Xavi says: This paper is interesting in many ways. By using multimodal fine-tuning of small(er) (i.e. 1B parameter) models like T5 the authors show that they can beat GPT3.5 in visual QA tasks, surpassing visual performance in several tasks.