gwern, first off I'd like to mention I absolutely love your work! It's always a pleasure to read your longform writeups on StyleGAN, even if half of it admittingly goes over my head. :)
Can I ask what was modified recently about the write up? I remember reading it a few days back but I see at the top that it was recently modified.
As an aside, I wonder when we'll see the first "fake artist" using TWDNE and Waifu2x then photoshopping/cleaning up some of the irregularities of the generations to output a bunch of "well drawn art". Someone who's decently talented could also use it as a shortcut to avoid a lot of time spent drawing faces/hairstyles and use the generated faces as a starting point.
I had a thought once, years ago, about trying to work out the math behind the bezier curves that make up a cartoon image of a human face - it seems like you'd have to be able to work it out to a series of control points relative to some reference point. I never had the free time to pursue it, and the fact that nobody else never did makes me think I must have been underthinking how complex that actually would be - still, I find it surprising that you have to apply the complexity of a neural network before you can generate a believable cartoon image of a human.
4 comments
[ 3.2 ms ] story [ 22.7 ms ] threadCan I ask what was modified recently about the write up? I remember reading it a few days back but I see at the top that it was recently modified.
As an aside, I wonder when we'll see the first "fake artist" using TWDNE and Waifu2x then photoshopping/cleaning up some of the irregularities of the generations to output a bunch of "well drawn art". Someone who's decently talented could also use it as a shortcut to avoid a lot of time spent drawing faces/hairstyles and use the generated faces as a starting point.
I've been writing it from the top on down. So the past few days I spent adding the preliminary whole-Danbooru2018 StyleGAN runs ( https://www.gwern.net/Faces#anime-faces-danbooru2018 https://www.gwern.net/Faces#anime-faces-ffhq-danbooru2018 ), the FFHQ transfer experiments ( https://www.gwern.net/Faces#ffhq-variations ), the appendix ( https://www.gwern.net/Faces#appendix ) with all my past GAN experiments (required quite a bit of spelunking through my Twitter & dusty source code repos), the encoding/editing of StyleGAN images ( https://www.gwern.net/Faces#reversing-stylegan-to-control-mo... ), and of course a whole bunch of formatting & copyediting.