BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Mode