Vision-Language Models

We’ve flipped the script on multimodal fusion, instead of forcing visual features into discrete text space, we map text embeddings into continuous visual space, eliminating costly alignment pre-training while achieving competitive performance.