1 paper
Huilin Zhou, Ruoxi Cheng, Yuhang Wang +4
Diffusion models achieve strong performance in text-to-image generation, but they can still produce unsafe content. Training-free safety guidance offers a practical inference-time…