BEM: Learning Bregman Divergences for Concept Erasure in Flow Matching Models

Key points

• Concept erasure removes harmful, private, or copyrighted concepts from text-to-image models while preserving their other abilities.1
• Leading models such as FLUX and Stable Diffusion 3 are trained with flow matching, and erasure methods for them fine-tune the model toward a guided teacher's velocities under a Euclidean loss.1
• Yet the squared distance is only one member of a larger family: Bregman divergences are exactly the losses under which flow matching recovers the correct velocity field, and each member shapes differently what is erased and what is preserved.1
• In the experiments we target the suppression of nudity, violence, real identities, and copyrighted characters.1

Sources (1)

• [1]BEM: Learning Bregman Divergences for Concept Erasure in Flow Matching Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 07:07 AM

  We introduce BEM (Bregman erasure matching), which is, to our knowledge, the first to learn this divergence through bi-level optimization, so that the fine-tuned model forgets the concept on unseen prompts while its unrelated generations stay unchanged.

  Concept erasure removes harmful, private, or copyrighted concepts from text-to-image models while preserving their other abilities.

Extractive summary: sentences quoted from the sources.