Latent Space
A compressed mathematical representation where AI models encode input data as points in multi-dimensional space.
Definition
Latent space is the compressed, abstract representation neural networks learn to encode data. In this space, similar items cluster together—photos of cats occupy nearby points, semantically similar sentences group close together.
Models map inputs to latent space during encoding and back to outputs during decoding. The latent space captures meaningful features abstractly, enabling generation, interpolation, and similarity comparisons.
Why It Matters
Latent space is where AI "understanding" lives. Embeddings, the basis of semantic search and RAG systems, are latent space representations. Understanding this explains how AI finds similar content.
Manipulating latent space enables creative applications. Interpolating between two images in latent space creates smooth transitions. Adding latent directions can transform attributes (age, style, emotion).
Examples in Practice
Stable Diffusion operates in latent space—compressing images to smaller representations for efficient processing, then decoding back to full resolution images.
OpenAI's CLIP creates shared latent space for images and text. Searching "sunset over mountains" finds images whose latent representations are near that text's representation.