SDMD improves few-step image generation by better matching data distributions
sDMD: Sinkhorn Distribution Matching Distillation
Computer Vision and Pattern RecognitionMachine Learning
Summary
Generating high-quality images quickly is difficult because detailed models usually require many slow steps. The researchers studied a method called distribution matching distillation (DMD), which trains faster models to mimic slower ones by focusing on local data details. They found that DMD sometimes misses important variety in images and struggles with rare cases. Their new method, sDMD, adds a way to gently guide samples towards the right areas without disrupting well-learned parts. This helps generate diverse and accurate images using fewer steps than before.
What this means in practice
- •For image generation developers: Create faster image generation tools using fewer steps with improved quality and diversity compared to previous distillation methods.
- •For computer vision engineers: Enhance image synthesis systems by reducing errors in rare or outlier cases during model distillation for better robustness.
Authors
Jaihoon Kim, Lingxiao Li, Soshi Shimada, Roshni Cooper, Minhyuk Sung, Yiwei Zhao
Abstract
Distribution Matching Distillation (DMD) distills multi-step diffusion models into few-step generators by minimizing reverse KL divergence. Despite its strong performance, DMD guides samples by local density alone, which introduces two limitations: teacher scores can be unreliable in regions of low teacher density, and the mode-seeking behavior of reverse KL can collapse the student onto a subset of the teacher modes. We first connect DMD to Wasserstein gradient flow, showing that its gradient follows from a fixed-point formulation with integral KL (IKL) as the energy functional. This perspective motivates augmenting IKL with an optimal transport term that addresses these limitations by guiding outlier samples toward the target and promoting mode coverage. We realize this term using Sinkhorn divergence estimated from finite batches. In practice, however, errors in the estimated Sinkhorn velocity can perturb in-domain samples for which IKL guidance is already accurate. To this end, we introduce sDMD, which adaptively activates the Sinkhorn correction based on discriminator outputs, focusing on outlier samples where IKL may be unreliable. With 2--4 sampling steps on FLUX.1-dev, Qwen-Image-20B, and Wan2.1-14B, sDMD outperforms distillation baselines on GenEval, DPG, WISE, and VBench while improving sample diversity over DMD2.