Diversity-Preserved Distribution Matching Distillation for Fast Visual Synthesis

TL;DR

This paper introduces DP-DMD, a novel distillation method that preserves sample diversity in few-step image synthesis without additional regularization, enhancing stability and visual quality.

Contribution

The authors propose a role-separated distillation approach that maintains diversity without perceptual or adversarial regularization, improving stability and simplicity.

Findings

DP-DMD preserves sample diversity effectively.

DP-DMD achieves competitive visual quality with few steps.

DP-DMD is more stable and simpler than existing methods.

Abstract

Distribution matching distillation (DMD) facilitates few-step image generation by aligning a distilled student with a reference multi-step teacher. In practice, however, optimizing DMD can reduce sample diversity in few-step synthesis, and existing remedies typically rely on perceptual or adversarial regularization, leading to stability and scalability challenges during training. Here, we describe diversity-preserved DMD (DP-DMD), a role-separated distillation method inspired by the complementary roles of early and late denoising steps. Specifically, the first distillation step is trained with a teacher-derived target-prediction objective (e.g., v-prediction) to preserve sample diversity, while the remaining steps are optimized with the standard DMD loss to refine perceptual quality. DP-DMD, with no perceptual or adversarial regularization, no additional modules, and no teacher-generated reference samples, preserves sample diversity while maintaining competitive visual quality under few-step sampling, providing a simple and stable alternative to other DMD variants.