Exploring the Capabilities of Large Multimodal Models on Dense Text