SafeDiffusion-R1: Online Reward Steering for Safe Diffusion Post-Training

Researchers introduced SafeDiffusion-R1, a post-training framework that uses online reinforcement learning to make diffusion models safer.

It combines Group Relative Policy Optimization with a CLIP-based steering reward to reduce unsafe generations from both positive and negative prompts.

The method reportedly generalizes across multiple harm categories while maintaining or improving image quality on models like SD v1.4.

Because it avoids paired supervision and reward-model tuning, it offers a more scalable and lower-cost path to diffusion safety.