Home

Awesome

Learning Dual-Level Implicit Representation for Real-World Scale Arbitrary Super-Resolution

1. RealArbiSR Dataset Preparation

Version 2

In version 2, we further refine the dataset quality and increase the size of x1.7/x2.3/x2.7/x3.3/x3.7 testset from 83 scenes to 100 scenes.

Dataset Version 2 is available at RealArbiSRdatasetv2 - Google Drive

The pretrained models and the PSNR results of RealArbiSR dataset Version 2 are listed below:

EDSR-DDIR-v2

RDN-DDIR-v2

MethodsPSNRx1.5x2.0x2.5x3.0x3.5x4.0
Bicubic34.8731.6129.8128.5627.6427.00
EDSR-LIIF36.5533.6331.7630.4929.4728.80
EDSR-LTE36.5633.6331.7530.4829.5228.84
EDSR-CiaoSR36.6733.8432.0130.7429.7529.01
EDSR-DDIR36.9134.0932.2030.9429.9429.19
RDN-LIIF36.6433.8431.9430.6929.6929.00
RDN-LTE36.6033.8031.9530.6729.7029.00
RDN-CiaoSR36.8534.0732.1830.8729.8629.10
RDN-DDIR37.0434.2832.3531.0530.0429.26
MethodsPSNRx1.7x2.3x2.7x3.3x3.7
Bicubic31.3128.5427.5126.4225.83
EDSR-LIIF33.3730.5729.3728.0227.32
EDSR-LTE33.4730.6429.4028.0227.30
EDSR-CiaoSR33.0430.5829.5028.2327.48
EDSR-DDIR33.7130.9729.7628.3727.64
RDN-LIIF33.4930.7129.5128.1627.43
RDN-LTE33.5430.8329.6128.2327.51
RDN-CiaoSR33.1630.8129.7428.4427.69
RDN-DDIR33.7731.0629.8528.4627.72

Version 1 (used in the original paper)

Dataset is available at RealArbiSRdataset - Google Drive.

Arrange dataset into the path like load/Train/... and load/Test/...

2. DDIR Code

Train

python train_realliif_deform.py --gpu [GPU] --config [CONFIG_NAME] --save_name [SAVE_NAME]

Test on Pretrained Models

The pretrained models (for Verision 1, used in the original paper) can be downloaded from the google drive links below:

EDSR-DDIR

RDN-DDIR

To test at all scale factors:

bash ./scripts/test-realsrarbi-deform.sh [MODEL_PATH] [GPU]

Citation

If you find this code useful in your work then please cite:

@article{li2024learning,
  title={Learning Dual-Level Deformable Implicit Representation for Real-World Scale Arbitrary Super-Resolution},
  author={Li, Zhiheng and Li, Muheng and Fan, Jixuan and Chen, Lei and Tang, Yansong and Zhou, Jie and Lu, Jiwen},
  journal={arXiv preprint arXiv:2403.10925},
  year={2024}
}

Contact

Please contact Zhiheng Li @ lizhihan21@mails.tsinghua.edu.cn if any issue.

Acknowledgements

This code is built on LIIF. We thank the authors for sharing their codes.