LPNSR: Optimal Noise-Guided Diffusion Image Super-Resolution Via Learnable Noise Prediction

This repository contains the official model weights for LPNSR, a prior-enhanced efficient diffusion framework for image super-resolution (SR).

• Paper: LPNSR: Optimal Noise-Guided Diffusion Image Super-Resolution Via Learnable Noise Prediction
• GitHub Repository: Faze-Hsw/LPNSR

Introduction

Diffusion-based image super-resolution (SR) often suffers from inconsistent performance due to random noise injection, especially when using a limited number of sampling steps. To address this, LPNSR introduces a theoretical framework for optimal noise injection.

Key Innovations:

• Optimal Noise Framework: Derives a closed-form analytical solution for optimal intermediate noise based on maximum likelihood estimation.
• Learnable Noise Predictor: Replaces standard random Gaussian noise with an LR-guided multi-input-aware noise predictor.
• Reduced Initialization Bias: Uses a high-quality pre-upsampling network to improve starting conditions.
• Efficient 4-Step Trajectory: A compact trajectory uniquely enables end-to-end optimization of the entire reverse chain, achieving state-of-the-art perceptual quality on synthetic and real-world datasets without large-scale text-to-image priors.

Quick Start

To use these weights, clone the official repository and follow the environment setup instructions.

Inference

Once the environment is set up and weights are placed in the pretrained/ folder, run:

python LPNSR/inference.py -i [image folder/image path] -o [output folder]

Online Demo

You can also launch a local Gradio web interface:

python LPNSR/app.py

Citation

If you find this work useful, please cite:

@article{lpnsr2026,
  title={LPNSR: Optimal Noise-Guided Diffusion Image Super-Resolution Via Learnable 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.