arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2508.06044cs.CV

NEP:基于下一编辑 token 预测的自回归图像编辑

NEP: Autoregressive Image Editing via Next Editing Token Prediction

State Key Laboratory of General Artificial Intelligence, BIGAI(一般人工智能国家重点实验室,BIGAI);Peking University(北京大学)

Huimin Wu, Xiaojian Ma, Haozhe Zhao, Yanpeng Zhao, Qing Li

AI总结:

针对现有图像编辑方法重建全图导致的计算成本高和非编辑区域偏差问题,提出基于自回归的下一编辑token预测(NEP)方法,通过预训练任意顺序T2I模型实现零样本局部区域编辑,达到SOTA水平并支持测试时缩放。

摘要:

文本引导的图像编辑涉及基于语言指令修改源图像,通常只需更改小的局部区域。然而,现有方法生成整个目标图像,而不是选择性地仅重新生成预期的编辑区域。这导致(1)不必要的计算成本,以及(2)倾向于重建非编辑区域的偏差,这损害了预期编辑的质量。为了解决这些限制,我们提出将图像编辑表述为基于自回归图像生成的下一编辑 token 预测(NEP),其中仅重新生成需要编辑的区域,从而避免对非编辑区域进行意外修改。为了实现任意区域编辑,我们提出预训练一个任意顺序的自回归文本到图像(T2I)模型。一旦训练完成,它能够进行零样本图像编辑,并且可以轻松适配 NEP 进行图像编辑,在广泛使用的图像编辑基准上实现了新的最先进水平。此外,我们的模型通过以零样本方式迭代地优化其生成,自然地支持测试时缩放(TTS)。项目页面为:https://nep-bigai.github.io/

英文摘要:

Text-guided image editing involves modifying a source image based on a language instruction and, typically, requires changes to only small local regions. However, existing approaches generate the entire target image rather than selectively regenerate only the intended editing areas. This results in (1) unnecessary computational costs and (2) a bias toward reconstructing non-editing regions, which compromises the quality of the intended edits. To resolve these limitations, we propose to formulate image editing as Next Editing-token Prediction (NEP) based on autoregressive image generation, where only regions that need to be edited are regenerated, thus avoiding unintended modification to the non-editing areas. To enable any-region editing, we propose to pre-train an any-order autoregressive text-to-image (T2I) model. Once trained, it is capable of zero-shot image editing and can be easily adapted to NEP for image editing, which achieves a new state-of-the-art on widely used image editing benchmarks. Moreover, our model naturally supports test-time scaling (TTS) through iteratively refining its generation in a zero-shot manner. The project page is: https://nep-bigai.github.io/

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑