Alias-Free Generative Adversarial Networks (GANs) address the issue of "texture sticking" in image synthesis by ensuring that details are correctly attached to surfaces rather than fixed to pixel coordinates. The problem arises from aliasing in the generator network, which causes unwanted information to leak into the hierarchical synthesis process. The authors propose a method that interprets all signals as continuous, leading to a fully equivariant generator that is translation and rotation invariant at subpixel scales. This approach improves the internal representations of the network, resulting in a generator that matches the FID of StyleGAN2 but with different internal structures. The new StyleGAN3 generator is more suitable for video and animation due to its natural hierarchical transformation properties. The key contributions include the use of continuous signal interpretation, improved upsampling filters, and handling of nonlinearities to suppress aliasing. The generator is redesigned to be equivariant to translation and rotation, with a focus on continuous signal processing and low-pass filtering. The results show that the new generator produces more natural motion and better maintains the illusion of a coherent 3D scene. The paper also discusses the limitations and future directions, including the potential for extending equivariance to scaling and other transformations. The implementation and pre-trained models are available for use.Alias-Free Generative Adversarial Networks (GANs) address the issue of "texture sticking" in image synthesis by ensuring that details are correctly attached to surfaces rather than fixed to pixel coordinates. The problem arises from aliasing in the generator network, which causes unwanted information to leak into the hierarchical synthesis process. The authors propose a method that interprets all signals as continuous, leading to a fully equivariant generator that is translation and rotation invariant at subpixel scales. This approach improves the internal representations of the network, resulting in a generator that matches the FID of StyleGAN2 but with different internal structures. The new StyleGAN3 generator is more suitable for video and animation due to its natural hierarchical transformation properties. The key contributions include the use of continuous signal interpretation, improved upsampling filters, and handling of nonlinearities to suppress aliasing. The generator is redesigned to be equivariant to translation and rotation, with a focus on continuous signal processing and low-pass filtering. The results show that the new generator produces more natural motion and better maintains the illusion of a coherent 3D scene. The paper also discusses the limitations and future directions, including the potential for extending equivariance to scaling and other transformations. The implementation and pre-trained models are available for use.