Latent Space

Latent space is the internal, abstract representation an AI model learns to encode its data. 'Latent' means hidden — these are the underlying features the model extracts that are not directly observable in the raw input. In this high-dimensional space, each input (a sentence, image, or sound) becomes a point or vector, and the geometry of the space carries meaning: similar concepts cluster together, and directions can correspond to interpretable attributes.

A classic illustration comes from word embeddings, where vector arithmetic such as king − man + woman ≈ queen works because the latent space has organized concepts along meaningful axes. In generative models, latent space is where creation happens: diffusion image models like Stable Diffusion denoise within a compressed latent space rather than raw pixels, and interpolating between two points in latent space produces a smooth blend between two outputs. Latent space is the bridge between raw data and the structured understanding that makes embeddings, generation, and similarity search possible.