The field of image safety and text-to-image models is rapidly evolving, with a growing focus on developing more effective and efficient methods for ensuring the safety of generated images. Recent research has highlighted the importance of fine-grained image safety distinctions, with a particular emphasis on identifying and mitigating subtle changes to images that can drastically alter their safety implications.

Noteworthy papers in this area include SafetyPairs, which introduces a scalable framework for generating counterfactual pairs of images that differ only in the features relevant to a given safety policy. T2I-RiskyPrompt is another significant contribution, providing a comprehensive benchmark for evaluating safety-related tasks in text-to-image models. SafeEditor proposes a unified MLLM for efficient post-hoc safety editing, enabling efficient safety alignment for any text-to-image model. Stop the Nonconsensual Use of Nude Images in Research raises important ethical concerns regarding the use of nonconsensually collected nude images in research, highlighting the need for more responsible and respectful practices in the field.