CoRE-UIR
Official pretrained models for CoRE-UIR: Prior-Guided Common and Residual Experts for Efficient All-in-One Remote Sensing Image Restoration.
[Paper] [Code] [MDVD-108K] [Project Page]
Model Description
CoRE-UIR is a prior-guided global-local framework for restoring diverse single and compound degradations in remote sensing imagery. Its primary structural innovation, the Common-and-Residual Expert (CoRE) block, combines a common dense expert for degradation-invariant restoration with sparsely routed low-rank residual experts for degradation-specific compensation.
Available Checkpoints
| Benchmark | DPE adapter | Restoration model | Routing |
|---|---|---|---|
| MDVD-108K | dpe_clip_b32_mdvd.ckpt | core_uir_mdvd.pth | 6 experts, Top-3 |
| MDRS-Landsat | dpe_clip_b32_mdrs.ckpt | core_uir_mdrs.pth | 4 experts, Top-1 |
The DPE files contain the trained adapter and classification head. The frozen openai/clip-vit-base-patch32 backbone is loaded separately from Hugging Face.
Usage
Install the official implementation:
git clone https://github.com/zzaiyan/CoRE-UIR.git
cd CoRE-UIR
pip install -r requirements.txt
Evaluate on MDVD-108K:
python test.py \
-opt configs/core_uir_mdvd.yml \
--weights hf://zzaiyan/CoRE-UIR/core_uir_mdvd.pth \
--gpu 0 \
--no-resume
Evaluate on MDRS-Landsat:
python test.py \
-opt configs/core_uir_mdrs.yml \
--weights hf://zzaiyan/CoRE-UIR/core_uir_mdrs.pth \
--gpu 0 \
--no-resume
The public configurations download the matching DPE adapter automatically and reuse files from the local Hugging Face cache.
Citation
@article{zhang2026coreuir,
title = {{CoRE-UIR: Prior-guided common and residual experts for efficient all-in-one remote sensing image restoration}},
author = {Zhang, Zaiyan and Yuan, Qiangqiang and Li, Jie and Lihe, Ziyang and Wan, Yu and Chen, Yuzeng and Su, Xin and Zhang, Liangpei},
journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
volume = {240},
pages = {410--429},
year = {2026},
publisher = {Elsevier},
doi = {10.1016/j.isprsjprs.2026.07.028}
}