GauGAN – yes it’s an artistic Generative Adversarial Network with a cute name – can draw a landscape that you specify in very general terms. It is so good it just won two awards at this year’s SIGGRAPH. Is this where art is heading?
GauGAN, NVIDIA’s real-time AI art application, is a neural network that has been trained on one million real landscape images and can now create a real looking landscape image from a user generated segmentation map.
A segmentation map is just a division of the drawing space into sky, mountains, sea and so on. The network converts the segmentation map into the realistic scene by filling it with appropriate detail that fits together correctly. In addition it uses style transfer to modify the image into a requested form – daytime, nighttime, oil painting etc and you can upload your own filters for added “creativity”.
You can see it in action in the following video:
If you feel inspired you can also try the software out in NVIDIA’s AI Playground.
I did and I have to say that the experience was addictive and troubling. I kept on creating new landscapes and trying different styles. Things got interesting when you specified something that was unlikely – like a mountain that had a hole in it through which you could see the sky or the sea.
There is a certain smeary, almost dreamy, quality about many of the results that you might associate with neural network generated art. All great fun, but what troubled me was what exactly was I doing? The “degrees of freedom” of the segmentation map is far lower than if I sit in front of a blank canvas with drawing tools – even more so with traditional digital drawing tools.
In the traditional case I am confronted with a blank canvas and have to have some sort of idea, mental image, plan – something. In this case I can create a starting point in no time at all and the blank canvas is gone. It is like trying to write poetry where you are given some verses and try to improve on them or inject some creativity into them.
At the SIGGRAPH presentation NVIDIA’s Gavriil Klimov said:
“As an artist it’s extremely valuable to be able to generate content quickly because artists need to iterate fast. GauGan allows us to generate our ideas at a speed that was previously not possible, and this will become a main tool for artists to use in the future.”Gavriil Klimov
This seems to be a production-line approach to creativity, which might work for some things but not all.
You can also get the source code from NVIDIA for non-commercial use which means you can play around with it all the more.