Text-to-image diffusion models have moved quickly from research prototypes to widely deployed tools. Along with their capabilities, they can also produce harmful or undesirable content, including NSFW material. This paper addresses that problem by rethinking negative prompting, a common technique used to steer models away from certain concepts.
The authors propose a method called unconditional anchoring and localized guidance. Rather than relying solely on negative prompts, the approach anchors generation to the unconditional model output and applies guidance in a localized way. The paper presents this as a more effective strategy for suppressing unwanted concepts.
The work is described in an arXiv preprint. The abstract does not include full experimental details, but the proposed technique points toward safer deployment of diffusion models by improving how they handle content suppression.