Multimodal models first convert each input into numerical embeddings: tokens for text, visual patches or image regions for images, and time–frequency features for audio. Special encoders preserve each modality’s useful structure, while projection layers map their embeddings into a shared representation space so related concepts—such as a spoken word, its written form, and a matching image—can align.
A transformer then fuses these representations. In early fusion, text, image, and audio tokens are combined before most processing; in late fusion, separate encoders process them and a later module combines their summaries. Cross-attention is a common middle ground: tokens from one modality attend to another, allowing the model to connect captions with objects, or speech with images. Training on paired or synchronized data teaches the shared space these correspondences. Would you also like to know how cross-attention aligns text, image, and audio tokens?
Get more accurate answers with Super Pandi, upload files, personalized discovery feed, save searches and contribute to the PandiPedia.
Let's look at alternatives: