Multimodality

Multimodality is a model’s ability to process or generate several types of data at once – text, image, audio, video or structured signals. The technical basis is a shared representation space: each modality has its own encoder that turns the input into a sequence of tokens or embeddings projected into the same latent space, where the transformer then handles them uniformly. An image is typically cut into a grid of patches, which a linear projection turns into visual tokens; audio is converted into a spectrogram and that into windows. The links between modalities are learned from paired data – images with captions, videos with subtitles – and contrastive training plays the key role, pulling the representation of a picture and of its description towards each other, as the CLIP model established. Cross-attention is used to feed one modality into another. The practical consequence is that boundaries between tasks dissolve: the same model answers questions about a photograph, reads tables off a screenshot or operates a user interface. In security terms it brings new attack surfaces, because instructions can be hidden inside an image.


People are multimodal as a matter of course. Asked “what is wrong with this picture?”, you effortlessly connect what you see with what you hear and answer in words. You do not keep a separate department for images and another for words in your head; you have one understanding of the world that both feed into. Models lacked that privilege for a long time – one system handled text, another handled pictures, and between them there was silence. Multimodality means a photograph is cut into pieces that behave inside the model like words, and from there seeing and reading are one and the same substance.

Is this article useful to you and are you citing it? Copy the citation