DanceOPD: On-Policy Generative Field Distillation for Unified Image Generation

Forum topic · 小凯 · 2026-06-27

Summary

DanceOPD is an on-policy generative field distillation framework for flow-matching image generation models, presented in arXiv paper 2606.27377 by Wei Zhou, Xiongwei Zhu, and colleagues including Tat-Seng Chua. Modern image generation models must unify text-to-image (T2I) synthesis, local editing, and global editing, but these capabilities often conflict—editing degrades T2I performance, and global and local editing interfere with each other. DanceOPD addresses this by defining each capability source as a velocity field over a shared flow state space. The framework routes each training sample to a capability field, queries a low-noise student-induced state, and trains with a simple velocity MSE objective. Students learn from fields queried on their own rollout states, enabling composition of expert capabilities, and the formulation can also absorb operator-defined fields such as classifier-free guidance (CFG). Experiments across T2I, editing, realism field absorption, and CFG absorption show improved multi-capability composition while enhancing target capabilities and preserving anchor generation quality.

Overview

Research areas: Computer Vision, Computational Linguistics, Machine Learning Authors: Wei Zhou, Xiongwei Zhu, Zelin Xu, Bo Dong, Lixue Gong, Yongyuan Liang, Meng Chu, Leigang Qu, Lingdong Kong, Wei Liu, Tat-Seng Chua Published: 2026-06-27 arXiv: 2606.27377

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, the authors 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
Trains with a simple velocity MSE objective

Each capability source is defined as a velocity field over the shared flow state space. The student learns from fields queried on its own rollout states, allowing it to compose expert capabilities. This formulation also absorbs operator-defined fields, such as classifier-free guidance (CFG).

Results

Comprehensive experiments on T2I, editing, realism field absorption, and CFG absorption demonstrate that the method improves multi-capability composition—enhancing target capabilities while maintaining anchor generation quality.

--- *Auto-collected on 2026-06-27*