MiniGPT-v2:Large Language Model as a Unified Interface for Vision-Language Multi-task Learning