CSU-JPG/IESBench
收藏资源简介:
IESBench是一个视觉中心的基准测试数据集,用于评估大型图像编辑模型的安全性。该数据集包含1054个视觉提示图像,这些图像通过视觉输入(如标记、箭头和视觉-文本提示)来推断用户意图,以模拟潜在的攻击场景。数据集覆盖了15个安全策略、116个属性和9个动作,旨在全面测试模型在面临恶意视觉编辑时的安全性和鲁棒性。每个数据点包括问题(意图描述)、图像路径、属性(可编辑目标)、动作(编辑操作)、类别(安全策略)、重写文本提示(LLM生成的文本提示)和图像ID等字段,以支持标准化评估和研究。
IESBench is a vision-centric benchmark dataset for evaluating the safety of large image editing models. It contains 1,054 visually-prompted images that infer user intent through visual inputs such as marks, arrows, and visual-text prompts, simulating potential attack scenarios. The dataset spans 15 safety policies, 116 attributes, and 9 actions, designed to comprehensively test the security and robustness of models against malicious visual editing. Each data point includes fields such as question (intent description), image-path, attributes (editable targets), action (edit operations), category (safety policies), rewrite (LLM-generated text prompt), and image_id, facilitating standardized evaluation and research.




