This repository is an official PyTorch implementation of the paper "Toward Real-World Super-Resolution via Adaptive Downsampling Models" which is accepted at TPAMI(link).

• Python 3.7
• PyTorch >= 1.6.0
• matplotlib
• imageio
• pyyaml
• scipy
• numpy
• tqdm
• PIL

In this project, we learn the adaptive downsampling model(ADM) with an unpaired dataset consisting of HR and LR images but not pixel aligned.

As HR is not responsible to pixel aligned with LR, we recommend to use DIV2K dataset as HR dataset, which is consist of clean and high-resolution images. You can download it from here (7.1GB). We note that experiments conducted on our paper used 400 HR images('0001.png'-'0400.png') as HR.

For LR datasets, you should put a bunch of images that undergo a similar downsampling process. (e.g. DPED, RealSR, or you own images from same camera setting)

Please put HR and LR datasets in datasets/. Again, different lengths between each dataset are acceptable, as we noted that two datasets are not responsible for pixel-aligned. However, we also note that total iterations of 1 epoch can differ along with your dataset size. Since our ADM learns average downsampling kernel along with LR datasets, please use available LR images with scene variety as much as possible for stable training(we recommend to use more than 200 for HD scale images). For more details on the effect of the number of training samples, please refer to our paper.

Let denote filename of HR dataset as Source, and LR dataset as Target.

Our ADM will retrieve downsampling between two datasets Source and Target, and generate downsampled version of Source with retrieved downsampling kernel. Basic usage of training is following:

cd src
CUDA_VISIBLE_DEVICES=0 python train.py --source Source --target Target --name save_name --make_down

Generated downsampled version of Source will be saved at ./experiments/save_name/down_results/. Note that you can use Source and generated downsampled version of Source as paired dataset in conventional SR settings.

Here we additionally support joint training with SR network, which use intermediate generated image as paired dataset hence does not require additional SR network training step. Usage of joint training is following:

cd src
CUDA_VISIBLE_DEVICES=0 python train.py --source Source --target Target --name save_name --joint

Default SR model is 'EDSR-baseline', but you can change with --sr_model.

If you have validation datasets for evaluating perforamce of SR, please locate them in datasets/. After then, you can measure performance of SR network by measuring PSNR on validation datasets. You can specify it with --test_lr filename_lr --test_hr filename_hr, where filename_hr and filename_lr should be paired images.

In case that you don't have validation paired datasets as common in real-world, you can visualize SR results by --test_lr filename_lr --realsr --save_results, and then SR results will be saved in ./experiments/save_name/sr_results/. Note that filename_lr can be same with Target.

You can check detailed usage of this repo in demo.sh .

Please note that experimental results reported in our paper are conducted in separate manner(i.e., we only generate downsampled images and train SR network with corresponding official implementation), so the results in joint training may slightly differs with number in the paper.

@ARTICLE{9521710,
    author={Son, Sanghyun and Kim, Jaeha and Lai, Wei-Sheng and Yang, Ming-Hsuan and Lee, Kyoung Mu},
    journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
    title={Toward Real-World Super-Resolution via Adaptive Downsampling Models}, 
    year={2021},
    volume={},
    number={},
    pages={1-1},
    doi={10.1109/TPAMI.2021.3106790}
 }

If you have any question, please email [email protected].