Modeling Caption Diversity in Contrastive Vision-Language Pretraining

arXiv:2405.00740v1 Announce Type: cross
Abstract: There are a thousand ways to caption an image. Contrastive Language Pretraining (CLIP) on the other hand, works by mapping an image and its caption to a single vector -- limiting how well CLIP-like models can represent the diverse ways to describe an image. In this work, we introduce Llip, Latent Language Image Pretraining, which models the diversity of captions that could match an image. Llip's vision encoder outputs a set of visual features that …

abstract arxiv clip cs.ai cs.cl cs.cv cs.lg diverse diversity image language mapping modeling pretraining type vector vision vision-language work