Re:Image Synthesis from Yahoo's open_nsfw
by gabrielgoh on Sunday October 23, 2016 @04:28AM
(#12772588)
the artifacts are explained in detail here
http://distill.pub/2016/deconv-checkerboard/
more compute will enable higher resolutions and less artifacts, but it means training a bigger GAN
Is it a compute/processing power thing? Does it take a lot more time to generate an image at say 5x resolution?
Also what I wondered about, quite a few of the images seem to have a sort of "canvas" like texture on them. Any idea why that is? Would a lot of the source images have canvas textures? Is it an artifact of the algorithm? It kinda feels like a highpass (edge enhance) filter as well. Or maybe they're just JPG artifacts?