This paper presents an adaptive discriminator augmentation (ADA) method to stabilize training of generative adversarial networks (GANs) with limited data. The method prevents discriminator overfitting by applying a diverse set of augmentations to the discriminator's input, while ensuring that none of the augmentations leak to the generated images. The approach does not require changes to loss functions or network architectures and is applicable both when training from scratch and when fine-tuning an existing GAN on another dataset. The authors demonstrate that good results are now possible using only a few thousand training images, often matching StyleGAN2 results with an order of magnitude fewer images. They also find that the widely used CIFAR-10 benchmark is a limited data benchmark, and improve the record FID from 5.59 to 2.42. The key problem with small datasets is that the discriminator overfits to the training examples; its feedback to the generator becomes meaningless and training starts to diverge. The authors propose a way to tackle this problem by employing versatile augmentations that prevent the discriminator from becoming overly confident. They design a diverse set of augmentations and an adaptive control scheme that enables the same approach to be used regardless of the amount of training data, properties of the dataset, or the exact training setup. The authors also introduce an adaptive discriminator augmentation method that dynamically adjusts the augmentation strength based on the degree of overfitting. This method uses two heuristics to determine the appropriate augmentation strength: one based on the difference between the training set and validation set, and another based on the sign of the discriminator outputs. The method is tested on several datasets and shows significant improvements in FID and other metrics. The authors also evaluate the method on several smaller datasets, including METFACES, BRECAHAD, and AFHQ, and find that ADA provides significant improvements in image quality and diversity. They also find that the method is effective in transfer learning scenarios, where a pre-trained GAN is used to train on a new dataset. The authors conclude that their method significantly improves the performance of GANs on limited data scenarios, and that the method is effective in both training from scratch and transfer learning. They also note that the method has a negligible impact on the energy consumption of training a single model, making it a practical solution for real-world applications.This paper presents an adaptive discriminator augmentation (ADA) method to stabilize training of generative adversarial networks (GANs) with limited data. The method prevents discriminator overfitting by applying a diverse set of augmentations to the discriminator's input, while ensuring that none of the augmentations leak to the generated images. The approach does not require changes to loss functions or network architectures and is applicable both when training from scratch and when fine-tuning an existing GAN on another dataset. The authors demonstrate that good results are now possible using only a few thousand training images, often matching StyleGAN2 results with an order of magnitude fewer images. They also find that the widely used CIFAR-10 benchmark is a limited data benchmark, and improve the record FID from 5.59 to 2.42. The key problem with small datasets is that the discriminator overfits to the training examples; its feedback to the generator becomes meaningless and training starts to diverge. The authors propose a way to tackle this problem by employing versatile augmentations that prevent the discriminator from becoming overly confident. They design a diverse set of augmentations and an adaptive control scheme that enables the same approach to be used regardless of the amount of training data, properties of the dataset, or the exact training setup. The authors also introduce an adaptive discriminator augmentation method that dynamically adjusts the augmentation strength based on the degree of overfitting. This method uses two heuristics to determine the appropriate augmentation strength: one based on the difference between the training set and validation set, and another based on the sign of the discriminator outputs. The method is tested on several datasets and shows significant improvements in FID and other metrics. The authors also evaluate the method on several smaller datasets, including METFACES, BRECAHAD, and AFHQ, and find that ADA provides significant improvements in image quality and diversity. They also find that the method is effective in transfer learning scenarios, where a pre-trained GAN is used to train on a new dataset. The authors conclude that their method significantly improves the performance of GANs on limited data scenarios, and that the method is effective in both training from scratch and transfer learning. They also note that the method has a negligible impact on the energy consumption of training a single model, making it a practical solution for real-world applications.