Latent Space

Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary

The compressed internal coordinate system a model learns, where each input becomes a vector and distance between points reflects similarity in meaning.

What is Latent Space?

A latent space is the compressed internal coordinate system a model learns, in which each input becomes a vector and distance between points reflects similarity. “Latent” means hidden: nobody designs the dimensions – the model invents them during training.

The space is far smaller than the raw input. A face photo is hundreds of thousands of pixel values; its latent representation might be a few hundred numbers that still capture what the face looks like.

How Latent Space Works

An encoder network maps each input to a point in the space. The bottleneck is the point: with so few numbers available, the model must keep what distinguishes one input from another and discard the rest.

What survives compression is structure. In a model trained on faces, inputs arrange themselves by pose, lighting, and age, and directions through the space correspond to meaningful variation even though no single coordinate is human-readable.

Word embeddings are a latent space for words. Generative models run the idea in reverse: pick a point in latent space and decode it into a new image, sound, or sentence.

Example of Latent Space

A variational autoencoder is trained on face photographs, compressing each image to a vector of, say, 200 numbers. Take two photos of different people and walk through what the space allows.

First, encode photo A and photo B into their two vectors. Second, average them, giving a point halfway between the two faces. Third, decode that midpoint: out comes a plausible new face blending both people.

Averaging raw pixels instead would produce a ghostly double exposure. The blend works only because the latent space is organized by similarity, so the midpoint of two faces still sits in face territory.

Related AI terms: Word Embedding · Embedding Layer · Autoencoder · Variational Autoencoder · Dimensionality Reduction

Did you like the Latent Space gist?

Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.

Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About

Read the Governor's Letter

Stay ahead with Governor's Letter, the newsletter delivering expert insights, AI updates, and curated knowledge directly to your inbox.

By subscribing to the Governor's Letter, you consent to receive emails from AI Guv.
We respect your privacy - read our Privacy Policy to learn how we protect your information.

Browse All AI Terms A–Z

Every term in the dictionary, in alphabetical order. Jump to a letter or scroll the full list.

A

B

C

D

E

F

G

H

I

J

K

L

M

N

O

P

Q

R

S

T

U

V

W

X

Y

Z