Latent space refers to a compressed, often lower-dimensional representation of input data learned by a model—typically an autoencoder, variational autoencoder (VAE), or other deep learning model. Each point in this space corresponds to a set of features that capture high-level or abstract characteristics of the original input.

In practice, the latent space acts as an intermediate representation that makes it easier to perform tasks like clustering, interpolation, generation, or anomaly detection. For example, in a VAE, the latent space is structured to allow smooth sampling and interpolation between data points, enabling controllable generation of new samples.

Latent spaces are crucial in representation learning, where the goal is to discover informative features automatically. In vision, latent vectors may capture object type or pose; in NLP, they might represent semantic meaning or syntax.

A well-structured latent space enables models to generalize better, compress information efficiently, and support creative or exploratory applications like style transfer, semantic search, and synthetic data generation.