All terms
Foundations
Latent Space
A compressed internal vector space where a model represents the patterns in its data.
Definition
Latent space is a kind of internal map (a space with many dimensions, each input stored as a long list of numbers) in which a model encodes its inputs as compact representations. Similar inputs land near each other and dissimilar ones far apart, so distances and directions can carry meaning. Autoencoders, diffusion models, and language-model embeddings all operate in such spaces, which makes latent representations useful for semantic search and other comparisons.