This is done with two models in most standard biencoder approaches. This is how multimodal embedding search works. We want to train a model such that the location of the text embeddings that represent an item and the image embeddings for that item are colocated.