The paper "Alias-Free Generative Adversarial Networks" by Tero Karras et al. addresses the issue of aliasing in generative adversarial networks (GANs), which causes fine details to appear fixed in pixel coordinates rather than being naturally transformed with the underlying coarse features. The authors trace the root cause to careless signal processing in the generator network, leading to aliasing. They propose a set of architectural changes that ensure continuous equivariance to translation and rotation, even at subpixel scales, by interpreting all signals in the network as continuous. These changes result in a generator that matches the FID score of StyleGAN2 but differs significantly in its internal representations, allowing for more natural hierarchical refinement of details. The new generator, called alias-freeGAN3, is shown to produce images with improved texture and motion coherence, making it better suited for applications such as video and animation generation. The paper also discusses the practical implementation of these changes, including the use of high-quality filters and the removal of per-pixel noise inputs.The paper "Alias-Free Generative Adversarial Networks" by Tero Karras et al. addresses the issue of aliasing in generative adversarial networks (GANs), which causes fine details to appear fixed in pixel coordinates rather than being naturally transformed with the underlying coarse features. The authors trace the root cause to careless signal processing in the generator network, leading to aliasing. They propose a set of architectural changes that ensure continuous equivariance to translation and rotation, even at subpixel scales, by interpreting all signals in the network as continuous. These changes result in a generator that matches the FID score of StyleGAN2 but differs significantly in its internal representations, allowing for more natural hierarchical refinement of details. The new generator, called alias-freeGAN3, is shown to produce images with improved texture and motion coherence, making it better suited for applications such as video and animation generation. The paper also discusses the practical implementation of these changes, including the use of high-quality filters and the removal of per-pixel noise inputs.