JANUS: A lightweight framework for jailbreaking text-to-image models via distribution optimization

The researchers propose JANUS, a text-to-image jailbreak framework that formulates jailbreaking as optimizing a structured prompt distribution under a black-box end-to-end reward. It replaces a high-capacity generator with a low-dimensional mixture strategy over two semantically anchored prompt distributions, reducing exploration cost while preserving the target semantics. On Stable Diffusion 3.5 Large Turbo, its ASR-8 increases from 25.30% to 43.15%, and CLIP and NSFW scores are also higher. The method is effective on both open-source and commercial text-to-image models, and the authors state that it exposes structural weaknesses in existing T2I safety pipelines.