Overview
Latent space is the 'hidden' representation learned by models like autoencoders or GANs. It contains the essential 'essence' of the data, stripped of noise and redundant details.
Generative AI
In generative models, we can 'sample' points from the latent space to create new, realistic data. Moving through the latent space can smoothly transform one image into another (e.g., changing a person's expression or hair color).
Difference from Embedding Space
While often used interchangeably, 'latent space' usually refers to the compressed representation in generative models, while 'embedding space' often refers to the semantic mapping of discrete items like words.