While it’s possible for these stacked deconvolutions to cancel out artifacts, they often compound, creating artifacts on a variety of scales. However, switching deconvolutional layers for resize-convolution layers makes the artifacts disappear. If gradient artifacts can affect an image being optimized based on a neural networks gradients in feature visualization, we might also expect it to affect the family of images parameterized by the generator as they’re optimized by the discriminator in GANs.