HERO: Rethinking Visual Token Early Dropping in High-Resolution Large Vision-Language Models

VLMRe-assign community