Remember when you couldn’t scroll through your Instagram feed without seeing an image from Prisma, the app that uses AI to make photographs look like paintings?
Well, we’re ready to go in the opposite direction.
Researchers from UC Berkeley‘s AI Research Lab (BAIR) have discovered a way to reconstruct photographs from paintings using a technique known as “image-to-image translation.” The researchers also applied their work to a number of other techniques, like converting paintings to different painters’ styles and replacing objects in photos.
The findings, presented in a paper titled “Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks,” shows examples of the technique in action.
Basically, the technique involves recognizing “special characteristics” in an image and translating them into another one. Here’s an example of a Monet converted into a photograph:
And here’s another example that applies Renoir’s style to an image from Beauty and the Beast:
Using the same techniques, the researchers were also able to turn images into different painting styles. By using a photograph as input, they were able to produce different painters’ styles as output.
They also applied the research to another technique called object transfiguration. Using this, they were able to turn apples into oranges and vice versa.
Of course, the technique is a work in progress and doesn’t always produce great results. Sometimes you’re left with something pretty unsettling, like this:
For now, the project is still in prototype stages. You can learn more about it here.