Multimodal Large Language Models (MLLMs) require highly representative visual inputs for efficient video understanding. Existing keyframe selection methods, however, often struggle to simultaneously ...