Reproduce “Example-Based Spatial Guidance for Training-Free 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
Younghwan Kil, Giung Nam, Joonhyeong Park, Jinwoo Shin, Juho Lee
EBSG provides training-free inference-time control for harmful content in diffusion models. It detects emerging harmful sub-concepts, localizes their regions, and redirects them toward paired safe alternatives. On seven I2P concept slices, it averages 73.6% Safety Rate in separate runs and 73.1% when erasing all seven simultaneously. This rate includes partial suppression and does not establish permanent erasure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.