CompCap: Improving Multimodal Large Language Models with Composite Captions