Well, we trained neural nets with thousands and thousands of hi-resolution laser scans of real people from a wide variety of ages and ethnic populations. Each person also had dozens of photos associated with them, taken from a variety of cameras, angles, and lighting conditions. We then created a neural net pipeline which works in reverse, given a random photo of anybody, we can backwards infer a high resolution mesh components from candidates in our original laser scan database. Given a series of candidate face parts represented as a probability cloud, the 'nets collapse the probability cloud into whatever target mesh and format requested. Finally, the candidate's photo is backwards distorted to fit the generated geometry. The amazing part is how fast it works; algorithmic methods, like the EyeToy/DigiMask system take time, whereas ours can literally keep up with a video camera's 30 FPS when generating real time quality lower density meshes.
