Ever since Google wowed us with the surreal artwork produced by Deep Dream, the internet has been buzzing about the use of neural networks in creating art. Artificial intelligence expert and game developer Alex J. Champandard has recently been showing off another example of neural networks generating art and the results are blowing us away.
Neural Doodle is a Python script that can do a number of things but we’re most impressed by how it can take our rubbish doodles and turn them into masterpieces in the style of other artists. You supply the script with a style image (e.g. The Cliffs at Etretat by Monet) and your own silly doodle. The script uses the style image to recreate your doodle in the same style, as in the following example:
It’s capable of more than turning sketches into fine art. It can take two pieces of art and take the style from one and apply it to the other. It can also create amazing seamless textures from photos. This is Neural Doodle’s attempt at creating an HD grass texture:
— Alex J. Champandard (@alexjc) April 25, 2016
All this in a few hundred lines of code? It’s a work-in-progress but it’s looking fantastic already. Right now it exists as a Python script but in the future the technology could be added to artistic packages. The following video isn’t real but shows a mock-up of how an application like GIMP could use a Neural Doodle plug-in:
For those of us who can’t paint to save our lives, we’ve found yet another use for AI research. It’s also fascinating to see how biology is inspiring technology. We’ve had Deep Mind’s AlphaGo using deep reinforcement learning and now systems modelling human neural networks in order to recreate art. Perhaps someday computers will do this with video too.
If you want to try out Neural Doodle, you can grab it at GitHub. You won’t always get perfect results and there’s still a bit of tinkering required to get it to work perfectly with each picture but when it does work it’s so good it’s almost scary.
Main image © Alex J. Champandard