Pixel to Gaussian: Ultra-Fast Continuous Super-Resolution with 2D Gaussian Modeling

🎉 Accepted by ICLR 2026 🎉

Pixel to Gaussian: Ultra-Fast Continuous Super-Resolution with 2D Gaussian Modeling

Long Peng1,3†, Anran Wu1,2†, Wenbo Li3*, Peizhe Xia1, Xueyuan Dai4, Xinjie Zhang5, Xin Di1, Haoze Sun6, Renjing Pei3, Yang Wang1,4*, Yang Cao1, Zheng-Jun Zha1. †Equal Contribution. *Corresponding Authors.

1USTC, 2AHU, 3Huawei Noah’s Ark Lab, 4Chang’an University, 5HKUST, 6THU.

[Arxiv Paper]  [Website Page]  [Google Drive] 

:rocket: :rocket: :rocket: News

• [x] Our paper has been accepted by ICLR 2026!
• [x] Arxiv Version has been released.
• [x] Test code has been released.
• [x] The pretrained model has been released.
• [x] Since this paper is completed within the company, the related code and pre-trained models cannot be released. To support the open-source community, we voluntarily and freely attempt to reproduce it in this repository. The reproduced results may have very slight deviations, but the difference stays within 0.05 dB.

TODO

• [ ] Release train code.

💡 Summary

We introduce ContinuousSR, a novel Pixel-to-Gaussian paradigm designed for fast and high-quality arbitrary-scale super-resolution. By explicitly reconstructing 2D continuous HR signals from LR images using Gaussian Splatting, ContinuousSR significantly improves both efficiency and performance. Through statistical analysis, we uncover the Deep Gaussian Prior (DGP) and propose a DGP-driven Covariance Weighting mechanism along with an Adaptive Position Drifting strategy. These innovations improve the quality and fidelity of the reconstructed Gaussian fields. Experiments on seven popular benchmarks demonstrate that our method outperforms state-of-the-art methods in both quality and speed, achieving a {19.5×} speed improvement and {0.90dB} PSNR improvement, making it a promising solution for ASSR tasks.

💡 Motivation and Framework

  Compared to other methods, the proposed ContinuousSR delivers significant improvements in SR quality across all scales, with an impressive 19.5× speedup when continuously upsampling an image across forty scales.

📃 Dependencies and Installation

• python=3.9
• pytorch=1.13
• basicsr==1.3.4.9
• Others:

git clone https://github.com/XingtongGe/gsplat.git
cd gsplat
pip install -e .[dev]

Get Started

Pretrained model

• After our thorough reproduction process, we are excited to open-source the pre-trained model!
• The best-performing version of our method can be downloaded here: ➡️ Pretrained Model
• Once downloaded, place the model in the designated folder, and you’ll be ready to run the demo and perform inference. 🚀

Demo

Here is an Demo

# scale represents the magnification factors for height and width respectively
python demo.py --input butterflyx4.png --model ContinuousSR.pth --scale 4,4 --output output.png

Inference

Here is an example command for inference

# test Set5 X4
python test.py --config ./configs/test/test-set5-4.yaml --model ContinuousSR.pth

Visual Examples

✉️ License

Licensed under a Creative Commons Attribution-NonCommercial 4.0 International for Non-commercial use only. Any commercial use should get formal permission first.

✉️ Acknowledgement

This repository is maintained by Long Peng and Anran Wu.

Citation

If you are interested in the following work, please cite the following paper.

@article{peng2025pixel,
  title={Pixel to Gaussian: Ultra-Fast Continuous Super-Resolution with 2D Gaussian Modeling},
  author={Peng, Long and Wu, Anran and Li, Wenbo and Xia, Peizhe and Dai, Xueyuan and Zhang, Xinjie and Di, Xin and Sun, Haoze and Pei, Renjing and Wang, Yang and others},
  journal={arXiv preprint arXiv:2503.06617},
  year={2025}
}