Concept erasure techniques have recently gained significant attention for their potential to remove unwanted concepts from text-to-image models. While these methods often demonstrate promising results in controlled settings, their robustness in real-world applications and suitability for deployment remain uncertain. In this work, we (1) identify a critical gap in evaluating sanitized models, particularly in assessing their performance across diverse concept dimensions, and (2) systematically analyze the failure modes of text-to-image models post-erasure. We focus on the unintended consequences of concept removal on non-target concepts across different levels of interconnected relationships including visually similar, binomial, and semantically related concepts. To address this, we introduce EraseBench, a comprehensive benchmark for evaluating post-erasure performance. EraseBench includes over 100 curated concepts, targeted evaluation prompts, and a robust set of metrics to assess both effectiveness and side effects of erasure. Our findings reveal a phenomenon of concept entanglement, where erasure leads to unintended suppression of non-target concepts, causing spillover degradation that manifests as distortions and a decline in generation quality.
You Only Erase Once: Erasing Anything without Bringing Unexpected Content
School of Computer Science and Engineering, Sun Yat-sen University, China(中山大学计算机科学与工程学院);Amazon(亚马逊);Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China(教育部机器智能与先进计算重点实验室)
RealEra: Semantic-level Concept Erasure via Neighbor-Concept Mining
Institute of Information Engineering, Chinese Academy of Sciences(中国科学院信息工程研究所);Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)(广东省人工智能与数字经济实验室(深圳));School of Cyber Security, University of Chinese Academy of Sciences(中国科学院大学网络空间安全学院)
SAGE: Exploring the Boundaries of Unsafe Concept Domain with Semantic-Augment Erasing
Faculty of Data Science, City University of Macau(澳门城市大学数据科学学院);Institute of Information Science, Beijing Jiaotong University(北京交通大学信息科学学院);Beijing Key Laboratory of Advanced Information Science and Network Technology(北京先进信息科学与网络技术重点实验室)
Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models
Michigan State University(密歇根州立大学);Intel(英特尔公司);University of Minnesota, Twin City(明尼苏达大学双城分校);MIT-IBM Watson AI Lab(麻省理工学院-IBM沃森人工智能实验室);IBM Research(IBM研究院)
Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models
Fudan University(复旦大学);Shanghai Key Lab of Intell. Info. Processing(上海智能信息处理重点实验室);School of Computer Science(计算机学院);Shanghai Collaborative Innovation Center on Intelligent Visual Computing(上海智能视觉计算协同创新中心)
DuMo: Dual Encoder Modulation Network for Precise Concept Erasure
School of Computer Science, Fudan University(复旦大学计算机科学技术学院);Shanghai Key Lab of Intell. Info. Processing(上海智能信息处理重点实验室);Shanghai Collaborative Innovation Center on Intelligent Visual Computing(上海智能视觉计算协同创新中心)
Semantic Steering for Controllable Generation: Tuning-Free Concept Erasure in Multimodal Diffusion Transformers
Institute of Information Engineering, Chinese Academy of Sciences(中国科学院信息工程研究所);School of Cyber Security, University of Chinese Academy of Sciences(中国科学院大学网络空间安全学院)
VCE: Safe Autoregressive Image Generation via Visual Contrast Exploitation
Fudan University(复旦大学);Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University(复旦大学计算机科学技术学院上海市智能信息处理重点实验室);Shanghai Collaborative Innovation Center on Intelligent Visual Computing(上海智能视觉计算协同创新中心);International Computer Science Institute(国际计算机科学研究所);UC Berkeley(加州大学伯克利分校)
Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters
University of Science and Technology of China(中国科学技术大学);The University of Manchester(曼彻斯特大学);Hefei University of Technology(合肥工业大学);Institute of Software Chinese Academy of Sciences(中国科学院软件研究所)
Uni-AdaVD: Universal Concept Erasure for Visual Generation via Orthogonal Value Decomposition
School of Computer Science and Information Engineering, Hefei University of Technology(合肥工业大学计算机与信息工程学院);School of Cyber Science and Technology, University of Science and Technology of China(中国科学技术大学网络空间安全学院);School of Artificial Intelligence and Data Science, University of Science and Technology of China(中国科学技术大学人工智能与数据科学学院);Institute of Software, Chinese Academy of Sciences(中国科学院软件研究所)