"The GAN is dead; long live the GAN! A Modern GAN Baseline"

This paper proposes a simplified GAN architecture with improved stability and performance, challenging the notion that GANs are difficult to train.

https://arxiv.org/abs/2501.05441

🤔 Original Problem:

→ GANs are notoriously unstable and prone to mode collapse, leading to a reliance on ad-hoc tricks and outdated architectures.

→ The field has been held back by the belief that GANs are inherently tricky to train and don't scale well.

-----

💡 Solution in this Paper:

→ The authors introduce R3GAN, a new baseline GAN that combines a regularized relativistic loss with modern network architectures.

→ They augment the relativistic pairing GAN loss with zero-centered gradient penalties to improve stability and diversity.

→ The well-behaved loss allows them to discard ad-hoc tricks and replace outdated backbones with modern architectures.

→ Key architectural choices include proper ResNet design, initialization, resampling, grouped convolution, and no normalization.

-----

🔑 Key Insights from this Paper:

→ Regularized relativistic GAN loss addresses mode dropping and non-convergence issues

→ Combining R1 and R2 gradient penalties is crucial for stable training

→ Modern ConvNet designs can significantly improve GAN performance

→ Normalization layers can be harmful in generative models

-----

📊 Results:

→ R3GAN surpasses StyleGAN2 on FFHQ-256 (FID 2.75 vs 3.78)

→ Achieves full mode coverage on StackedMNIST (1000/1000 modes)

→ Outperforms many SOTA GANs and diffusion models on CIFAR-10 and ImageNet