Paper Overview
Field: NLP Authors: Wei Zhou, Xiongwei Zhu, Zelin Xu Posted: 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. The core idea is to ensure that each capability learns from the model's own generated distribution, rather than from an off-policy or external source. This on-policy design avoids misalignment and prevents the different capabilities from conflicting with each other during training.
Key Points
- Unifies three capabilities in one model: text-to-image generation, local editing, and global editing.
- Identifies mutual interference: editing degrades T2I quality; global and local editing conflict with each other.
- Proposes on-policy generative field distillation so each capability is learned from the model's own output distribution.
*Auto-collected on 2026-06-27*