Safe and Reliable Diffusion Models via Subspace Projection

Summary

SAFER is a method for removing target concepts from text-to-image diffusion models so they do not reappear in subtle forms. It identifies a concept-specific subspace in the text embedding space and projects prompt embeddings onto its complementary subspace, using textual inversion from a reference image and a subspace expansion strategy. The authors report that it consistently erases unwanted concepts while preserving generation quality.