JA EN

Latent Space

A compressed, lower-dimensional representation space where generative models encode the essential features of high-dimensional data such as images.

A latent space is an abstract, lower-dimensional representation space into which high-dimensional data (images, text, audio) is compressed by a learned encoder. A 512x512 RGB image contains roughly 786,000 dimensions, but an autoencoder's bottleneck may compress it into a few hundred dimensions capturing essential structure.

A key property is that semantically similar data points map to nearby locations. In a face image latent space, attributes like smile intensity or age correspond to specific directions, enabling manipulation through vector arithmetic.

Understanding latent spaces is central to image generation, style transfer, semantic editing, and anomaly detection. t-SNE and UMAP are commonly used to visualize latent spaces in 2D for model analysis.

Related Terms

Related Articles