image embeddings

Steerable ViT Features for Text-Guided Vision

A new vision representation lets text steer ViT features toward specific objects without giving up generic visual utility.