Awhile ago I tried to train a GAN to let you generate your own ‘original’ Bob Ross paintings. It requiredsome time creating the dataset and lots of tracing, but in the end only worked decently for winter paintings- I’m guessing because of their less complicated compositions and colors palettes.
Random facts: Bob Ross is missing his left index finger and avoids including human evidence in his paintings (eg no chimneys).
They were trained in the paintings data (hence why I collected it!). That quality is from, I believe, the gan having some difficulty figuring out how to color each pixel. But I’m not sure entirely on that, though in the time since I made that, GANs have increased by magnitudes in output quality.
Random facts: Bob Ross is missing his left index finger and avoids including human evidence in his paintings (eg no chimneys).
GAN: https://twitter.com/jdwlbr/status/1131244682317484032?s=20
Data: https://github.com/jwilber/Bob_Ross_Paintings