Comparison between a regular GAN and the resize-convolution propsed in the article http://distill.pub/2016/deconv-checkerboard/ as an alternative to the standard deconvolution (transposed convolution) in the generator to get rid of the checkboard artifacts.
The upsampling was made using the 2d unpooling function in Chainer which is very similar to a nearest-neighbor interpolation.
Training dataset of CIFAR-10.
We see that the artifacts are less noticeable in the resize-convolutions and that the images look more natural, but that the difference becomes less obvious as the training goes on.
Left images are images generated using a regular GAN and right images are generated using the resize-convolutions technique.
