StepX-OPD merges distillation stages for faster diffusion models
Original title: StepX-OPD: Rethinking Multi-Teacher On-Policy Distillation in Diffusion Models
StepX-OPD lets one student learn all capabilities and few-step sampling using its own rollout states, scoring highest among four-step methods on SD3.5-Medium and FLUX.1-dev while using 37% fewer GPU-hours.
StepX-OPD combines full-step capability consolidation and step distillation into one stage. A student model receives supervision from frozen specialists on states visited during its own rollouts. Evaluations show it exceeds the 40-step Flow-OPD baseline and reduces unwanted text rendering in images.
- StepX-OPD surpasses 40-step Flow-OPD on both SD3.5-Medium and FLUX.1-dev.
- Training uses about 37% fewer GPU-hours than sequential DiffusionOPD plus step distillation.
- Variants lower unwanted prompt words rendered into images.
- Highest normalized average score among four-step methods.
From the original article
1 min read at huggingface.co
A deployable text-to-image diffusion model must hold several capabilities at once, such as instruction following, text rendering and human preference, and must sample in a few steps. Capabilities and few-step sampling are built in separate stages: a full-step model first consolidates capabilities through multi-teacher on-policy distillation, then serves as a fixed teacher for step distillation.
By Chenru Wang, Xuanyang Zhang, 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.