LION: Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge | Read Paper on Bytez