EraseBench
收藏资源简介:
EraseBench是由谷歌研究团队创建的多维基准测试数据集,旨在全面评估文本到图像生成模型中的概念擦除技术。该数据集包含超过100个不同概念和1000多个定制提示,涵盖了视觉相似性、艺术风格、二项式关系和子集-超集关系等多个维度。数据集的创建过程包括从大型语言模型和结构化源(如ImageNet)中提取语义知识,并通过人工验证确保概念的质量和可行性。EraseBench的应用领域主要集中在评估和提升概念擦除技术的鲁棒性,解决在擦除特定概念时对相关或视觉相似概念产生的连锁影响问题,从而提高生成模型的可靠性和安全性。
EraseBench is a multi-dimensional benchmark dataset created by the Google Research team, aiming to comprehensively evaluate concept erasure techniques in text-to-image generation models. This dataset contains over 100 distinct concepts and more than 1,000 custom prompts, covering multiple dimensions such as visual similarity, artistic style, binomial relationships, and subset-superset relationships. The dataset creation process involves extracting semantic knowledge from large language models and structured sources such as ImageNet, and ensuring the quality and feasibility of concepts through manual validation. The main application areas of EraseBench focus on evaluating and improving the robustness of concept erasure techniques, addressing the cascading impact on related or visually similar concepts when erasing specific concepts, thereby enhancing the reliability and safety of generation models.

- 1EraseBench: Understanding The Ripple Effects of Concept Erasure Techniques谷歌研究, 麦吉尔大学, 莱斯大学, 谷歌Deepmind · 2025年



