--- license: other license_name: per-method-see-below library_name: pytorch pipeline_tag: image-to-image tags: - sar-to-eo - image-translation - remote-sensing - baselines - benchmark - geoset --- # GeoSET: retrained comparison methods The sixteen comparison methods of Tables 2 and 3 of the GeoSET paper ([arXiv:2609.37496](https://arxiv.org/abs/2609.37496)), retrained by the GeoSET authors on the training split of each public benchmark: 61 checkpoints, one folder per dataset and method. Seg-CycleGAN is evaluated on SpaceNet6 only. These are not the original authors' checkpoints. Training and inference code, pinned upstream commits and patches: [`baselines/`](https://github.com/KAIST-VICLab/GeoSET/tree/main/baselines) in the GeoSET repository. Metrics of every checkpoint: [MODEL_ZOO.md](https://github.com/KAIST-VICLab/GeoSET/blob/main/MODEL_ZOO.md). ## Layout `baselines///` holds the weight files and a `README.md` with the file roles, SHA-256 digests, the paper's metrics for that checkpoint, the inference commands and its licence. | Method | Venue | qxs-saropt | sar2opt | sar2eo | spacenet6 | Weights licence | |---|---|---|---|---|---|---| | pix2pix | CVPR'17 | [folder](qxs-saropt/pix2pix) | [folder](sar2opt/pix2pix) | [folder](sar2eo/pix2pix) | [folder](spacenet6/pix2pix) | BSD | | CycleGAN | ICCV'17 | [folder](qxs-saropt/cyclegan) | [folder](sar2opt/cyclegan) | [folder](sar2eo/cyclegan) | [folder](spacenet6/cyclegan) | BSD | | pix2pixHD | CVPR'18 | [folder](qxs-saropt/pix2pixhd) | [folder](sar2opt/pix2pixhd) | [folder](sar2eo/pix2pixhd) | [folder](spacenet6/pix2pixhd) | BSD | | SPADE | CVPR'19 | [folder](qxs-saropt/spade) | [folder](sar2opt/spade) | [folder](sar2eo/spade) | [folder](spacenet6/spade) | CC BY-NC-SA 4.0 | | DDPM (SR3) | TPAMI'22 | [folder](qxs-saropt/ddpm-sr3) | [folder](sar2opt/ddpm-sr3) | [folder](sar2eo/ddpm-sr3) | [folder](spacenet6/ddpm-sr3) | no upstream licence | | SD2.1 FT | CVPR'22 | [folder](qxs-saropt/sd21-ft) | [folder](sar2opt/sd21-ft) | [folder](sar2eo/sd21-ft) | [folder](spacenet6/sd21-ft) | CreativeML Open RAIL++-M | | BBDM | CVPR'23 | [folder](qxs-saropt/bbdm) | [folder](sar2opt/bbdm) | [folder](sar2eo/bbdm) | [folder](spacenet6/bbdm) | MIT | | ControlNet | ICCV'23 | [folder](qxs-saropt/controlnet) | [folder](sar2opt/controlnet) | [folder](sar2eo/controlnet) | [folder](spacenet6/controlnet) | CreativeML Open RAIL++-M | | HI-Diff | NeurIPS'23 | [folder](qxs-saropt/hi-diff) | [folder](sar2opt/hi-diff) | [folder](sar2eo/hi-diff) | [folder](spacenet6/hi-diff) | Apache-2.0 | | ResShift | NeurIPS'23 | [folder](qxs-saropt/resshift) | [folder](sar2opt/resshift) | [folder](sar2eo/resshift) | [folder](spacenet6/resshift) | S-Lab License 1.0 | | StegoGAN | CVPR'24 | [folder](qxs-saropt/stegogan) | [folder](sar2opt/stegogan) | [folder](sar2eo/stegogan) | [folder](spacenet6/stegogan) | no upstream licence | | CondDiff | GRSL'23 | [folder](qxs-saropt/conddiff) | [folder](sar2opt/conddiff) | [folder](sar2eo/conddiff) | [folder](spacenet6/conddiff) | no upstream licence | | E3Diff | GRSL'24 | [folder](qxs-saropt/e3diff) | [folder](sar2opt/e3diff) | [folder](sar2eo/e3diff) | [folder](spacenet6/e3diff) | no upstream licence | | cBBDM | GRSL'25 | [folder](qxs-saropt/cbbdm) | [folder](sar2opt/cbbdm) | [folder](sar2eo/cbbdm) | [folder](spacenet6/cbbdm) | MIT | | Seg-CycleGAN | GRSL'25 | – | – | – | [folder](spacenet6/seg-cyclegan) | no upstream licence | | C-DiffSET | TCSVT'26 | [folder](qxs-saropt/c-diffset) | [folder](sar2opt/c-diffset) | [folder](sar2eo/c-diffset) | [folder](spacenet6/c-diffset) | CreativeML Open RAIL++-M | Datasets: QXS-SAROPT (3,999 test pairs, 256 × 256), SAR2Opt (627 test pairs, center 512 × 512 crop), SAR2EO (first 4,000 of the 21,260 test pairs, 256 × 256) and SpaceNet6 (495 test pairs, 256 × 256), with the split lists of the GeoSET repository. ## Download ```bash hf download JeonghyeokDo/GeoSET --include "baselines/sar2opt/e3diff/*" --local-dir hf ``` Copy the files to `$CKPT_ROOT///` and run `baselines//test.sh ` of the GeoSET repository (see the cell README). ## Licences Each checkpoint carries the terms of the code it was trained with. The texts are in [`licenses/`](licenses): | Methods | Terms | Text | |---|---|---| | pix2pix, CycleGAN | BSD (junyanz/pytorch-CycleGAN-and-pix2pix) | [LICENSE-pix2pix.txt](licenses/LICENSE-pix2pix.txt), [LICENSE-cyclegan.txt](licenses/LICENSE-cyclegan.txt) | | pix2pixHD | BSD (NVIDIA) | [LICENSE-pix2pixhd.txt](licenses/LICENSE-pix2pixhd.txt) | | SPADE | CC BY-NC-SA 4.0: non-commercial, adaptations under the same licence | [LICENSE-spade.txt](licenses/LICENSE-spade.txt) | | BBDM, cBBDM | MIT; both checkpoints contain the CompVis latent-diffusion VQ-f4 autoencoder (MIT) | [LICENSE-bbdm.txt](licenses/LICENSE-bbdm.txt), [LICENSE-cbbdm.txt](licenses/LICENSE-cbbdm.txt), [LICENSE-latent-diffusion.txt](licenses/LICENSE-latent-diffusion.txt) | | HI-Diff | Apache-2.0 | [LICENSE-hi-diff.txt](licenses/LICENSE-hi-diff.txt) | | ResShift | S-Lab License 1.0: non-commercial use only | [LICENSE-resshift.txt](licenses/LICENSE-resshift.txt) | | SD2.1 FT, ControlNet, C-DiffSET | CreativeML Open RAIL++-M: derivatives of Stable Diffusion 2.1-base; the use-based restrictions of Attachment A apply to any use and redistribution. Training code: MIT (C-DiffSET) or Apache-2.0 (diffusers) | [LICENSE-CreativeML-OpenRAIL++-M.txt](licenses/LICENSE-CreativeML-OpenRAIL++-M.txt), [LICENSE-c-diffset.txt](licenses/LICENSE-c-diffset.txt), [LICENSE-diffusers.txt](licenses/LICENSE-diffusers.txt) | | DDPM (SR3), StegoGAN, CondDiff, E3Diff, Seg-CycleGAN | the upstream code publishes no licence | [NO-UPSTREAM-LICENSE.md](licenses/NO-UPSTREAM-LICENSE.md), [LICENSE-SoftPool.txt](licenses/LICENSE-SoftPool.txt) (bundled in the E3Diff code) | The training data are not redistributed; dataset terms are listed in [docs/DATASETS.md](https://github.com/KAIST-VICLab/GeoSET/blob/main/docs/DATASETS.md) of the GeoSET repository. ## Citation ```bibtex @article{do2026geoset, title={GeoSET: Generalist Foundation Model for SAR-to-EO Image Translation}, author={Do, Jeonghyeok and Kim, Munchurl}, journal={arXiv preprint arXiv:2609.37496}, year={2026} } ```