From Evaluation to Evolution: Benchmarking Image Editing and Evolving Image Models

Computer Science > Computer Vision and Pattern Recognition arXiv:2610.41210 (cs) [Submitted on 30 Oct 2026] Title: From Evaluation to Evolution: Benchmarking Image Editing and Evolving Image Models.

计算机科学 > 计算机视觉与模式识别 arXiv:2610.41210 (cs) [提交于 2026 年 10 月 30 日] 标题：从评估到进化：图像编辑基准测试与进化图像模型。

Abstract: Real-world image editing requires coordinating multiple visual transformations with content preservation and reference constraints, yet aggregate scores can readily obscure unmet local requirements. Effective evaluation must therefore identify local failures and translate them into priorities for subsequent model improvement.

摘要：现实世界的图像编辑需要协调多种视觉变换，同时兼顾内容保留和参考约束，然而汇总评分往往会掩盖未满足的局部需求。因此，有效的评估必须识别局部失败，并将其转化为后续模型改进的优先级。

We introduce WeGenEditBench, comprising 2,740 evaluation instances and 148 task types across single-image, text-intensive, and multi-reference editing. Pre-generated checklists shared across models decompose each request into atomic requirements; item-level deductions and visual evidence are established first and then summarized into scores for edit execution, preservation, visual quality, and task-specific correctness.

我们引入了 WeGenEditBench，包含 2,740 个评估实例和 148 种任务类型，涵盖单图编辑、文本密集型编辑和多参考编辑。跨模型共享的预生成清单将每个请求分解为原子需求；首先建立项目级扣分和视觉证据，然后汇总为编辑执行、内容保留、视觉质量和任务特定正确性的评分。

To support the experience-acquisition component of recursive self-improvement (RSI), we develop WeGenEvoStudio, a unified multi-agent harness for RSI-oriented learning that integrates model-conditioned failure diagnosis, retrieval-grounded data construction, supervised adaptation, and re-evaluation.

为了支持递归自我改进（RSI）的经验获取组件，我们开发了 WeGenEvoStudio，这是一个面向 RSI 学习的统一多智能体框架，集成了模型条件下的故障诊断、基于检索的数据构建、监督适应和重新评估。

Among 15 evaluated configurations, GPT-Image-2 leads single-image editing with an Overall score of 4.076, while Seedream-5.0-Pro leads text-intensive editing at 4.291. Evolution experiments across four generation and editing benchmarks demonstrate substantial targeted gains: on their report-selected Top-5 single-image targets, FLUX.2 [dev] improves Overall from 1.083 to 3.517, while Qwen-Image-Edit-2511 gains 13.2%.

在 15 种评估配置中，GPT-Image-2 以 4.076 的总分在单图编辑中领先，而 Seedream-5.0-Pro 以 4.291 的分数在文本密集型编辑中领先。跨四个生成和编辑基准的进化实验证明了显著的针对性提升：在报告选定的前 5 个单图目标上，FLUX.2 [dev] 将总分从 1.083 提高到 3.517，而 Qwen-Image-Edit-2511 提升了 13.2%。

Together, our benchmark and harness provide a foundation for developing more reliable image models through interpretable evaluation and feedback-driven learning.

总之，我们的基准测试和框架通过可解释的评估和反馈驱动的学习，为开发更可靠的图像模型奠定了基础。