“The GAN Is Dead; Long Live the GAN! R3GAN: A Modern GAN Baseline”, Nick Huang, Aaron Gokaslan, Volodymyr Kuleshov, James Tompkin2024-11-06 (StyleGAN; similar)⁠:

There is a widely-spread claim that GANs are difficult to train, and GAN architectures in the literature are littered with empirical tricks. We provide evidence against this claim and build a modern GAN baseline in a more principled manner.

First, we derive a well-behaved [zero-centered gradient-penalty] regularized relativistic GAN loss that addresses issues of mode dropping and non-convergence that were previously tackled via a bag of ad-hoc tricks. We analyze our loss mathematically and prove that it admits local convergence guarantees, unlike most existing relativistic losses. Second, this loss allows us to discard all ad-hoc tricks and replace outdated backbones used in common GANs with modern architectures. Using StyleGAN-2 as an example, we present a roadmap of simplification and modernization that results in a new minimalist baseline—R3GAN. [Briefly, we find that proper ResNet design^{17, 67}, initialization⁹⁹, and resampling^{29, 31, 32, 100} are important, along with grouped convolution^{95, 5} and no normalization.^{31, 34, 14, 88, 4}]

Despite being simple, our approach surpasses StyleGAN-2 on FFHQ, ImageNet, CIFAR, and Stacked MNIST datasets, and compares favorably against state-of-the-art GANs and diffusion models.

Code: Github.

Similar Links: