Brief

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

Researchers have developed a new post-training quantization (PTQ) framework specifically designed for visual autoregressive models (VAR). This framework addresses two main challenges: high reconstruction errors in attention-value products and a mismatch between calibration data sampling frequencies and predicted probabilities. The proposed solution includes a shift-and-sum quantization method and a resampling strategy for calibration data, which collectively improve performance in various image generation tasks. AI