无辜画面，仇恨故事：多轮视觉故事生成中的仇恨意图评估与检测
原标题：Innocent Panels, Hateful Stories: Evaluating and Detecting Hateful Intent in Multi-Turn Visual Story Generation
AI 摘要
该研究指出，前沿文生图系统（如Gemini和GPT-Image）支持多轮对话生成一致的角色和场景，使得制作仇恨性视觉故事（有序图像组）变得廉价且可扩展。作者引入了HatefulStoryPrompts数据集，包含55个仇恨故事、330个多轮配置，并评估了五个前沿模型，发现所有模型都能完成超过80%的故事，最强模型达到99.0%。现有审核系统在检测组级仇恨含义时表现不佳，专用安全模型召回率最高仅34.9%，强视觉语言模型为67.5%。作者提出了交互感知监控和生成后防御方法，分别达到97.3%和92.6%的召回率，以及80.2%的生成后检测召回率。
正文节选
Computer Science > Computer Vision and Pattern Recognition Title:Innocent Panels, Hateful Stories: Evaluating and Detecting Hateful Intent in Multi-Turn Visual Story Generation View PDF HTML (experimental) Abstract:Picture books and comics have long been used to disseminate hateful narratives because they are easily understood even by children, as exemplified by the notorious Nazi propaganda picture book \emph{Der Giftpilz}. Recently, frontier text-to-image (T2I) systems such as Gemi