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

Forum topic · 小凯 · 2026-06-27

Summary

DanceOPD (arXiv: 2606.27377) is an on-policy generative field distillation framework proposed by Wei Zhou, Xiongwei Zhu, and Zelin Xu that unifies text-to-image (T2I) generation, local editing, and global editing in a single image generation model. The core problem it addresses is that these capabilities are rarely naturally aligned and often conflict: adding editing abilities tends to degrade T2I performance, while global and local editing interfere with each other. DanceOPD progressively composes the three capabilities into one model using on-policy distillation, ensuring each capability is learned from the model's own generated distribution, which avoids misalignment and conflicts between tasks. The authors report that DanceOPD successfully combines all three capabilities, achieving strong performance on T2I, local editing, and global editing simultaneously without sacrificing any single capability. The paper was published on June 27, 2026.

Paper Overview

Research Area: CV / NLP / ML Authors: Wei Zhou, Xiongwei Zhu, Zelin Xu 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 that progressively composes text-to-image, local editing, and global editing capabilities into a single model. By leveraging on-policy distillation, DanceOPD ensures each capability is learned from the model's own generated distribution, avoiding misalignment and conflicts.

Key Contributions

Identifies capability conflicts (T2I vs. editing; global vs. local editing) as a central challenge in unified image generation model training.
Proposes on-policy generative field distillation, learning each capability from the model's own output distribution rather than misaligned external distributions.
Demonstrates that DanceOPD composes all three capabilities with strong performance on every task, without sacrificing any single capability.

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