SACE: Concept Erasure at the Semantic Singularity in Visual Autoregressive Models
Researchers have introduced SACE, a novel concept erasure framework designed for visual autoregressive (VAR) models. This framework addresses safety concerns in text-to-image generation by proposing the Semantic Singularity Axiom, which identifies the precise scale at which semantic concepts are embedded. SACE utilizes Incremental Semantic Saliency Analysis to validate this axiom and employs an Entropy-Regularized Erasure Objective coupled with a preservation loss to ensure concept erasure without compromising model integrity or introducing visual artifacts. Experiments show SACE effectively removes targeted concepts with minimal training and preserves essential benign priors. AI
IMPACT Enhances safety and control in text-to-image generation models by enabling precise concept erasure.