eDMD: Endpoint Distribution Matching for Few-Step Diffusion Distillation
Training against final outputs keeps the distillation signal from being distorted by intermediate denoising states.
eDMD trains one shared auxiliary “fake” model exclusively on the student’s final outputs, then uses it to correct every denoising step. The authors prove that its distribution-matching term vanishes if and only if the final output distribution matches the teacher’s. They also strengthen teacher supervision with short multi-step trajectories over the final noise interval. Across SD3-Medium, FLUX.2-klein 4B and 9B, and Qwen-Image-2.1, they report consistent gains over existing distribution-matching methods for four-step generation.
HF Daily Papers' note
eDMD trains one shared auxiliary “fake” model exclusively on the student’s final outputs, then uses it to correct every denoising step. The authors prove that its distribution-matching term vanishes if and only if the final output distribution matches the teacher’s. They also strengthen teacher supervision with short multi-step trajectories over the final noise interval. Across SD3-Medium, FLUX.2-klein 4B and 9B, and Qwen-Image-2.1, they report consistent gains over existing distribution-matching methods for four-step generation.
HF Daily Papers' note
score 5