SafeDiffusion-R1: Online Reward-Guided Safety Post-Training for Diffusion Models

SafeDiffusion-R1 proposes an online reinforcement learning post-training framework based on GRPO, reducing unsafe content generated by diffusion models through a reward guidance mechanism in the CLIP embedding space. The method requires no supervised paired data or reward model fine-tuning, reducing the proportion of inappropriate content from 48.9% for SD v1.4 to 18.07%, and the nudity detection count from 646 to 15. GenEval compositional generation quality rises from 42.08% to 47.83%, and the safety gains generalize to out-of-domain unsafe prompts across seven harm categories.