Paper 2606.27377
DanceOPD: On-Policy Generative Field Distillation
- Published
- Jun 2026
- Research lab
- Independent
- Citations
- 2
- GitHub
- 395 stars
01 In brief
Summary
DanceOPD is an on-policy generative field distillation framework for flow-matching image generation models, designed to compose multiple capabilities (e.g., text-to-image, local editing, global editing) into a single student model.
The method treats each frozen capability source as a velocity field over a shared state space and addresses three key challenges: target-field ambiguity, state-distribution mismatch, and trajectory-query correlation.
It uses hard-routed sample-wise field matching, queries the selected field on a single low-noise state from the student's own rollout, and trains with a plain velocity MSE objective.
Experiments on T2I and editing composition, local and global editing composition, realism-field absorption, and CFG absorption show that DanceOPD improves target capabilities while preserving anchor generation quality.
For example, it improves GEditBench by 8.1% over the best OPD baseline in T2I+Edit composition, and by 16.1% over the best baseline in local+global edit composition.
Ablations confirm the importance of hard routing, on-policy querying, and single semantic-side queries.
02 From the paper
Abstract
Modern image generation demands a single model that unifies diverse capabilities, including text-to-image (T2I), local editing, and global editing. However, these capabilities are rarely naturally aligned and often conflict. For instance, editing tends to degrade T2I performance, while global and local editing interfere with each other. Consequently, effectively composing these capabilities has become a central challenge for image generation model training. To tackle this, we introduce DanceOPD, an on-policy generative field distillation framework for flow-matching models that routes each sample to one capability field, queries one low-noise student-induced state, and trains with a simple velocity MSE objective. With each capability source defined as a velocity field over the shared flow state space, the student learns from fields queried on its own rollout states to compose expert capabilities. This formulation also absorbs operator-defined fields such as classifier-free guidance. Comprehensive experiments on T2I, editing, realism-field absorption, and CFG absorption show that our approach improves multi-capability composition, strengthening target capabilities while preserving anchor generation quality. We believe this work establishes a practical route for generative field distillation in flow-matching models.