EditAgent expands the action and observation spaces of image-editing agents
The arXiv paper EditAgent routes each editing step to general-purpose editors or specialized methods, then attaches a verification record to every candidate. It generates candidates in parallel and evaluates them in one listwise call; the authors report that experts preferred this approach over sequential refinement in 78.5% of decided comparisons, using no more than half as many judge calls.
The paper offers a concrete method for improving agentic image editing by broadening available tools and giving judges richer information, with reported gains on two open-weight editors across three benchmarks.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.39664 (cs)
[Submitted on 29 Oct 2026]
Title:EditAgent: Rethinking Agentic Image Editing
Authors:Haiyang Xu, Tianyu Wang, Jialu Li, Jiteng Mu, Yuanjun Xiong, Jianming Zhang, Zhuowen Tu, Mingze Xu
View a PDF of the paper titled EditAgent: Rethinking Agentic Image Editing, by Haiyang Xu and 7 other authors