Latent space

Latent space is the abstract space — usually of far fewer dimensions than the raw input — in which a model represents the compressed, encoded “essence” of its input data. Instead of working with raw pixels or tokens, the model maps each input to a point (a vector) in this space, where each dimension captures some learned feature rather than a human-labeled one — hence latent, meaning hidden. The useful property is that geometry becomes meaning: inputs that are semantically similar land near each other, and directions in the space can correspond to interpretable changes (the classic “king − man + woman ≈ queen” with word embeddings). Autoencoders, diffusion models, and embedding models all rely on it; generative models (a GAN, say) work by sampling or steering points in latent space and then decoding them back into images, text, or audio.