The field of text-to-image generation is moving towards developing more responsible and controlled generative models. Recent research has focused on improving the safety and reliability of these models by removing unwanted concepts and reducing reliance on spurious correlations. Innovative approaches have been proposed to achieve fine-grained erasure of target concepts while preserving related concepts, and to generate training samples that reduce the model's reliance on spurious correlations. Additionally, new techniques have been developed to extract intrinsic concepts from single images and to quantify the fundamental limits of perfect concept erasure. Noteworthy papers include:

• SAFER, which proposes a novel approach for thoroughly removing target concepts from diffusion models,
• FADE, which introduces adjacency aware unlearning in diffusion models to preserve related concepts,
• ICE, which presents a framework for automatically extracting intrinsic concepts from a single image, and
• Fundamental Limits of Perfect Concept Erasure, which investigates the theoretical bounds of concept erasure and proposes an approach to achieve optimal erasure.