[📄 arXiv 2511.05009] [💻 GitHub Project] [🍍知乎]
Nov. 2025 🔥 New SOTA! UHDRes achieves state-of-the-art performance across five UHD restoration tasks (4K & 8K)
with only 0.40M parameters — outperforming previous methods up to +2.3dB PSNR while using 10× fewer parameters.
We propose UHDRes, a UHD image restoration framework solely based on the frequency domain and large-kernel convolution. Our method achieves state-of-the-art performance while maintaining high computational efficiency.
- CUDA 11.8 (or later)
- PyTorch 2.0.0 (or later)
# git clone this repository
git clone https://github.com/Zhao0100/UHDRes.git
cd UHDRes
# create new anaconda env
conda create -n uhdres python=3.9
source activate uhdres
# install PyTorch
pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txtUHD-LL, UHD-Haze, 8KDehaze-mini, UHDBlur, 4K-Rain13k
We provide pretrained models for UHD-LL, UHD-Haze, UHD-Blur, and 4K-Rain13k.
| Task | Model |
|---|---|
| UHD-LL | model |
| UHD-Haze | model |
| 8KDehaze-mini | model |
| UHD-Blur | model |
| 4K-Rain13k | model |
bash train.shbash test.sh
@ARTICLE{UHDRes,
title={UHDRes: Ultra-High-Definition Image Restoration via Dual-Domain Decoupled Spectral Modulation},
author={S. Zhao and W. Lu and B. Wang and T. Wang and K. Zhang and H. Zhao},
year={2025},
eprint={2511.05009},
archivePrefix={arXiv},
primaryClass={eess.IV},
url={https://arxiv.org/abs/2511.05009},
}