We observe that despite their hierarchical convolutional nature, the synthesis process of typical generative adversarial networks depends on absolute pixel coordinates in an unhealthy manner. This manifests itself as, e.g., detail appearing to be glued to image coordinates instead of the surfaces of depicted objects. We trace the root cause to careless signal processing that causes aliasing in the generator network. Interpreting all signals in the network as continuous, we derive generally applicable, small architectural changes that guarantee that unwanted information cannot leak into the hierarchical synthesis process. The resulting networks match the FID of StyleGAN2 but differ dramatically in their internal representations, and they are fully equivariant to translation and rotation even at subpixel scales. Our results pave the way for generative models better suited for video and animation.
核心贡献 · Key contributions
识别出 GAN 生成器中的混叠是纹理粘附于像素坐标的根本原因。 Identified aliasing in GAN generators as root cause of texture sticking to pixel coordinates.
提出连续信号解释和架构修改以消除混叠。 Proposed continuous signal interpretation and architectural changes to eliminate aliasing.
在 StyleGAN3 中实现了亚像素级别的完全平移和旋转等变性。 Achieved full translation and rotation equivariance at subpixel scales in StyleGAN3.
匹配 StyleGAN2 的 FID，同时显著改变内部表示。 Matched FID of StyleGAN2 while dramatically altering internal representations.
为更适合视频和动画的生成模型铺平道路。 Paved way for generative models better suited for video and animation.
引入实用近似，如非线性操作的 2 倍上采样和 Kaiser 窗滤波器。 Introduced practical approximations like 2x upsampling for nonlinearities and Kaiser-windowed filters.
局限 · Limitations
假设训练数据已正确抗混叠；混叠数据（如像素艺术）会导致困难。 Assumes training data is properly antialiased; aliased data (e.g., pixel art) causes difficulties.