tencent/DiffSpot
收藏资源简介:
DiffSpot是一个用于评估视觉语言模型(VLMs)在网页界面中检测细微视觉差异能力的数据集。它专注于“找不同”任务,通过生成4,400对截图来测试模型,其中3,900对包含基于13种CSS属性操作符和3个难度等级(简单、中等、困难)的单个视觉变化,另外500对作为无差异控制样本用于测量幻觉。数据集采用代码驱动方法生成:通过突变自包含HTML中目标元素的CSS属性,重新渲染截图,并记录变化属性、元素和幅度作为真实标签。接地门机制确保像素差异仅局限于目标元素,从而提供精确的屏幕变化描述。任务要求模型在给定前后截图的情况下,以开放形式描述变化,并使用LLM作为评判进行评分。评估结果显示,即使最佳模型也只能识别40.7%的真实变化,困难层级的召回率低于23%,表明当前VLMs在细粒度视觉感知方面仍面临挑战。数据集支持英语、日语和中文,适用于图像到文本和视觉问答任务。
DiffSpot is a dataset designed to evaluate the ability of vision-language models (VLMs) to spot fine-grained visual differences in web interfaces. It focuses on a spot-the-difference task, comprising 4,400 screenshot pairs: 3,900 with a single visual change balanced across 13 CSS-property operators and three difficulty tiers (easy, medium, hard), plus 500 no-diff controls for hallucination measurement. The dataset is code-driven: it mutates one CSS property of a target element in self-contained HTML, re-renders screenshots, and records the changed property, element, and magnitude as ground truth. A grounding gate ensures that rendered pixel differences are confined to the target element, providing an exact label of what changed on screen. The task requires models to describe the change in natural language given both screenshots, scored by an LLM-as-Judge against structured mutation logs. Evaluation shows that even the best model identifies only 40.7% of true changes, with hard-tier recall below 23% for all models, highlighting the challenge for VLMs in fine-grained visual perception. The dataset supports English, Japanese, and Chinese languages, and is categorized under image-to-text and visual-question-answering tasks.




