Concept
Latent Space
Learned high-dimensional space where each point encodes the abstract features of an input.
Definition
Latent space is the compressed internal representation a model uses to reason. Distances in this space correspond to semantic similarity; interpolation between points often produces meaningful intermediate outputs. Diffusion image models do most of their work in a compressed latent rather than raw pixels.
Common use cases
- Interpolation
- Style mixing
- Anomaly detection