Zhen Huang

Contrastively trained vision-language models like CLIP, have made remarkable progress in learning joint image-text representations, but still face challenges in compositional understanding. They often exhibit a “bag-of-words” behavior—struggling to capture the object relations, attribute-object b...