News for hackers. Stuff that matters.
Re:Creating photorealistic images with neural networks and a... by gwern on Friday February 17, 2017 @11:53AM (#13669684)
"Thanks to this post http://distill.pub/2016/deconv-checkerboard/ I found out about the alternative deconvolution layer and implemented that one."

Yeah, I saw the same thing with WGAN. ConvTranspose2d is not that great for upscaling because it creates artifacts. That said, the post actually recommends doing a 'subpixel' convolution for upscaling (something like, do a convolution out to channels for each pixel, then use PixelShuffle to map it back into a 3-channel image), not doing a bilinear/nearest-neighbor + Conv2d(3,3).

(A GAN would probably also deliver better colorizing results in general.)

There are no comments on this comment yet.