Diversity-Preserved Distribution Matching Distillation for Fast Visual Synthesis

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.

DP-DMD is a simple way to make fast diffusion models more diverse. Standard DMD can generate good images in only a few steps, but it often reduces diversity. DP-DMD solves this by splitting the roles of the steps: the first step is trained to preserve variety, while the later steps use DMD to improve image quality. This keeps generation fast and stable, without extra adversarial or perceptual losses.