StepFun's StepX-OPD merges capability learning and few-step distillation for diffusion models
stepfun-ai · hf · 2026-11-07
StepFun and Westlake University released StepX-OPD, rethinking multi-teacher on-policy distillation for text-to-image diffusion models.
- Conventional pipelines are two-stage: a full-step model consolidates capabilities via multi-teacher distillation, then serves as a fixed teacher for step distillation; every full-step model scores worse once compressed, so stage two can't pick up where stage one ended
- StepX-OPD merges the stages: one student acquires capabilities while learning few-step sampling, with frozen specialists supervising states its own rollouts actually visit
- Since KL velocity mismatch concentrates near the clean end, the student's velocity is additionally regressed on the low-noise part of the rollout
The authors report it outperforms sequential approaches while cutting training cost.
More from Multimodal
- Wabi already offers shareable AI spaces, answering calls aimed at Meta — joonasvirtanen · 2026-11-07
- Premiere Pro Beta Adds Generative Remove for Tracking and Erasing Objects on Timeline — Mr_AllenT · 2026-11-07
- Phone-captured scenes reconstructed at stunning quality with Echo-3 — risingodegua · 2026-11-07
- AI rotoscoping workflow today: LTX 2.5 Alpha LoRA + SAM3 + BiRefNet, layered ProRes — PurzBeats · 2026-11-07
- Custom H3 I2V/REF2V ComfyUI workflow: 20s video in 22 minutes on a 3090, controlling 5 characters — EQLMonk · 2026-11-07
- Flux Max test: mantis shrimp vs white shark prompt fails hilariously, rest looks cinematic — flowersslop · 2026-11-07