[ICCV 2025] PAN-Crafter: Learning Modality-Consistent Alignment for PAN-Sharpening

Jeonghyeok Do¹     Sungpyo Kim¹     Geunhyuk Youk¹     Jaehyup Lee^2†     Munchurl Kim^1†

^†Co-corresponding authors

¹Korea Advanced Institute of Science and Technology, South Korea

²Kyungpook National University, South Korea

---

This repository is the official PyTorch implementation of "PAN-Crafter: Learning Modality-Consistent Alignment for PAN-Sharpening". PAN-Crafter achieves state-of-the-art results on multiple datasets, outperforming the recent PAN-Sharpening methods.

Network Architecture

---

📧 News

• Jul 26, 2025: Youtube video about PAN-Crafter is uploaded ✨
• Jul 18, 2025: Codes of PAN-Crafter are released 🔥
• Jun 26, 2025: PAN-Crafter accepted to ICCV 2025 🎉
• May 30, 2025: This repository is created

---

Reference

@inproceedings{do2025pancrafter,
  title={PAN-Crafter: Learning Modality-Consistent Alignment for PAN-Sharpening},
  author={Do, Jeonghyeok, Kim, Sungpyo, Youk, Geunhyuk, Lee, Jaehyup, and Kim, Munchurl},
  booktitle={Proceedings of the IEEE/CVF international conference on computer vision},
  year={2025}
}

---

Contents

• Requirements
• Data Preparation
• Training
• Results
• License
• Acknowledgement

⚙️ Requirements

• Python >= 3.9.19
• PyTorch >= 2.4.0
• Platforms: Ubuntu 22.04, CUDA 11.8
• We have included a dependency file for our experimental environment. To install all dependencies, create a new Anaconda virtual environment and execute the provided file. Run conda env create -f requirements.yaml.

📁 Data Preparation

Pancollection Dataset

We follow the evaluation setup from PanCollection.

Download the datasets here and arrange them as follows:

Pancollection
    ├── WV3
    │   ├── train_wv3.h5
    │   ├── valid_wv3.h5
    │   ├── train_wv3_pan.h5
    │   ├── valid_wv3_pan.h5
    │   ├── reduced_examples_h5
    │   │   ├── test_wv3_multiExm1.h5
    │   │   └── test_wv3_multiExm1_pan.h5
    │   └─── full_examples_h5
    │       ├── test_wv3_OrigScale_multiExm1.h5
    │       └── test_wv3_OrigScale_multiExm1_pan.h5
    │
    ├── QB
    │   ├── train_qb.h5
    │   ├── valid_qb.h5
    │   ├── train_qb_pan.h5
    │   ├── valid_qb_pan.h5
    │   ├── reduced_examples_h5
    │   │   ├── test_qb_multiExm1.h5
    │   │   └── test_qb_multiExm1_pan.h5
    │   └─── full_examples_h5
    │       ├── test_qb_OrigScale_multiExm1.h5
    │       └── test_qb_OrigScale_multiExm1_pan.h5
    │
    └── GF2
        ├── train_gf2.h5
        ├── valid_gf2.h5
        ├── train_gf2_pan.h5
        ├── valid_gf2_pan.h5
        ├── reduced_examples_h5
        │   ├── test_gf2_multiExm1.h5
        │   └── test_gf2_multiExm1_pan.h5
        └─── full_examples_h5
            ├── test_gf2_OrigScale_multiExm1.h5
            └── test_gf2_OrigScale_multiExm1_pan.h5

Note: Files with _pan.h5 (e.g., train_wv3_pan.h5, test_qb_multiExm1_pan.h5) contain the panchromatic image (I_{pan}^{lr}),
which has been down-sampled by a factor of 4 to match the spatial resolution of the multispectral image (I_{ms}^{lr}).

Training

# Download code
git clone https://github.com/KAIST-VICLab/PAN-Crafter
cd PAN-Crafter

# Train PAN-Crafter on WV3
python main.py --config ./config/pancrafter_wv3.yaml

# Train PAN-Crafter on QB
python main.py --config ./config/pancrafter_qb.yaml

# Train PAN-Crafter on GF2
python main.py --config ./config/pancrafter_gf2.yaml

Results

Please visit our project page for more experimental results.

All evaluation metrics were measured using the official MATLAB code from the DLPan-Toolbox.

License

The source codes can be freely used for research and education only. Any commercial use should get formal permission from the principal investigator (Prof. Munchurl Kim, [email protected]).

Acknowledgements

This work was supported by National Research Foundation of Korea (NRF) grant funded by the Korean Government [Ministry of Science and ICT (Information and Communications Technology)] (Project Number: RS- 2024-00338513, Project Title: AI-based Computer Vision Study for Satellite Image Processing and Analysis, 100%).

This repository is built upon SkateFormer and U-Know-DiffPAN.