DanceOPD: On-Policy Generative Field Distillation
Wei Zhou, Xiongwei Zhu, Zelin Xu, Bo Dong, Lixue Gong, Yongyuan Liang, Meng Chu, Leigang Qu, Lingdong Kong, Wei Liu, Tat-Seng Chua
arXiv:2606.27377 · 81▲ · cs.CV, cs.CL, cs.LG
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
DanceOPD is an on-policy distillation method for flow-matching image generators that combines multiple frozen capability models (text-to-image, local editing, global editing) into one student. Each training sample is hard-routed to exactly one teacher's velocity field, that field is queried on a single low-noise state from the student's own rollout (stop-gradient), and the student is trained with plain velocity MSE.
Why it matters
It offers a way to ship one model that does T2I plus editing without the usual capability interference, where adding editing degrades generation quality or global and local edits fight each other. The single-query design also cuts per-step training cost compared to dense trajectory supervision, and the same objective absorbs classifier-free guidance into the model so part of the guidance runs in a single forward pass.
Practical takeaway
If you are post-training a flow-matching model to merge specialist checkpoints, try hard sample-wise routing to one teacher rather than averaging teacher outputs or merging weights, and query on student-generated states biased toward the low-noise end. Watch the CFG absorption caveat: if a model already absorbed guidance scale alpha, applying external CFG beta at inference multiplies to roughly alpha*beta and over-guides.
Key result
On T2I plus editing composition (Z-Image base), DanceOPD improved the GEditBench-EN average by 8.1% over the best reproduced on-policy distillation baseline and 8.5% over the edit-only source, while slightly exceeding the T2I source on GenEval (0.849 overall). Evaluations rely on GEditBench and GenEval automated metrics rather than human studies.
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.