The algorithm was given this photo of buildings, left, and a copy of Vincent Van Gogh’s “The Starry Night.” In about an hour it taught itself to mimic Van Gogh’s style, and apply it to the photo of the buildings. (University of Tuebingen)
For great artists, creating a masterpiece is the culmination of a career. Years of practice, creative musings and experimentation with styles build up to the genesis of something truly original and timeless.
A story is often told about Pablo Picasso charging an enormous sum for a portrait. The physical act of drawing it took only a few moments, so the subject complained to Picasso. He is said to have responded, “Madame, it took me my entire life.”
While the great art of yesteryear was an exhaustive process to create, today the style of those masters can be mimicked in minutes. Last week German researchers released a paper detailing how a computer algorithm could be used to pump out images borrowing the styles of the world’s greatest artists.
The algorithm took the top left image and applied the style of the small, inset images to create works with very similar styles. (University of Tuebingen)
The paper includes new images that look remarkably like the works of Vincent Van Gogh, Edvard Munch and Picasso. All the images were created in about an hour. As the researchers optimize their work, the images should be able to be produced even faster. Something that was rare and difficult could become quick and easy to create. (The researchers have not publicly released the algorithm yet.)
The findings are the latest out of the red-hot field of deep learning, in which computers identify and classify patterns in huge data sets. The computer’s thought processes are designed to mirror the way the human brain works.
The researchers create their new images by taking one image, such as a bland photo of a row of houses, and another image, a painting from a great artist. They then use what’s called a convolutional neural network to create a new image, in which the style of the great artist is applied to the other photo.
Here’s a demonstration of how the network creates the new image. (University of Tuebingen)
“The key finding of this paper is that the representations of content and style in the convolutional neural network are separable. That is, we can manipulate both representations independently to produce new, perceptually meaningful images,” write the authors. Their paper has been submitted to Nature Communications. Leon Gatys, a PhD student at the University of Tuebingen is its lead author.
This is another example of how machines are becoming more capable of matching and exceeding human capabilities. Of course, the machines don’t have a style all their own. They can only copy existing styles. True creativity is a greater challenge that machines have yet to master.
Matt McFarland is the editor of Innovations. He's always looking for the next big thing. You can find him on Twitter and Facebook.