Kunal Jain

CLIP models suffer from a surprisingly strong "center bias," causing them to miss important objects outside the image's central region, even when those objects are crucial for accurate vision-language understanding.