eDMD: few-step diffusion distillation via endpoint distribution matching

Tong Zhao · hf · 2026-11-16

eDMD (Endpoint Distribution Matching Distillation), released on Hugging Face, improves few-step diffusion distillation: it trains a shared fake model only on the student's final outputs and adds precise teacher supervision to better align the few-step student's output distribution with the teacher's, enabling 4-step generation. Matching at the endpoint level targets final generation quality more directly than aligning intermediate steps.

Original post →

More from Multimodal

Multimodal channel →