, Angtian Wang, Haotian Yang, Canyu Zhang, Siyuan Yuan, Xingchang Huang, Bo Liu, Yizhi Wang, Yiding Yang, Chongyang Ma, Gordon Guocheng Qian Published: 2026-09-28 arXiv: 2609.35768

Abstract

Modern video diffusion models require tens of denoising evaluations over long spatiotemporal token sequences. Distribution Matching Distillation (DMD) reduces the number of function evaluations (NFE) to just a few. However, DMD samples can degrade during training, exhibiting progressive oversaturation and artifacts. The authors trace this instability to critic errors, which enter successive student updates and accumulate over time.

They introduce Projected Distribution Matching Distillation (PDMD) to filter critic errors. PDMD projects out the component of the DMD update parallel to the student-critic endpoint residual. At a fixed noisy query, the paper proves that this residual is an unbiased estimate of the critic's endpoint error. Under high-dimensional assumptions, this projection removes a constant fraction of the critic error while only discarding ideal DMD signal that vanishes.

Empirically, the projection stabilizes training and improves sample quality in cases where DMD degrades and produces unnatural textures. PDMD requires only a one-line code modification to DMD, with no additional losses, networks, data, model passes, or multi-stage training.

Key Results

Wan2.1: PDMD reaches a VBench total score of 83.73 at 4 NFE, outperforming equally trained DMD by 1.03 points.
MiniMax-H3 joint video-audio generation: PDMD achieves a VideoGen-Eval visual total score of 83.17, 0.41 higher than the strongest distillation baseline, and achieves the best scores on all six audio metrics.
User studies prefer PDMD in visual quality, motion, and audio quality.

Why It Matters

Stabilizes DMD training by removing accumulating critic error via a principled projection step.
Practically free: a single-line code change, no extra cost in models, data, or training stages.
Enables few-step (4 NFE) video generation with state-of-the-art visual and audio quality.

--- *Auto-collected on 2026-09-30*