Some recent models, such as BLIP, BLIP-2, and InstructBLIP approach 
VQA as a generative task.