steffen-jung / SpectralGAN

Code accompanying the AAAI 2021 paper "Spectral Distribution Aware Image Generation".
23 stars 1 forks source link

SpectralGAN: Spectral Distribution Aware Image Generation

Motivation

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

Our method adds an additional discriminator increasing the spectral fidelity.

Spectral fidelity without our method

DCGAN

Spectral fidelity with our method

SpectralDCGAN

Requirements

Tested on:

Usage Example

Train a new model:

python Training.py \
   --device cuda:0 \
   --name Debugging \
   --experiments_folder /path/to/folder \
   --data_folder /path/to/data_folder \
   --epochs 100 \
   --img_size 64 \
   --img_nc 3 \
   --loss lsgan \
   --d_spectral linear

Continue training:

python Training.py \
   --device cuda:0 \
   --data_folder /path/to/data_folder \
   --epochs 50 \
   --img_size 64 \
   --img_nc 3 \
   --loss lsgan \
   --d_spectral linear \
   --checkpoint /path/to/previous/runfolder

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