# CondDiff · SAR2Opt

CondDiff (GRSL'23) retrained by the GeoSET authors on the SAR2Opt split. These are the weights behind
the CondDiff row for SAR2Opt in Table 2 of the GeoSET paper
([arXiv:2609.37496](https://arxiv.org/abs/2609.37496)).

- Method: Bai, Pu and Xu, "Conditional diffusion for SAR to optical image translation", GRSL 2023.
- Implementation: [Coordi777/Conditional-Diffusion-for-SAR-to-Optical-Image-Translation](https://github.com/Coordi777/Conditional-Diffusion-for-SAR-to-Optical-Image-Translation) at commit `8326f54adf3a8d717792cb7abc7d3d2b7dfac565`.
- Training and inference code: [`baselines/conddiff`](https://github.com/KAIST-VICLab/GeoSET/tree/main/baselines/conddiff) in the GeoSET repository.

## Files

| File | Bytes | SHA-256 | Role |
|---|---:|---|---|
| `ema_final.pt` | 662,459,343 | `3075a059144ef972d3b0129fded5d4d94193e5406bb98e5b30e6afd3fbb64653` | UNet EMA (0.9999) weights at step 50,000 (512 × 512 model) |

## Results (GeoSET paper, Table 2)

SAR2Opt test split, 627 pairs, centre 512 × 512 crop.

| FID↓ | DISTS↓ | KID↓ | DINO↑ | LPIPS↓ | SSIM↑ | PSNR↑ |
|---:|---:|---:|---:|---:|---:|---:|
| 211.8 | 0.415 | 0.1379 | 0.343 | 0.686 | 0.248 | 12.48 |

## Inference

```bash
git clone https://github.com/KAIST-VICLab/GeoSET && cd GeoSET
export BASELINES_ROOT=/path/to/upstream_code DATA_ROOT=/path/to/data WORK_DIR=/path/to/work CKPT_ROOT=/path/to/checkpoints
bash baselines/conddiff/fetch.sh
python baselines/data/diff_data.py conddiff sar2opt
hf download JeonghyeokDo/GeoSET --include "baselines/sar2opt/conddiff/*" --local-dir hf
mkdir -p $CKPT_ROOT/conddiff/sar2opt && cp hf/baselines/sar2opt/conddiff/ema_final.pt $CKPT_ROOT/conddiff/sar2opt/
bash baselines/conddiff/test.sh sar2opt        # writes $WORK_DIR/results/conddiff/sar2opt/<stem>.png
python scripts/evaluate.py --dataset sar2opt --pred-dir $WORK_DIR/results/conddiff/sar2opt
```

## Licence

The CondDiff repository contains no licence file; it is derived from [openai/guided-diffusion](https://github.com/openai/guided-diffusion) (MIT). These weights were trained by the GeoSET authors with that code; the GeoSET repository distributes only its changes to the code.