Last June we talked about a project from the University of Washington that explored the possibility of creating "digital puppets" of various people, including actors and politicians. Now, a new development from Stanford University raises the stakes with a platform capable of capturing and replacing expressions in real time on the face of anyone in a conventional video, using just a webcam.
Fakes are all over the web. Slowly, the line separating the real from the fictional is getting thinner with each new generation of design and editing tools, combined with the power of algorithms, artificial intelligence and machine learning. Modifying a photo is so easy that anyone can do it, but a true master of the commands can leave us confused for weeks, or even months. When it comes to video, alterations are still too obvious to fool a trained eye, but we've already seen what Hollywood can do by replacing dead or elderly actors with digital avatars. What follows is even more surprising, and why not, more concerning.
The Face2Face project is made up of researchers from Stanford University, the Max Planck Institute, and the University of Erlangen-Nuremberg. Basically, it proposes real-time facial capture and its later recreation by replacing the original expressions. In this way, the system makes a person say something that never actually came out of their mouth, assuming other details such as training and voice have been solved. If Face2Face's results aren't surprising on their own, then everyone will give in to the fact that it works with conventional RGB videos (for example, YouTube), and uses a simple webcam as the source to detect replacement expressions.
Face2Face is still far from perfection, although time is on its side, and all the "errors" in rendering that appear around the mouth should gradually disappear later on. Using politicians in Face2Face's demonstrations definitely has greater impact. False statements from our leaders are much more common than their truths, but if we now have something like Face2Face, well...
