The coordinate system in which latent codes live, where distance and direction correspond to semantic or structural variation in the data.
What It Is
Latent space organizes all latent codes for a model family: VAE priors, diffusion noise schedules viewed across timesteps, or shared multimodal alignment spaces. Geometry matters—nearby points should correspond to similar decoded outputs after the decoder or denoising steps.
Why It Matters
Understanding latent space separates representation learning from generation mechanics. It explains why interpolation works, why regularization targets KL terms, and why encoders and denoising generators are paired in diffusion and autoencoder stacks.
Simple Example
Two face latents averaged in latent space decode to a blended face after the decoder runs. Moving along one principal direction might change smile intensity if the space disentangles factors—quality depends on training, not on the word latent alone.
Common Confusions
Latent space is not the tokenizer vocabulary or embedding table—that is discrete input indexing. It is also not parameter space (weights); latents are activations or stochastic codes during inference. A single latent vector is a point in latent space, not the space itself.