we do not have a full checkpoint conversion validation, if you encounter pipeline loading failure and unsidered output, please contact me via bili_sakura@zju.edu.cn

CUT-OpenEarthMap-SAR

CUT (Contrastive Unpaired Translation) models for SAR ↔ optical image translation. Trained on OpenEarthMap data with anti-aliased ResNet generators.

Model variants

Model	Direction	Epoch
opt2sar	Optical → SAR	20
sar2opt	SAR → Optical	15
seman2opt	Semantic → Optical	25
seman2opt_pesudo	Semantic (pseudo) → Optical	195
seman2sar	Semantic → SAR	25
seman2sar_pesudo	Semantic (pseudo) → SAR	200

Usage

Use with pytorch-image-translation-models and the openearthmap_sar community pipeline:

from PIL import Image

from examples.community.openearthmap_sar import load_openearthmap_sar_pipeline

pipeline = load_openearthmap_sar_pipeline(
    checkpoint_dir="/path/to/CUT-OpenEarthMap-SAR",
    model_name="sar2opt",  # choose one: opt2sar, sar2opt, seman2opt, seman2opt_pesudo, seman2sar, seman2sar_pesudo
    device="cuda",
)

source = Image.open("/path/to/sar.png").convert("RGB")
output = pipeline(source_image=source, output_type="pil")
output.images[0].save("cut_sar2opt.png")

CLI:

python -m examples.community.openearthmap_sar \
    --checkpoint-dir BiliSakura/CUT-OpenEarthMap-SAR \
    --model sar2opt \
    --input sar.png \
    --output out.png

Pass source_image as PIL.Image. The generator uses anti-aliased down/upsampling to match the original CUT training.

Repository layout

{model_name}/
  generator/
    config.json
    diffusion_pytorch_model.safetensors

Architecture

Parameter	Value
in_channels	3
out_channels	3
base_filters	64
n_blocks	9
norm	InstanceNorm

Citation

CUT (architecture):

@inproceedings{park2020cut,
  title={Contrastive Learning for Unpaired Image-to-Image Translation},
  author={Park, Taesung and Efros, Alexei A and Zhang, Richard and Zhu, Jun-Yan},
  booktitle={ECCV},
  year={2020}
}

OpenEarthMap-SAR (dataset & baseline):

@ARTICLE{11303033,
  author={Xia, Junshi and Chen, Hongruixuan and Broni-Bediako, Clifford and Wei, Yimin and Song, Jian and Yokoya, Naoto},
  journal={IEEE Geoscience and Remote Sensing Magazine},
  title={OpenEarthMap-SAR: A benchmark synthetic aperture radar dataset for global high-resolution land cover mapping [Software and Data Sets]},
  year={2025},
  volume={13},
  number={4},
  pages={476-487},
  keywords={Translation;Semantic segmentation;Source coding;Urban planning;Land surface;Geoscience and remote sensing;Benchmark testing;Software;Sustainable development;Synthetic aperture radar},
  doi={10.1109/MGRS.2025.3599512}
}

Credits

Models trained on the OpenEarthMap-SAR benchmark dataset. Thanks to the authors for the dataset and CUT baseline.