Freeze Discriminator: A Simple Baseline for Fine-tuning GANs
share
Generative adversarial networks (GANs) have shown outstanding performance on a broad range of computer vision problems, but often require enormous training data and computational resources. Several works propose a transfer learning scheme to handle this issue, but they are prone to overfitting or too restrictive to learn the distribution shift. In this paper, we find that simply fine-tuning the networks while freezing the lower layers of the discriminator surprisingly works well. The simple baseline, freeze D, significantly outperforms the prior methods in both unconditional and conditional GANs, under StyleGAN and SNGAN-projection architectures and Animal Face, Anime Face, Oxford Flower, CUB-200-2011, and Caltech-256 datasets. Code and results are available in https://github.com/sangwoomo/freezeD.
READ FULL TEXT