SACE: Concept Erasure at the Semantic Singularity in Visual Autoregressive Models

The article introduces the Semantic Singularity Axiom and the SACE framework, which addresses the challenges of concept erasure in visual autoregressive (VAR) models. SACE employs an Entropy-Regularized Erasure Objective and a restorative preservation loss to achieve effective concept erasure while maintaining model integrity, demonstrating improved performance over existing techniques with minimal training overhead. This advancement is significant for practitioners as it enhances safety alignment in text-to-image synthesis, allowing for more reliable manipulation of generated content without compromising visual quality.