ESRGAN-pytorch

This repository implements a deep-running model for super resolution. Super resolution allows you to pass low resolution images to CNN and restore them to high resolution. We refer to the following article.
ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks

architecture

[Overall Architecture]
[Basic block]

Test Code

python test.py --lr_dir LR_DIR --sr_dir SR_DIR

Prepare dataset

Use Flicker2K and DIV2K

cd datasets
python prepare_datasets.py
cd ..

custom dataset

Make dataset like this; size of hr is 128x128 ans lr is 32x32

datasets/
    hr/
        0001.png
        sdf.png
        0002.png
        0003.png
        0004.png
        ...
    lr/
        0001.png
        sdf.png
        0002.png
        0003.png
        0004.png
        ...

how to train

run main file

python main.py --is_perceptual_oriented True --num_epoch=10
python main.py --is_perceptual_oriented False --epoch=10

Sample

we are in training on this code and train is not complete yet. this is intermediate result.