Consistency-Preserving Concept Erasure via Unsafe-Safe Pairing and Directional Fisher-weighted Adaptation

One-line Summary

The paper introduces PAIR, a framework for concept erasure in text-to-image models that maintains semantic consistency by aligning unsafe concepts with safe alternatives.

Plain-language Overview

This research addresses the challenge of removing undesirable content from text-to-image models while keeping the rest of the image consistent. The proposed method, called PAIR, pairs unsafe content with safe alternatives to guide the erasure process. By doing so, it ensures that the visual and semantic quality of the image remains intact even after removing the unwanted content. The new approach outperforms existing methods by better preserving the structure and meaning of the original image.

Technical Details