Take a wondrous and slightly nauseating journey through western art history in a new video by Alexander Mordvintsev, which uses machine learning to create an endless vortex of paintings that zoom into one another. A software engineer at Google, Mordvintsev created DeepDream, a computer vision program that uses neural networks to interpret and generate new, often creepy images.
For this latest video, he trained neural networks to generate a wild stream of paintings, where you get closer and closer to an artwork until its fine, high-resolution brushstrokes reveal an entirely new artwork, which camouflages another mini artwork, and so forth. Mordvinsev’s video lasts just four minutes, which is a good length, as the effect, mesmerizing as it is, is dizzying.
“We explore various image parameterizations for neural input optimization-based image generation, and demonstrate usefulness of these parameterizations on a few creative image generation tasks,” Mordvintsev wrote of the video.
The clip was submitted to an online art gallery launched for a recent workshop hosted by The Neural Information Processing Systems (NIPS) Foundation, a nonprofit that explores research in neural information processing systems. It was one of three that Mordvintsev created, along with one that similarly zooms in to various Google Doodles as well as one that examines morphing patterns on bunnies.
Although this video’s source material is limited, focusing on works by the most famous Impressionists and Post-Impressionists, it highlights interesting connections between the colors and textures of different paintings. Monet’s portrait of Camille Monet, for instance, is found within a fold of a jacket worn by one of Cézanne’s card players; the nose of Joseph Roulin, painted by van Gogh, holds a Cézanne still life.