Official PyTorch implementation of the BMVC 2022 paper "Dual Pyramid Generative Adversarial Networks for Semantic Image Synthesis ". The code allows the users to reproduce and extend the results reported in the study. Please cite the paper when reporting, reproducing or extending the results.

[Arxiv] [Poster]

This repository implements the DP_GAN model, which generates realistic looking images from semantic label maps. In addition, many different images can be generated from any given label map by simply resampling a noise vector (first two rows of the figure below). The model also allows to just resample parts of the image (see the last two rows of the figure below). Check out the paper for details, as well as the appendix, which contains many additional examples.

First, clone this repository:

git clone https://github.com/sj-li/DP_GAN.git
cd DP_GAN

The code is tested for Python 3.7.6 and the packages listed in environment.yml. The basic requirements are PyTorch and Torchvision. The easiest way to get going is to install the dp_gan conda environment via

conda env create --file environment.yml
conda activate dp_gan

For Cityscapes or ADE20K, please follow the instructions for the dataset preparation as outlined in https://github.com/NVlabs/SPADE.

Please refer to the scripts in ./scripts

Our method outperforms previous methods on almost all metrics

If you use this work please cite

@inproceedings{schonfeld_sushko_iclr2021,
  title={Dual Pyramid Generative Adversarial Networks for Semantic Image Synthesis},
  author={Shijie Li, Ming-Ming Cheng, Juergen Gall},
  booktitle={British Machine Vision Conference},
  year={2022}
}   

This code is based on https://github.com/boschresearch/OASIS

This project is open-sourced under the AGPL-3.0 license. See the LICENSE file for details.

For a list of other open source components included in this project, see the file 3rd-party-licenses.txt.