StepX-OPD: one student learns every skill and few-step sampling at once

TL;DR. A production text-to-image model needs several skills (instruction following, text rendering, human preference) and must sample in a few steps. Today those are built in two stages: multi-teacher on-policy distillation (OPD, where the student is supervised on states its own rollouts visit) first merges the skills into a full-step model, then step distillation compresses that model to a few steps. The authors find every full-step model loses score once compressed, and stage two cannot start until stage one ends. StepX-OPD merges the stages: frozen specialist teachers supervise a few-step student on the states its own short rollout visits. On SD3.5-Medium and FLUX.1-dev, four-step StepX-OPD gets the best normalized score among four-step methods and beats the forty-step Flow-OPD model, with about 37% fewer GPU-hours on SD3.5-Medium than the sequential pipeline.

flowchart LR
  P["Prompt<br/><small>text-to-image request</small>"] --> S["Few-step student<br/><small>4-step rollout</small>"]
  S --> V["Visited states<br/><small>student's own trajectory</small>"]
  T["Frozen specialists<br/><small>text, instruction, preference</small>"] --> K["Velocity matching<br/><small>KL on visited states</small>"]
  V --> K
  K -->|low-noise weight| S
  S --> O["4-step model<br/><small>all skills, 37% fewer GPU-hours</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class P,T input
  class S core
  class V,K loop
  class O exit
  linkStyle 5 stroke:#f08c00,stroke-width:2px

Key points

• The two-stage pipeline leaks quality. Every full-step model the authors tested scores lower after step distillation, so capability work is partly lost at compression time.
• Supervise the trajectory the student runs. Teachers grade the states the four-step student visits, not a long forty-step trajectory it will never run in deployment. This is the on-policy argument applied to the number of steps.
• Weight the clean end. Decomposing the rollout KL shows velocity mismatch matters most near the low-noise end, so the student also regresses its velocity on that part of the rollout.
• A side effect of merging skills. Integrating a text-rendering teacher makes models write prompt words into images that never asked for text. A variant reduces this.
• Results. Best normalized average among four-step methods on both SD3.5-Medium and FLUX.1-dev, above the forty-step Flow-OPD model, about 37% fewer GPU-hours than DiffusionOPD followed by step distillation.

Relation to prior wiki pages

• Extends the OPD cost story. The OPD vs GRPO page found distillation is only cheap when the teacher is close to the student. StepX-OPD's frozen specialists are domain teachers of the same family, the cheap regime.
• Confirms "compress where you deploy". Like the Extrapolation Cliff result (05-14, which found a teacher-student gap threshold above which on-policy distillation collapses), StepX-OPD shows the student should be supervised in its own operating regime, here four steps rather than forty.
• Diffusion-only so far. Whether the same merge works for LLMs (skill distillation plus reasoning-length compression in one stage) is untested.

Concept page: knowledge-distillation.