ScaleFormer

This folder contains the implementation of the paper Cross-Scale Pansharpening via ScaleFormer and the PanScale Datasets (CVPR 2026).

Environment Setup

Recommended: Python 3.10+ with CUDA-enabled PyTorch.

Install a minimal dependency set:

pip install torch torchvision numpy pyyaml pillow opencv-python tifffile scipy tqdm tensorboardX einops rotary-embedding-torch matplotlib

Dataset Format

PanScale is publicly available on Hugging Face:

• https://huggingface.co/datasets/kecao/PanScale

Download example:

pip install -U huggingface_hub
huggingface-cli download kecao/PanScale --repo-type dataset --local-dir ./datasets/hddataset

Configuration

Edit one of the YAML files before running (for example options/option_jilin.yml). All option files are now organized under the options/ folder:

• options/option_jilin.yml

• options/option_landsat.yml

• options/option_skysat.yml

• options/option_full_jilin.yml

• options/option_full_landsat.yml

• options/option_full_skysat.yml

• algorithm: model name (must match a python file under model/)

• checkpoint: root directory for checkpoints

• log_dir: training log directory

• data_dir_train: training split root (contains ms/ and pan/)

• data_dir_eval: validation/test split root (contains ms/ and pan/)

• test.data_dir: dataset used by test scripts

• test.model: relative checkpoint path under checkpoint

• test.save_dir: output directory for fused images

Scale settings used in this project:

• Jilin: upsacle: 4
• Landsat: upsacle: 2
• SkySat: upsacle: 2.5

Training

Examples:

python runners/main_jilin.py --option_path options/option_jilin.yml
python runners/main_landsat.py --option_path options/option_landsat.yml
python runners/main_skysat.py --option_path options/option_skysat.yml

The training loop saves:

• latest.pth
• bestPSNR.pth
• bestSSIM.pth
• additional best checkpoints for no-reference metrics

under:

<checkpoint>/<algorithm>_<upsacle>_<timestamp>/

Inference (Reduced Resolution)

python runners/test_jilin.py
python runners/test_landsat.py
python runners/test_skysat.py

Inference (Full Resolution / Large Images)

python runners/test_full_jilin.py
python runners/test_full_landsat.py
python runners/test_full_skysat.py

Citation

@article{cao2026cross,
  title={Cross-Scale Pansharpening via ScaleFormer and the PanScale Benchmark},
  author={Cao, Ke and He, Xuanhua and Li, Xueheng and Zhu, Lingting and Wang, Yingying and Ma, Ao and Zhang, Zhanjie and Zhou, Man and Xie, Chengjun and Zhang, Jie},
  journal={arXiv preprint arXiv:2603.00543},
  year={2026}
}