This repository contains the code used for Image Generation using the Wasserstein Patch Prior, as well as a U-Net implementation in Torch to compare model performance when training on datasets with and without synthetic data. It is based on Elnekave and Weiss's generative patch distribution matching (paper, repository) and builds upon an earlier implementation by Moritz Piening.

Some examples showcasing the functionality of the codebase can be found in /examples.
All results are based on the following three publicly available datasets which should be downloaded into /data

• DRIVE (available e.g. at Kaggle)
• CrackForest (available e.g. at Kaggle)
• WHU Building Dataset - Satellite dataset I (available on their website)

To cut on data preparation and image generation time, feel free to contact me for the dataset folder and synthetic images I prepared.