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

TL;DR

This paper introduces SafeDiffusion-R1, an online reinforcement learning method that improves diffusion model safety by steering content away from unsafe prompts without requiring supervised unsafe data.

Contribution

It proposes a novel online reward steering mechanism using CLIP embeddings and Group Relative Policy Optimization to enhance safety without catastrophic forgetting.

Findings

Reduces inappropriate content to 18.07% from 48.9% in baseline

Decreases nudity detections to 15 from 646 baseline

Improves generative quality from 42.08% to 47.83% on GenEval

Abstract

Diffusion models have been widely studied for removing unsafe content learned during pre-training. Existing methods require expensive supervised data, either unsafe-text paired with safe-image groundtruth or negative/positive image pairs, making them impractical to scale. Furthermore, offline reinforcement learning and supervised fine-tuning approaches that generate synthetic data offline suffer from catastrophic forgetting, degrading generation quality. We propose a novel online reinforcement learning framework that addresses both data scarcity and model degradation through post-training with Group Relative Policy Optimization (GRPO) on both negative and positive text prompts. To eliminate the need for fine-tuning specialized safe/unsafe reward models, we introduce a \textit{steering reward mechanism} that exploits an inherent property of CLIP embeddings: steering text representations toward positive safety directions and away from negative ones in the embedding space. Our online-policy approach enables the model to learn from diverse prompts, including explicit unsafe content, without catastrophic forgetting. Extensive experiments demonstrate that our method reduces inappropriate content to 18.07\% (vs. 48.9\% for SD v1.4) and nudity detections to 15 (vs. 646 baseline) while improving compositional generation quality from 42.08\% to 47.83\% on GenEval. Remarkably, these safety gains generalize to out-of-domain unsafe prompts across seven harm categories, achieving state-of-the-art performance without supervised paired data or reward tuning. Github: https://github.com/MAXNORM8650/SafeDiffusion-R1.