Physics arXiv Blog:

Yihao and co take a different approach using two different techniques. The first involves the kind of warping already tried but on a local scale rather than an entirely global one.

The algorithm divides the face into regions such as the eyes, mouth, nose etc and measures the distortion in each of these areas independently. In particular, it measures any change in the height-to-width ratio of each facial feature.

It then applies the same warping to each part of the target face. And to ensure that this warping is realistic, it constrains the changes using the measured height-to-width ratio. That ensures that the facial feature never becomes too large or small in relation to the rest of the face.

The second technique produces accurate shadows on the distorted face. To do this, Yihao and co have created a computer model of the facial muscle groups involved in different expressions and the way these change the topology of the face. The algorithm works out which groups have been used to create a certain expression and then adjusts the shadows in the distorted face accordingly.

While there will surely be valuable uses for this technology, I can guarantee that it will devolve into silliness 90% of the time.