Learning Dual-Level Implicit Representation for Real-World Scale Arbitrary Super-Resolution [ECCV24]

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

Methods	PSNR	x1.5	x2.0	x2.5	x3.0	x3.5	x4.0
Bicubic		34.87	31.61	29.81	28.56	27.64	27.00
EDSR-LIIF		36.55	33.63	31.76	30.49	29.47	28.80
EDSR-LTE		36.56	33.63	31.75	30.48	29.52	28.84
EDSR-CiaoSR		36.67	33.84	32.01	30.74	29.75	29.01
EDSR-DDIR		36.91	34.09	32.20	30.94	29.94	29.19
RDN-LIIF		36.64	33.84	31.94	30.69	29.69	29.00
RDN-LTE		36.60	33.80	31.95	30.67	29.70	29.00
RDN-CiaoSR		36.85	34.07	32.18	30.87	29.86	29.10
RDN-DDIR		37.04	34.28	32.35	31.05	30.04	29.26

Methods	PSNR	x1.7	x2.3	x2.7	x3.3	x3.7
Bicubic		31.31	28.54	27.51	26.42	25.83
EDSR-LIIF		33.37	30.57	29.37	28.02	27.32
EDSR-LTE		33.47	30.64	29.40	28.02	27.30
EDSR-CiaoSR		33.04	30.58	29.50	28.23	27.48
EDSR-DDIR		33.71	30.97	29.76	28.37	27.64
RDN-LIIF		33.49	30.71	29.51	28.16	27.43
RDN-LTE		33.54	30.83	29.61	28.23	27.51
RDN-CiaoSR		33.16	30.81	29.74	28.44	27.69
RDN-DDIR		33.77	31.06	29.85	28.46	27.72

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:

@inproceedings{li2025learning,
  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 Lu, Jiwen and Zhou, Jie},
  booktitle={European Conference on Computer Vision},
  pages={352--368},
  year={2025},
  organization={Springer}
}

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.