Distribution Matching Distillation

Also known as DMD

Overview

Distribution Matching Distillation transfers the behavior of a slow diffusion or flow-model teacher to a student that generates images, video, or other data in only a few steps. Many methods use teacher and student score estimates to form distribution-matching gradients, while alternatives learn density ratios adversarially or use consistency and flow-map objectives. Recent approaches also combine matching with diversity-preserving constraints, preference rewards, or temporal guidance for autoregressive video. The central tension is to gain speed without sacrificing sample quality, diversity, or long-range dynamics, especially when critic errors or limited student capacity distort training.

Latest papers

CardsList

34 papers7 papers in the last four weeks, down 30% on the four weeks before. 0.1% of all new papers.

  1. 1.x-Distill: Breaking the Diversity, Quality, and Efficiency Barrier in Distribution Matching Distillation

    Apr 5, 2026Haoyu Li, Tingyan Wen, Lin Qi +6Diffusion Model DistillationDiffusion Sampling

  2. Salt: Self-Consistent Distribution Matching with Cache-Aware Training for Fast Video Generation

    Apr 3, 2026Xingtong Ge, Yi Zhang, Yushi Huang +6Diffusion Model DistillationVideo Generation

  3. DUO-VSR: Dual-Stream Distillation for One-Step Video Super-Resolution

    Mar 23, 2026Zhengyao Lv, Menghan Xia, Xintao Wang +1Video Super-ResolutionDiffusion Model Distillation