Disney built a neural network to automatically change an actor’s age

The process of using digital effects to alter actors’ age is an increasingly common practice in Hollywood (despite its many criticisms), but for every uncannily younger Robert De Niro in The Irishman, hundreds of hours of painstaking work is required from behind-the-scenes VFX artists and engineers. As neural networking becomes more ubiquitous and accessible, however, entertainment industry heavy-hitters have an opportunity to harness the tech to streamline the process while simultaneously creating more photorealistic results. Right now, there are few entertainment companies as heavy-hitting as Disney—which is why it will probably come as little surprise to hear that the House of Mouse just announced it is harnessing some cutting-edge AI to “re-age” actors.

[Related: Disney is turning to robots to pull off dangerous aerial feats.]

Yesterday, Disney Research Studios revealed the Face Re-Aging Network (or FRAN), its latest advancement in VFX work which utilizes neural networking to create “the first practical, fully automatic, and production-ready method for re-aging faces in video images.” According to Disney’s own count, at least 12 films and television series utilized re-aging tech during 2022, a number that’s steadily grown since its first introduction only a few years ago. “Photorealistic digital re-aging of faces in video is becoming increasingly common in entertainment and advertising. But the predominant 2D painting workflow often requires frame-by-frame manual work that can take days to accomplish, even by skilled artists,” the team explains in the abstract of their research paper. FRAN’s solution to the issue relies on a multistep neural network process that eases labor time and costs while improving the effect’s realism.

Watch Disney’s video rundown below:

As Gizmodo and Ars Technica also explained earlier today, FRAN utilizes a number of steps to create its re-aged subjects. First, the team used a program called StyleGAN2 to randomly generate thousands of synthetically aged faces between 18- and 85-years-old. Once that database was established, machine learning tools aged and de-aged these artificial portraits, which were then fed into another neural network, FRAN. From there, FRAN can apply what it learned to videos of actual people, adding or subtracting years regardless of angle, position, movement, or lighting. Afterwards, artists can go in and manually touch-up frames and transitions as needed.

[Related: Machine learning can measure mood on social media.]

The results still aren’t perfect, but they’re arguably about as convincing as anything seen so far in movies like Rogue One or Ant-Man and The Wasp. If nothing else, the final products will only improve over time as neural networks like FRAN get smarter with their techniques.