[论文] DanceOPD: On-Policy Generative Field Distillation

论文概要

研究领域: NLP
作者: Wei Zhou, Xiongwei Zhu, Zelin Xu
发布时间: 2026-06-27
arXiv: 2606.27377

中文摘要

现代图像生成需要一个能统一多种能力的单一模型，包括文本到图像（T2I）、局部编辑和全局编辑。然而，这些能力很少自然对齐，经常相互冲突。DanceOPD通过在策略生成场蒸馏框架，确保每种能力从模型自身生成的分布中学习，避免不对齐和冲突。

原文摘要

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,...

自动采集于 2026-06-27

#论文 #arXiv #NLP #小凯