TropComplique / multi-scale-gradient-gan

Generation of high resolution fashion images

Multi-Scale Gradient GAN

This an implementation of MSG-GAN: Multi-Scale Gradient GAN for Stable Image Synthesis.

Notes

This implementation differs from the one in the original paper.

1. To regularize the discriminator I use R1 penalty.
2. To improve sample diversity I use PacGAN discriminator.
3. I use a stylegan-like generator (or you could call it a self modulation):
   latent vector injection with AdaIN layers.

How to train

1. Put your images in a folder.
2. Edit training config in the beginning of train.py file.
3. Run python train.py for training.
4. Run tensorboard --logdir=summaries/run00/ to view losses.
5. Use generation.ipynb to generate samples with a trained model.

Example of generated samples

For training I used FEIDEGGER dataset. All hyperparameters that I used are in train.py.
Note that the generated samples came from different epochs.

Requirements

1. pytorch 1.2
2. tqdm, Pillow 6.1
3. tensorboard 1.14

Acknowledgments

This code is based on

1. akanimax/BMSG-GAN
2. rosinality/style-based-gan-pytorch

Relevant papers

1. On Self Modulation for Generative Adversarial Networks
2. A Style-Based Generator Architecture for Generative Adversarial Networks
3. PacGAN: The power of two samples in generative adversarial networks
4. Which Training Methods for GANs do actually Converge?