LPNSR: Prior-Enhanced Diffusion Image Super-Resolution Dataset

This repository contains the evaluation datasets and testing data associated with the paper LPNSR: Prior-Enhanced Diffusion Image Super-Resolution via LR-Guided Noise Prediction.

Project Links

• Paper: arXiv:2603.21045
• GitHub Repository: Faze-Hsw/LPNSR

Dataset Description

This dataset collection is used to evaluate image super-resolution models on both synthetic and complex real-world degradations. It contains pairs of Low-Quality (LQ) and Ground-Truth (GT) high-resolution images.

The LPNSR approach utilizes an LR-guided multi-input-aware noise predictor instead of random Gaussian noise for partial diffusion initialization, allowing for efficient 4-step inference.

Sub-Datasets Included:

• imagenet512: Contains 3,000 synthetic image pairs used for validation/testing.
• RealSR: Contains 100 image pairs featuring real-world degradations captured from actual camera sensors.
• RealSet80: Contains 80 real-world highly degraded images without Ground-Truth references.

How to Use

Load the paired super-resolution datasets (imagenet512, RealSR) using the Hugging Face datasets library:

from datasets import load_dataset

# Load the imagenet512 subset
dataset_imagenet = load_dataset("mirpri/LPNSR-dataset", name="imagenet512")

# Load the RealSR subset
dataset_realsr = load_dataset("mirpri/LPNSR-dataset", name="RealSR")

# Check the properties of the first pair
print(dataset_imagenet['train'][0])
# Keys will map to 'image' (for LQ) and 'ground_truth' (for GT). 

Citation

If you find this dataset or the LPNSR framework useful, please cite our paper:

@article{lpnsr2026,
  title={LPNSR: Prior-Enhanced Diffusion Image Super-Resolution via LR-Guided Noise Prediction},
  author={Huang, Shuwei and Liu, Shizhuo and Wei, Zijun},
  journal={arXiv preprint arXiv:2603.21045},
  year={2026},
  eprint={2603.21045},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

Acknowledgement

This project is based on ResShift, BasicSR, SwinIR, and Real-ESRGAN.