It’s sort of an advanced Mr/Mrs Potato Head. The program has a bunch of face pieces collected from real photos. It does its best to match combinations of those pieces to the sketch.
The give away is the sketch area. It looks like those were the training sketches/tracings. One was done for each picture. The closer a user sketch line is to the training sketch line, the greater the chance of using a piece of the associated photo.
(all speculation. I’m bored enough to post, but not read the paper).
