Latent Space
The compressed, abstract representation space a model maps data into — where similar things sit close together.
A latent space is the lower-dimensional, learned representation space into which a model encodes data — each input becomes a point (vector) whose position captures its meaningful features. Similar inputs land near each other, and directions in the space often correspond to interpretable factors.
Worked example: an autoencoder or embedding model maps faces into a latent space where moving along one direction adds a smile and another adds glasses; a diffusion model generates by denoising in a VAE’s latent space rather than raw pixels, which is far cheaper. Gotcha: a latent space is only as good as what shaped it — a plain autoencoder’s latent space is not smooth enough to sample from (garbage between points), which is why VAEs enforce structure; and latent directions being interpretable is a useful tendency, not a guarantee.