From Face to Numbers

The embedding stage runs the aligned face through a model that outputs a fixed-length vector of numbers, for example a few hundred values. This vector is the embedding. Its useful property is geometric: embeddings of the same person sit close together in that numeric space, while embeddings of different people sit far apart. Because of this, the system can compare faces by measuring distances between vectors instead of comparing raw images, and it never needs to store the original photographs to do it.