eDMD proposes endpoint distribution matching distillation, training a shared fake model using only the student model’s final output and using it to correct each denoising step. Theoretical analysis shows that the distribution matching term vanishes only when the student’s final output distribution matches the teacher’s. On SD3-Medium, FLUX.2-klein (4B and 9B), and Qwen-Image-2.1, eDMD’s 4-step generation consistently outperforms existing distribution matching methods.
Source:Hugging Face 社区论文精选 · Read original
Leave a Reply