Concept erasure aims to strip harmful, private, or copyrighted concepts out of text-to-image models without damaging their other capabilities. A new arXiv preprint, BEM: Learning Bregman Divergences for Concept Erasure in Flow Matching Models, addresses this problem for the current generation of leading models.
According to the abstract, the paper targets flow-matching models, specifically citing FLUX and Stable Diffusion 3. The proposed approach, BEM, learns Bregman divergences to perform erasure, though the abstract provided is truncated and does not include experimental results or implementation details.
Because the source is a single preprint abstract, no independent comparisons or corroborating findings are available. Readers should treat the claims as preliminary.