Latent Space

A 512×512 face photo is roughly 800,000 pixel values — but the space of plausible faces is far smaller. Faces vary along a limited set of axes: pose, lighting, age, expression, hairstyle. A photo is high-dimensional; 'a face' is not.

A latent space is a learned, compact coordinate system for that hidden structure: each image becomes a small vector — say 128 numbers instead of 800,000 — and a decoder turns the vector back into pixels. It's compression, but it compresses meaning, not just bytes.