The actor co-wrote a case study on the ‘neural style transfer’ method used in the short film she directed, Come Swim
Kristen Stewart has co-authored a research paper about a technique involving artificial intelligence utilised in her most recent film. ‘Neural style transfer’ is used to apply the aesthetic of an image onto another.
Stewart wrote Bringing Impressionism to Life with Neural Style Transfer in Come Swim, alongside Adobe research engineer Bhautik J Joshi and producer David Shapiro, exploring the technique used in her directorial debut. Come Swim was recently shown at Sundance.
Neural style transfer was used in the film’s opening and closing scenes to add a ‘painting’ effect to footage. The paper describes the short film as “a poetic, impressionistic portrait of a heartbroken man underwater”, “grounded in a painting of a man rousing from sleep”, which “evokes the thoughts an individual has in the first moments of waking”.
Stewart and the team who worked on the 17-minute film originally tried to tune the algorithm to create the desired effect, but instead found it was simpler to modify the images from the film, cropping and adding texture that the algorithm would then pick up.
Come Swim is described in a synopsis as “a diptych of one man’s day; half impressionist and half realist portraits”.
The paper is available on Cornell University Library’s online scientific research repository arXiv.org.