$\mu$StyleGAN

This is a simplified PyTorch implementation of StyleGAN2 introduced in Analyzing and Improving the Image Quality of StyleGAN paper

Idea

The main purpose is to implement a smaller version of the StyleGAN using another interpolation method and evaluate the quality of result model built by my own.

This is a study project to understand what is GANs, how they can be designed and how can be trained.

The code is based on https://github.com/labmlai/annotated_deep_learning_paper_implementations/tree/master/labml_nn/gan/stylegan

Tools used

• PyTorch
• Numpy
• scikit-learn
• matplotlib (just to visualize training results)

Dataset

• Santiel 5K Truecolor Landshapes-4041

Issues

• Because of small computational power the model can't be trained longer to achieve desired image quality
• Currently looking for an opportunity to run model in distributed mode (multiple gpu's/multiple nodes/combined)
• A simple web-based API for image generation is currently unimplemented.

The main differences from original StyleGAN

• The network generates 32 $\times$ 32 px instead of 1024 $\times$ 1024 px.
• Smaller architecture: mapping network (5 layers used instead of 8 in original), generator and discriminator ($\log_2(32) = 5$ blocks instead of $\log_2(1024) = 10$)
• Noise injection was performed in two steps:
  • 1. Generate the base noise before mapping styles.
  • 2. In each style block add randomly generated noise instead of using pregenerated outter.
• Default interpolation method is nearest-exact instead of bilinear so produced images are more smooth.

Results

Loss

Image 1 - Generator and discriminator train losses per epochs

Generated samples

Image 2 - training results per epochs: 1200 epochs (a), 1500(b), 3200(c), 3500(c), 4200(e)

Generated and original images comparison

Image 3 - Comparison of zoomed generated (left) and real (right) images