Latent space

The hidden vector space where geometry encodes meaning.

A dreamy aerial photograph of a misty valley at dawn, softly glowing clustered forms floating in fog with similar shapes gathered together - a map of meaning.

Latent space is the abstract, usually much smaller space in which a model holds the compressed “essence” of its input. Rather than work with raw pixels or words, the model maps each input to a point (a vector) in this space, where each axis captures some feature the model learned on its own rather than one a human labeled, hence latent, meaning hidden. The useful part is that geometry turns into meaning: similar inputs land near each other, and moving in a particular direction can correspond to a meaningful change (the classic “king − man + woman ≈ queen” with word embeddings). Many kinds of models rely on it: autoencoders, diffusion models, embedding models, and generative models (a GAN , say) work by picking or nudging a point in latent space and then translating it back out into an image, text, or audio.