AI RESEARCH
Diversity-Preserved Distribution Matching Distillation for Fast Visual Synthesis
arXiv CS.CV
•
ArXi:2602.03139v2 Announce Type: replace 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