Home

Awesome

SegVAE

| Project Page | Conference Paper | ArXiv |

Controllable Image Synthesis via SegVAE.
Yen-Chi Cheng, Hsin-Ying Lee, Min Sun, and Ming-Hsuan Yang.
In European Conference on Computer Vision (ECCV), 2020.

PyTorch implementation for our SegVAE. With the proposed VAE-based framework, we are able to learn how to generate diverse and plausible semantic maps given a label-set. This provides flexible user editing for image synthesis.

Usage

Prerequisites

Installation

git clone https://github.com/yccyenchicheng/SegVAE.git
cd SegVAE

Dataset

We experimented on two datasets: CelebAMask-HQ and HumanParsing. You could download the datasets following their instructions.

Or you can download the dataset we used from this link: https://drive.google.com/drive/folders/1ah26mxO3rFLTMcVbgldEzIj28Poq1lf8?usp=sharing.

Create a folder named data/segvae/ and put the downloaded .zip files under data/segvae/, and unzip them:

mkdir -p data/segvae
mv ~/Downloads/celebamaskhq.zip data/segvae
mv ~/Downloads/humanparsing.zip data/segvae
cd data/segvae
unzip celebamaskhq.zip
unzip humanparsing.zip
cd ../../

Training

To train the model, run

python train.py --batch_size [batch_size] --dataset [humanparsing or celebamaskhq]

For example,

python train.py --batch_size 16 --dataset celebamaskhq

The log files will be written into logs/segvae_logs.

Then you could run

tensorboard --logdir logs/segvae_logs --port 6006

and go to http://127.0.0.1:6006 to see the visualization of training logs in the browser.

Citation

Please cite our paper if you find the code, or paper useful for your research.

Controllable Image Synthesis via SegVAE
Yen-Chi Cheng, Hsin-Ying Lee, Min Sun, and Ming-Hsuan Yang
European Conference on Computer Vision (ECCV), 2020

@inproceedings{cheng2020segvae,
    title={Controllable Image Synthesis via SegVAE},
    author={Cheng, Yen-Chi and Lee, Hsin-Ying and Sun, Min and Yang, Ming-Hsuan},
    booktitle = {European Conference on Computer Vision},
    year={2020},
}

Acknowledgement

This code borrows heavily from SPADE. We also thank COCO-GAN, BigGAN-PyTorch for the FID calculation and Spectral Norm implementation.