Training-Free Semantic Correction for Autoregressive Visual Models

The paper introduces Gazer, a training-free framework for enhancing autoregressive visual models (AVMs) by integrating multimodal large language model feedback into the generation process. Gazer operates in two stages: Reflective Diagnosis, which identifies semantic errors from intermediate states, and Semantic Correction, which adjusts the generation trajectory to better align with the target prompt. Experimental results show that Gazer significantly improves semantic alignment and compositional accuracy in image and video synthesis tasks without the need for additional training, providing a valuable tool for practitioners looking to enhance AVM output quality.