Image Inpainting via Generative Multi-column Convolutional Neural Networks

by Yi Wang, Xin Tao, Xiaojuan Qi, Xiaoyong Shen, Jiaya Jia.

Results on Places2, CelebA-HQ, and Paris streetview with rectangle masks.

Results on Places2 and CelebA-HQ with random strokes.

Introduction

This repository is for the NeurIPS 2018 paper, 'Image Inpainting via Generative Multi-column Convolutional Neural Networks'.

If our method is useful for your research, please consider citing:

@inproceedings{wang2018image,
  title={Image Inpainting via Generative Multi-column Convolutional Neural Networks},
  author={Wang, Yi and Tao, Xin and Qi, Xiaojuan and Shen, Xiaoyong and Jia, Jiaya},
  booktitle={Advances in Neural Information Processing Systems},
  pages={331--340},
  year={2018}
}

Our framework

Partial Results

    

    

More results

Prerequisites

• Python3.5 (or higher)
• Tensorflow 1.4 (or later versions, excluding 2.x) with NVIDIA GPU or CPU
• OpenCV
• numpy
• scipy
• easydict
• Pytorch 1.0 with NVIDIA GPU or CPU
• tensorboardX

Installation

git clone https://github.com/shepnerd/inpainting_gmcnn.git
cd inpainting_gmcnn/tensorflow

or

cd inpainting_gmcnn/pytorch

For tensorflow implementations

Testing

Download pretrained models through the following links (paris_streetview, CelebA-HQ_256, CelebA-HQ_512, Places2), and unzip and put them into checkpoints/. To test images in a folder, you can specify the folder address by the opinion --dataset_path, and set the pretrained model path by --load_model_dir when calling test.py.

For example:

python test.py --dataset paris_streetview --data_file ./imgs/paris-streetview_256x256/ --load_model_dir ./checkpoints/paris-streetview_256x256_rect --random_mask 0

or

sh ./script/test.sh

Training

For a given dataset, the training is formed of two stages. We pretrain the whole network with only confidence-driven reconstruction loss first, and finetune this network using adversarial and ID-MRF loss along with the reconstruction loss after the previous phase converges.

To pretrain the network,

python train.py --dataset [DATASET_NAME] --data_file [DATASET_TRAININGFILE] --gpu_ids [NUM] --pretrain_network 1 --batch_size 16

where [DATASET_TRAININGFILE] indicates a file storing the full paths of the training images.

Then finetune the network,

python train.py --dataset [DATASET_NAME] --data_file [DATASET_TRAININGFILE] --gpu_ids [NUM] --pretrain_network 0 --load_model_dir [PRETRAINED_MODEL_PATH] --batch_size 8

We provide both random stroke and rectangle masks in the training and testing phase. The used mask type is indicated by specifying --mask_type [rect(default)|stroke] option when calling train.py or test.py.

A simple interactive inpainting GUI

A GUI written using tkinter is given in `painter_gmcnn.py`. Start it by calling ```shell sh ./script/vis_tool.sh ```

Other pretrained models

CelebA-HQ_512 trained with stroke masks.

For pytorch implementations

The testing and training procedures are similar to these in the tensorflow version except some parameters are with different names.

Testing

A pretrained model: CelebA-HQ_256.

Training

Compared with the tensorflow version, this pytorch version would expect a relatively smaller batch size for training.

Other versions

Checkout the keras implementation of our paper by Tomasz Latkowski here.

Disclaimer

• For the provided pretrained models, their performance would degrade obviously when they are evaluated by a mask whose unknown areas are too large.
• As claimed in the paper, for the large datasets with thousands of categories, the model performance is unstable. Recent GAN using large-scale techniques may ease this problem.
• We did not give the full implementation of ID-MRF (in this repo) described in our original paper. The step of excluding s is omitted for computational efficiency.
• In the pytorch version, a different GAN loss (wgan hinge loss with spectral normalization) is adopted.

Acknowledgments

Our code is partially based on Generative Image Inpainting with Contextual Attention and pix2pixHD. The implementation of id-mrf loss is borrowed from contextual loss.

Contact

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