eDMD matches endpoints for few-step diffusion distillation
Original title: eDMD: Endpoint Distribution Matching for Few-Step Diffusion Distillation
One shared fake model is trained only on final student outputs and improves 4-step generation.
Endpoint Distribution Matching Distillation trains a single shared fake score model exclusively on a student's final outputs and uses it to correct every denoising step. The authors prove the matching term vanishes if and only if the student's final distribution matches the teacher's, regardless of intermediate states. A precise teacher replaces single-step guided predictions with short multi-step trajectories on the final noise interval. Experiments on SD3-Medium, FLUX.2-klein (4B and 9B) and Qwen-Image-2.1 report better 4-step results than existing distribution matching methods.
- Existing methods train the fake model on intermediate states or all steps.
- eDMD uses one shared fake model on final student outputs only.
- The precise teacher uses short trajectories on the final noise interval.
- Evaluated models include SD3-Medium, FLUX.2-klein and Qwen-Image-2.1.
- Reported gains are for 4-step generation.
- AI models
- SD3-MediumFLUX.2-kleinQwen-Image-2.1
From the original article
1 min read at huggingface.coDistribution matching distillation (DMD) has emerged as a leading paradigm for distilling diffusion models into few-step students, where an auxiliary fake score model tracks the student's distribution and the discrepancy between the teacher and fake predictions drives the student toward the teacher.
By Tong Zhao, Xianfang Zeng, Yucheng Han · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.