This is pretty insane. I’m surprised that no one is falling all over this. I loved his description of GANs. I had no clue these were a thing before this article, but the whole generator/critic model seems like it would be incredibly useful for MANY things involving artistic endeavors, not just restoring photos. For instance, I can imagine it being used as a vastly improved way to generate music. Part of the issue with algorithms that generate music is that they just don’t sound great, and I think it’s because art is very open-ended. However, there are rules and guidelines that musicians follow depending on the genre, and human players discover these rules by listening to existing music. Existing algorithms to generate music have, of course, relied on literal prior art, but the critic is usually the human. If the critic were the computer instead, then that could open a lot of doors.
Hello, creator here. Great minds think alike :) This thought about music has been on my mind too. I think it’ll be a much more ambitious leap from what’s been done currently but it seems like in principle it should work as you describe it more or less.
If you’re interested in GAN art, there’s somebody I really like to follow on Twitter- Helena Sarin. https://twitter.com/glagolista
And I’m posting my work likewise in a steady drip here:
There’s not too many of us yet as far as I know but I do think it’s going to be a very interesting next decade in terms of doors opening wide open in the arts and creative space with this stuff!
The WaveNet examples have some fascinating piano clips that were generated one sample at a time. Definitely worth a listen! For actually creating music I found the recent audio style transfer technique pretty impressive. At times it sounds just like a bad vocoder, but the results from a whistled Indiana Jones theme from 1:47 onwards is really fun.