To Erase, or Not to Erase: Robust Training-Free Concept Erasure with Preservation aware Adaptive Ranked Subspace Expansion

arXiv published the paper “To Erase, or Not to Erase”, proposing a training-free erasure framework called PARSE. This framework aims to address the challenge of balancing erase robustness with model efficiency in existing technologies. Its core mechanisms include: dynamically identifying target-induced erasure and preservation concepts; using perceptually preserved projections to edit the cross-attention value space for subspace operations; and adaptively expanding the erasure subspace when there is no conflict to handle triggers. Additionally, the study introduced the Balance Erasure Utility Score (BEUS), which balances attack success rate with image generation fidelity. Experiments show that PARSE achieves multi-concept robust erasure in NSFW, artistic style, and object erasure tasks without sacrificing the quality of the edited images.