Could there be any way to create beautiful landscape paintings without touching a brush and getting you hands dirty? Yes, with code, some training data and GANs.
This week about 100 PhD students are attending WASP Summer School 2019. The AI and AS track have their own schedules but they run in parallel. During Tuesday the AI-class got introduced to a colorful challenge – to create landscape masterpieces using neural networks.
– The Bob Ross Challenge let the students combine their creative and technical skills and there’s a lot of training data available for them to use. And well, there’s also some nostalgia in choosing Bob Ross, says professor Amy Loutfi.
In small groups the students went in to the challenge.
– I actually had some time during summer so already I tried the challenge. It’s a nice experiment, and to generate images is quite new. I’ll try to create some more pictures during this week and maybe we’ll end up with some crazy modern art painting before Friday, says Mohamadreza Faridghasemnia, working with Quantao Yang, PhD students at Örebro University.
– We started by finding some images. We’re planning to try transfer learning since it will be hard to get a lot of images – there’s only about 300-400 images and that’s tricky to get a good network on. So we’ll try to find a pretrained network that is trained to produce landscape images, like Bob Ross’, and then try to finetune them for Bob Ross paintings. It would be fun to do something nice but we’ll see. It’s quite ambitious to do this in just one week, says Albin Heimerson, Johan Grönqvist and Alexander Dürr from Lund University.
On Friday, the last day of summer school, there will be an exhibition and of course – there is a prize to be collected.