by Yi Wang, Lu Qi, Ying-Cong Chen, Xiangyu Zhang, Jiaya Jia.

This repository gives the implementation of our semantic image synthesis method in ICCV 2021 paper, 'Image synthesis via semantic synthesis'.

git clone https://github.com/dvlab-research/SCGAN.git
cd SCGAN/code

To use this code, please install PyTorch 1.0 and Python 3+. Other dependencies can be installed by

pip install -r requirements.txt

Please refer to SPADE for detailed execution.

1. Downloading pretrained models, then putting the folder containing model weights in the folder ./checkpoints.

2. Producing images with the pretrained models.

python test.py --gpu_ids 0,1,2,3 --dataset_mode [dataset] --config config/scgan_[dataset]_test.yml --fid --gt [gt_path] --visual_n 1

For example,

python test.py --gpu_ids 0,1,2,3 --dataset_mode celeba --config config/scgan_celeba-test.yml --fid --gt /data/datasets/celeba --visual_n 1

3. Visual results are stored at ./results/scgan_[dataset]/ by default.

Using train.sh to train new models. Or you can specify training options in config/[config_file].yml.

Our proposed dynamic computation units (spatial conditional convolution and normalization) are extended from conditionally parameterized convolutions [1]. We generalize the scalar condition into a spatial one and also apply these techniques to normalization.

If our research is useful for you, please consider citing:

@inproceedings{wang2021image,
  title={Image Synthesis via Semantic Composition},
  author={Wang, Yi and Qi, Lu and Chen, Ying-Cong and Zhang, Xiangyu and Jia, Jiaya},
  booktitle={ICCV},
  year={2021}
}

This code is built upon SPADE, Imaginaire, and PyTorch-FID.

[1] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. In NeurIPS. 2019.

Please send email to yiwang@cse.cuhk.edu.hk.