Pik-Fix: Restoring and Colorizing Old Photo

The official code implementation of WACV2023 paper "Pik-Fix: Restoring and Colorizing Old Photos".

Real Old Photo Data

Download

Due to copyright restrictions, our real old photo dataset is only available upon personal inquiry. To request access, please email derrickxu1994@gmail.com with the subject "Pik-Fix Data Inquiry" and include your name and affiliation in the message. Then we will reply to you with the download link in 3 days.

File Structure

There will be three folders in the shared data link: real_old_data, real_old_ref, and texture2. The first one contains 200 image pair with old photos and the repared ones. The second folder contains the reference images we used in our paper. The third folder contains the necessary texture files you will need to generate fake old photos. These files should be orgnaized as follows:

- Pik-Fix
    - data
        - real_old_data
        - real_old_ref
        - texture2
    - models
    - datasets
    - utils
    - ...

Installation

conda create -n pikfix python=3.7
conda activate pikfix
conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch
pip install -r requirements.txt
python setup.py develop

Synthetic Data Generation

There are two ways To generate synthetic old photos from good quality images, please first change the configuration file in hypes_yaml/data_generation.yaml to modify your data path, output path, and generation configurations. Next, run the following command:

python utils/data_generation.py --hypes_yaml hypes_yaml/data_generation.yaml