kstew.jpg

The Twilight star helped co-write a paper on machine learning and discusses the style transfer technique. An example of how the technique works (image via ArXiv)

 

From watching her play Bella in the Twilight series or just trying to read her stone face, it’s hard to tell what Kristen Stewart really likes. Most fans will be surprised to learn she has an interest in AI so much she recently co-wrote a paper on machine learning. The paper, released on ArXiv, discusses a technique known as “style transfers,” which she uses in certain scenes for her upcoming movie, Come Swim.

 

The technique takes an image and transforms it into the artistic technique and color profile of another image. Think of it as giving your selfies a Van Gough filter to imitate the iconic art style. The process is currently being used by popular apps, like Prisma. How the system works is through deep neural networks to find the “content” of the picture and the “style” of the other and putting them together to create a new image.

 

Stewart, who directs the film, used this technique to create trippy, dream-like scenes for the movie. But she didn’t want to copy the style of Picasso or Rembrandt. Rather she used her own paintings to get the look she wanted. It wasn’t easy to recreate what she pictured in her head, especially spanning over multiple frames.

 

Stewart’s team eventually got the effect they wanted by modifying the images instead of experimenting with the algorithm itself. What they did is cropped and added pieces of texture to the pictures making it easier for the algorithm to pick up the influences when creating the final image. They also experimented with how often the technique should be applied, or “style transfer ratio.” The image’s texture improved when increasing iterations, but after rough 256 times, no further improvements could be seen.

 

Aside from Stewart, special effects engineer Bhautik J Joshi and producer David Shapiro worked on the paper. Since ArXiv is an online repository, it hasn’t been peered reviewed, which will send off some alarms for some researchers. Whether or not Stewart’s peers take her research seriously shouldn’t matter. Not only does this paper show off a cool and unique technique for filmmaking, but it also shows how big name celebrities are getting involved with technology. Who knows, maybe Brad Pitt secretly likes robots and plans to build one. At least, we can hope that’s what’s happening.

 

Have a story tip? Message me at: cabe(at)element14(dot)com

http://twitter.com/Cabe_Atwell