Reproduce “RASteer: Retain-Aware Activation Steering for Concept Erasure in Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yongliang Wu, Haori Lu, Yulun Wu, Jinqi Luo, Xingyu Zhu, Yaoyao Liu
RASteer edits text-to-image diffusion steering directions to reduce damage to retained concepts while erasing a target. It uses retain concepts to identify shared activation features, then adaptively reduces their contribution during generation. On SD-v1.4, NudeNet detected 13 exposed body parts on I2P prompts, versus 24 with CASteer. These results concern tested Stable Diffusion models and predefined retain sets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.