StepX-OPD Rethinking Multi-Teacher On-Policy Distillation in Diffusion Models

Capability integration and trajectory compression, trained together on the student’s own rollout.

Chenru Wang1,*, Xuanyang Zhang1, Yucheng Han1, Wenbo Nie1,*, Wei Cheng1, Xianfang Zeng1, Gang YU1, Chi Zhang2

1StepFun·2Westlake University

*Work done during an internship at StepFun.

Spillover examples

Prompt words rendered as image text

These prompts neither request nor forbid text in the image. The figures below collect cases where words from the prompt are nevertheless drawn into the picture.

Examples of unintended prompt-word rendering.
Further examples. Further examples of unintended prompt-word rendering across compared models.
Examples of unintended prompt-word rendering.
Further examples. More examples showing how prompt-word recurrence varies across model outputs.