ICML2025

Minimalist Concept Erasure in Generative Models

Yang Zhang, Er Jin, Yanfei Dong, Yixuan Wu, Philip Torr, Ashkan Khakzar, Johannes Stegmaier, Kenji Kawaguchi

Abstract

Figure 1 : Minimalist concept erasure results on FLUX, the latest rectified flow model with 12 billion parameters. We propose minimalist concept erasure, an approach that applies just enough changes to unwanted concepts, so they become unrecognizable. We can effectively remove inappropriate content like NSFW, weapons, and tackle copyright issues by removing protected IPs and art styles while maintaining the model performance.