This paper points out the limitations of the safety mechanism of text-image diffusion models, which fail to account for individual user preferences, and proposes a Personalized Safety Alignment (PSA) framework. PSA integrates user profiles into the diffusion process to adapt the model's behavior to individual safety criteria while maintaining image quality. It incorporates user-specific safety preferences using a novel dataset, Sage, and integrates the profiles through a cross-attention mechanism. Experimental results demonstrate that PSA outperforms existing methods in suppressing harmful content, generates content that better aligns with user constraints, and achieves higher Win Rate and Pass Rate scores. The code, data, and models are publicly available.
