Look at the photo grid about halfway down this article <https://www.theregister.co.uk/2018/12/14/ai_created_photos/>: as I understand it, the main facial features are taken from the photos down the leftmost column; these are then resynthesized into a new orientation, expression and other embellishments (including hair colour/style, gender and apparent age!) to match the photos across the topmost row. Or something like that. Apparently, Nvidia is planning to publish its source code. Kids, you may not be able to afford the hardware to try this at home ...