steffen-jung/SpectralGAN-StyleGAN2

Code accompanying the AAAI 2021 paper "Spectral Distribution Aware Image Generation" to finetune StyleGAN2.

SpectralGAN: Spectral Distribution Aware Image Generation

• Authors: Steffen Jung and Margret Keuper
• Paper: https://arxiv.org/abs/2012.03110
• Code to train DCGAN-like architectures in PyTorch: https://github.com/steffen-jung/SpectralGAN
• This code is an extension of stylegan2-ada

Motivation

Commonly used Generative Adversarial Networks (GANs) are not able to learn the distribution of real datasets in the frequency domain.

Our method adds an additional discriminator increasing the spectral fidelity.

Spectral fidelity without our method

Spectral fidelity with our method

Requirements

Same as in stylegan2-ada.

Usage Example

Usage is the same as explained in stylegan2-ada, however, you need to use train_spectral.py instead of train.py.

Example to finetune StyleGAN2:

python train_spectral.py \
   --resume /path/to/network.pkl \
   --data /path/to/tfrecords \
   --gpus 4 \
   --outdir /path/to/folder \
   --kimg 7000 \
   --cfg stylegan2 \
   --aug noaug \
   --metrics fid50k_full

Citation

@inproceedings{Jung2021SpectralGAN,
   title     = {Spectral Distribution Aware Image Generation},
   author    = {Steffen Jung and Margret Keuper},
   booktitle = {Thirty-Fifth AAAI Conference on Artificial Intelligence},
   year      = {2021}
}

References

• This code is an extension of "StyleGAN2 with adaptive discriminator augmentation (ADA)": https://github.com/NVlabs/stylegan2-ada