got mad at the circles GAN for mode collapsing repeatedly. but then i read (most of) ‘s anime faces write-up and learned maybe why: GANS don’t like line art! https://www.gwern.net/Faces
they like photographs which are just a bunch of blurry color textures up close. but line art and monochrome images seem to not work as well
i read it last night and learned a lot. but if you’re trying to fall asleep i cannot recommend it, because the page is like twelve feet long and interesting. i was up like an hour later than i meant to be lol
interesting 🤔 My data set is more like MNIST than FFHQ actually
That's certainly what a lot of the lowcap ordinary DCGANs produce: much muddy texture that people mostly imagine output variation in because of color and sample #. Interestingly, bad GANs can work ok on MNIST but not other things. I call this the synthesis competence hierarchy.
— Andrew Burton (@ForABurton) April 5, 2021