One space, many kinds of input
The technique that makes multimodal systems work is training separate encoders, one per modality, so that their outputs land in the same shared vector space. A text encoder turns a sentence into a vector. An image encoder turns a picture into a vector of the same dimensionality. Crucially, these encoders are trained together, with a shared objective, so that related concepts end up near each other in that space regardless of which modality they came from.
That means the embedding for the word 'dog', a photo of a dog, and, in systems that include audio, the sound of a bark, can all end up positioned close together in the same space. None of these inputs look alike at the raw data level, a string of characters, a grid of pixels, a waveform, but after encoding, they can sit near each other as points in the same geometric space, which is what makes it possible to compare them at all.