MORAIDICTIONNAIRE IA
Génératif

Multimodal AI

AI models that understand and combine text, images, sound and video within one shared space of meaning.

Multimodal AI refers to artificial intelligence systems able to perceive, understand and generate several modalities at once — text, images, audio, video, even physiological signals. Where a classic model only "reads" text, a multimodal model behaves like a human endowed with several senses: it looks at a photo, reads a caption and listens to a comment to build a single, unified understanding.

The core idea: a shared representation space

The heart of multimodal AI is projecting heterogeneous data into one shared vector space (an embedding space). A word, a pixel and a sound are each turned into comparable vectors. The model then learns to bring together representations that describe the same reality — for instance the word "cat", a photo of a cat and the sound of a meow.

A dominant approach is contrastive learning, which maximizes the similarity between an image and its matching text while pushing apart mismatched pairs:

$$\mathcal{L} = -\log \frac{\exp(\text{sim}(I, T)/\tau)}{\sum_{k} \exp(\text{sim}(I, T_k)/\tau)}$$

where $\text{sim}$ is a cosine similarity and $\tau$ a temperature.

Modalities and uses

Input Output Example use
Image + text Text Describing an X-ray
Text Image Illustrative generation
Audio + text Text Context-aware voice assistant
Video Text Automatic summarization

Why it matters

Meaning often emerges from the crossing of sources: a chart without a caption is ambiguous, and so is a sentence without an image. By fusing signals, these models reduce errors and unlock applications in healthcare, education and accessibility.

Multimodality moves AI closer to how we actually perceive the world: through several senses at once.

Explore the full AI dictionary →