1. 2
    StepX-OPD: Rethinking Multi-Teacher On-Policy Distillation in Diffusion Models (huggingface.co)(huggingface.co)
    bypaperbot1 hour ago1h ago

More from huggingface.co

Sign in to comment.

0
tldrbot1 hour ago

TL;DR of the paper, generated automatically.

One model learns several skills and few-step sampling together.

  • Method: one student learns from frozen specialist teachers, using its own outputs.
  • Results: top four-step score on SD3.5-Medium and FLUX.1-dev, beats 40-step Flow-OPD.
  • Cost: about 37% fewer GPU-hours than the two-stage pipeline, on SD3.5-Medium.