QFormer

The paper discusses the challenges of extending large-scale pretraining and instruction tuning to vision-language models due to the diversity in visual inputs. The authors propose a more efficient method for…