New RISEBench++ benchmark: reasoning-informed visual editing still stumps top models
An arXiv paper introduces RISEBench++, a benchmark for reasoning-informed visual editing that spans six reasoning dimensions — temporal, causal, spatial, logical, counterfactual and hybrid — decomposed into 12 subcategories and 65 fine-grained task types, with 1,000 human-annotated test cases released in English and Chinese. The authors also present RISE-Agent, a training-free agentic framework, and evaluate 58 visual editing approaches (34 open-source models, 19 closed-source models and 5 agentic methods); even the strongest, GPT-Image-2.5 Sunburst, reaches only 56.6% accuracy.
It shifts evaluation from whether an edit simply looks good toward whether a model can actually reason through complex instructions, giving a measurable baseline for the next round of editing models.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.12343 (cs)
[Submitted on 8 Oct 2026]
Title:Reasoning-Informed Visual Editing
Authors:Xue Yang, Peiyuan Zhang, Yilun Zhu, Qihao Yang, Mingxin Liu, Xiangyu Zhao, Ziqian Fan, Zhaokai Wang, Yan Li, Yifan Yang, Xu Yang, Xiaosong Jia, Yue Zhou, Zhihang Zhong, Junchi Yan
View a PDF of the paper titled Reasoning-Informed Visual Editing, by Xue Yang and 14 other authors