GeoSET: Generalist Foundation Model for SAR-to-EO Image Translation

Paired synthetic aperture radar (SAR) and electro-optical (EO) imagery is increasingly available across sensors, resolutions, and geographic regions. Yet existing SAR-to-EO image translation (SET) methods are typically trained on a single, limited-scale dataset, producing models specialized to particular sensing conditions.

We introduce GeoSET, the first generalist model for SET, built around a single pretrained parent that is adapted to downstream datasets under a common protocol. We curate over 3 million high-quality SAR–EO pairs from a collection of more than 10 million SAR observations, spanning diverse sensors, spatial resolutions, and ground sampling distances. To bridge the modality gap between SAR observations and a pretrained image generator, we develop a speckle-robust SAR encoder and pretrain the conditional generator on this heterogeneous corpus.

The resulting parent supports efficient adaptation across downstream datasets through low-rank adaptation (LoRA), updating only 0.60% of the generator parameters and requiring approximately one hour per dataset. Across six downstream benchmarks, GeoSET achieves state-of-the-art results in FID and DISTS with full fine-tuning or LoRA, demonstrating effective transfer across heterogeneous SAR–EO domains.

Pretrain once, adapt to many

One parent model, pretrained on 3M+ curated SAR–EO pairs, adapted to each benchmark with LoRA or full fine-tuning.

Grid with five benchmark rows (QXS-SAROPT, SAR2Opt, SAR2EO, SpaceNet6, CAP-BSG) and nine columns: input SAR, SD2.1 fine-tuning, BBDM, E3Diff, cBBDM, C-DiffSET, GeoSET with LoRA, GeoSET with full fine-tuning, and ground-truth EO. — **Qualitative comparison on SAR-to-EO image translation benchmarks.** Columns (g)–(h): GeoSET with LoRA and full fine-tuning; (i): ground-truth EO.

More scenes Fig. 5 · four benchmarks

Grid with four benchmark rows (QXS-SAROPT, SAR2Opt, SAR2EO, SpaceNet6) and nine columns: input SAR, SD2.1 FT, BBDM, E3Diff, cBBDM, C-DiffSET, GeoSET LoRA, GeoSET full FT, and ground-truth EO. — **Qualitative comparison on QXS-SAROPT, SAR2Opt, SAR2EO and SpaceNet6.** All methods receive the same SAR observation.

All methods, same scenes

Every retrained baseline and both GeoSET variants on the same SAR input, 20 scenes per benchmark. Select a tile to compare it with the ground truth.

Input SAR GeoSET (ours) Ground-truth EO Baseline (retrained)

Use the arrows or the ← → keys to change scene · click a tile to open the comparison sliderSwipe or use the arrows to change scene · tap a tile to open the comparison slider

Quantitative results

Across full fine-tuning and LoRA, GeoSET achieves the best reported FID on all six benchmarks and the best DISTS on five.

Radar chart of FID and DISTS on QXS-SAROPT, SAR2Opt, SAR2EO, SpaceNet6, CAP-BSG and KOMPSAT, normalized per dataset and metric, for CondDiff, E3Diff, cBBDM, C-DiffSET, GeoSET (LoRA) and GeoSET (full FT). The two GeoSET curves are the outermost on most axes. — **Cross-dataset comparison.** FID and DISTS per benchmark, normalized as 100 × best/value (outer ring = best). FID and DISTS are the primary metrics.

Table 2. Quantitative comparison on QXS-SAROPT and SAR2Opt. We retrain (evaluate) all competing methods on the same training (test) splits.

Bold bestUnderline second best↓ lower is better · ↑ higher is better

Show

Scroll for more columns

Method	Venue	QXS-SAROPT							SAR2Opt
Method	Venue	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑
General image-to-image translation methods
pix2pix	CVPR'17	174.6	0.373	0.1713	0.275	0.665	0.203	12.33	261.9	0.347	0.2164	0.277	0.657	0.199	13.39
CycleGAN	ICCV'17	115.5	0.362	0.0802	0.283	0.647	0.278	13.25	139.1	0.323	0.0343	0.374	0.642	0.188	12.68
pix2pixHD	CVPR'18	85.7	0.298	0.0492	0.403	0.573	0.358	16.13	146.3	0.283	0.0654	0.475	0.567	0.268	15.95
SPADE	CVPR'19	90.7	0.292	0.0607	0.366	0.599	0.320	14.53	142.5	0.265	0.0518	0.447	0.597	0.234	14.47
DDPM (SR3)	TPAMI'22	43.8	0.311	0.0189	0.425	0.620	0.359	14.04	122.5	0.295	0.0437	0.497	0.610	0.313	13.65
SD2.1 FT	CVPR'22	19.1	0.257	0.0042	0.489	0.561	0.348	15.40	71.8	0.211	0.0094	0.600	0.541	0.293	16.24
BBDM	CVPR'23	76.6	0.270	0.0479	0.414	0.568	0.352	15.34	143.1	0.290	0.0671	0.466	0.590	0.276	15.29
ControlNet	ICCV'23	50.4	0.307	0.0211	0.458	0.604	0.297	13.42	140.5	0.350	0.0480	0.479	0.643	0.217	11.73
HI-Diff	NeurIPS'23	324.3	0.539	0.3269	0.215	0.692	0.457	17.10	319.8	0.473	0.2357	0.277	0.692	0.384	17.36
ResShift	NeurIPS'23	140.2	0.334	0.0872	0.295	0.607	0.217	14.20	141.7	0.304	0.0515	0.435	0.597	0.177	14.31
StegoGAN	CVPR'24	106.8	0.384	0.0707	0.261	0.658	0.254	12.96	149.8	0.332	0.0396	0.362	0.652	0.162	12.39
SAR-to-EO image translation (SET) methods
CondDiff	GRSL'23	88.6	0.355	0.0537	0.310	0.730	0.213	11.55	211.8	0.415	0.1379	0.343	0.686	0.248	12.48
E3Diff	GRSL'24	47.8	0.278	0.0167	0.379	0.530	0.302	16.44	104.7	0.232	0.0306	0.541	0.529	0.249	16.09
cBBDM	GRSL'25	50.6	0.246	0.0284	0.492	0.539	0.372	16.02	222.3	0.377	0.1521	0.413	0.571	0.361	17.05
C-DiffSET	TCSVT'26	19.9	0.233	0.0055	0.522	0.526	0.380	16.92	78.1	0.214	0.0138	0.601	0.529	0.314	16.81
GeoSET (LoRA)	–	19.3	0.254	0.0050	0.544	0.564	0.326	14.67	74.6	0.200	0.0066	0.626	0.535	0.278	15.50
GeoSET (full FT)	–	16.9	0.244	0.0032	0.558	0.553	0.332	15.07	71.1	0.196	0.0052	0.614	0.532	0.279	15.66

Table 3. Quantitative comparison on SAR2EO and SpaceNet6. We retrain (evaluate) all competing methods on the same training (test) splits.

Bold bestUnderline second best↓ lower is better · ↑ higher is better

Show

Scroll for more columns

Method	SAR2EO							SpaceNet6
Method	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑
General image-to-image translation methods
pix2pix	169.9	0.315	0.1551	0.400	0.573	0.455	18.06	166.7	0.242	0.1091	0.589	0.419	0.459	17.07
CycleGAN	360.7	0.522	0.3903	0.369	0.683	0.300	15.17	119.3	0.219	0.0436	0.660	0.373	0.478	16.88
pix2pixHD	106.0	0.254	0.0638	0.543	0.439	0.619	21.68	175.1	0.246	0.1136	0.682	0.361	0.525	19.21
SPADE	98.5	0.260	0.0646	0.504	0.501	0.562	19.81	179.7	0.242	0.1228	0.662	0.385	0.489	17.77
DDPM (SR3)	85.0	0.272	0.0390	0.529	0.471	0.479	13.49	219.1	0.333	0.1616	0.419	0.622	0.111	12.45
SD2.1 FT	67.7	0.270	0.0282	0.511	0.487	0.459	15.46	124.7	0.196	0.0537	0.612	0.348	0.513	19.29
BBDM	63.7	0.231	0.0245	0.613	0.418	0.601	19.08	230.2	0.381	0.1647	0.470	0.468	0.449	18.03
ControlNet	87.8	0.315	0.0424	0.509	0.523	0.418	12.93	131.3	0.279	0.0635	0.615	0.413	0.448	14.45
HI-Diff	315.5	0.556	0.3007	0.155	0.647	0.672	21.68	261.2	0.300	0.2036	0.499	0.403	0.589	20.61
ResShift	130.6	0.273	0.0768	0.418	0.510	0.504	19.29	155.4	0.230	0.0858	0.540	0.367	0.421	18.58
StegoGAN	384.7	0.550	0.4343	0.387	0.698	0.180	12.49	113.1	0.225	0.0361	0.649	0.388	0.457	16.04
SAR-to-EO image translation (SET) methods
CondDiff	113.2	0.317	0.0639	0.461	0.551	0.334	11.12	160.1	0.274	0.1012	0.514	0.517	0.198	14.68
E3Diff	55.5	0.228	0.0148	0.550	0.446	0.514	20.58	110.6	0.197	0.0440	0.696	0.361	0.462	18.32
cBBDM	52.7	0.191	0.0214	0.675	0.369	0.650	21.73	280.4	0.347	0.2620	0.450	0.419	0.447	19.72
Seg-CycleGAN	–	–	–	–	–	–	–	131.4	0.229	0.0611	0.649	0.386	0.471	16.62
C-DiffSET	64.0	0.252	0.0257	0.527	0.462	0.512	17.09	132.3	0.197	0.0614	0.611	0.346	0.520	19.43
GeoSET (LoRA)	29.5	0.202	0.0042	0.691	0.440	0.547	19.19	87.9	0.189	0.0162	0.733	0.338	0.513	17.95
GeoSET (full FT)	24.5	0.181	0.0032	0.730	0.399	0.564	19.72	94.1	0.192	0.0183	0.724	0.342	0.515	17.99

Table 8. Quantitative comparison on CAP-BSG and KOMPSAT. We retrain (evaluate) all competing methods on the same training (test) splits.

Bold bestUnderline second best↓ lower is better · ↑ higher is better

Show

Scroll for more columns

Method	CAP-BSG							KOMPSAT
Method	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑
General image-to-image translation methods
pix2pix	245.8	0.402	0.2086	0.254	0.626	0.251	13.77	255.6	0.346	0.2892	0.298	0.669	0.116	10.33
CycleGAN	100.0	0.327	0.0485	0.365	0.569	0.353	16.84	160.8	0.359	0.1105	0.299	0.653	0.139	10.14
pix2pixHD	212.7	0.335	0.1673	0.301	0.570	0.367	17.23	223.2	0.379	0.2004	0.305	0.649	0.128	11.24
SPADE	187.5	0.354	0.1407	0.287	0.593	0.348	16.20	226.3	0.377	0.2106	0.303	0.670	0.157	11.47
DDPM (SR3)	132.0	0.403	0.0815	0.297	0.646	0.373	12.41	119.8	0.344	0.0695	0.365	0.645	0.121	10.21
SD2.1 FT	98.6	0.348	0.0406	0.321	0.587	0.380	15.86	100.7	0.333	0.0535	0.359	0.635	0.150	11.23
BBDM	145.2	0.346	0.0789	0.305	0.577	0.360	16.16	154.3	0.343	0.1101	0.322	0.656	0.145	9.74
ControlNet	120.5	0.426	0.0466	0.290	0.660	0.303	13.10	115.5	0.358	0.0672	0.357	0.671	0.137	9.63
HI-Diff	200.0	0.496	0.1328	0.234	0.697	0.463	16.75	277.8	0.523	0.2664	0.223	0.756	0.226	12.39
ResShift	201.2	0.378	0.1282	0.280	0.611	0.212	15.11	247.9	0.348	0.2193	0.306	0.652	0.115	11.12
StegoGAN	97.3	0.319	0.0464	0.367	0.567	0.345	16.89	325.4	0.497	0.3565	0.222	0.713	0.092	12.70
SAR-to-EO image translation (SET) methods
CondDiff	212.1	0.418	0.1347	0.246	0.734	0.234	10.50	115.6	0.349	0.0672	0.352	0.692	0.082	8.70
E3Diff	87.2	0.324	0.0337	0.346	0.543	0.332	16.79	116.5	0.333	0.0762	0.388	0.600	0.147	11.74
cBBDM	214.6	0.365	0.1580	0.280	0.578	0.437	17.91	245.3	0.434	0.2300	0.279	0.666	0.189	11.92
C-DiffSET	113.0	0.333	0.0517	0.310	0.573	0.382	17.03	144.9	0.344	0.0943	0.315	0.640	0.144	11.17
GeoSET (LoRA)	64.7	0.318	0.0154	0.402	0.566	0.343	16.85	72.0	0.299	0.0294	0.446	0.625	0.152	10.64
GeoSET (full FT)	58.9	0.316	0.0109	0.402	0.565	0.352	16.60	69.8	0.295	0.0292	0.447	0.620	0.153	10.78

Pretraining corpus Table 1

Table 1. Stage 2 SAR–EO pretraining corpus. Drop rates are computed relative to the original SAR sample counts and include source-specific pair eligibility selection and quality filtering. Crop-equivalent counts account for native image size and redundancy and determine the source sampling mixture.

Scroll for more columns

Source	Orig. samples	Kept pairs	Drop (%)	Crop equivalents	Mix (%)
GUSO	589,143	585,424	0.63	2,341,696	38.73
TerraMesh	8,194,048	1,741,434	78.75	1,741,434	28.80
SARLO-80	87,870	77,932	11.31	1,246,912	20.62
SAR-1M	1,130,379	688,458	39.09	688,458	11.39
3MOS	113,074	111,496	1.40	27,874	0.46
Total	10,114,514	3,204,744	68.32	6,046,374	100.00

Ablations and analysis Tables 5, 4 and 7

Table 5. Cumulative component comparison on SAR2EO and SpaceNet6.

Bold best↓ lower is better

Scroll for more columns

Configuration	SAR2EO			SpaceNet6
Configuration	FID↓	DISTS↓	LPIPS↓	FID↓	DISTS↓	LPIPS↓
SD2.1 Full FT	67.7	0.270	0.487	124.7	0.196	0.348
FLUX.2 LoRA (Backbone)	71.2	0.275	0.481	117.4	0.201	0.361
+ Pretraining corpus	50.5	0.250	0.470	107.9	0.194	0.351
+ SAR encoder	39.8	0.220	0.448	94.2	0.193	0.347
+ Speckle aug. (GeoSET)	29.5	0.202	0.440	87.9	0.189	0.338

Each step improves FID, DISTS and LPIPS on both datasets (SAR2EO FID 71.2 → 50.5 → 39.8 → 29.5). The last row is the complete GeoSET (LoRA) configuration, as in Table 3.

Table 4. Stage 1 encoder diagnostics and Stage 3 adaptation efficiency. (a) Reconstruction PSNR on original and perturbed SAR inputs before and after encoder training. EO reports reconstruction with the frozen pretrained codec. (b) Computational costs of full fine-tuning and LoRA.

Scroll for more columns

(a) Stage 1 Encoder Reconstruction

Source	Original SAR			Perturbed SAR			EO
Source	Before	After	Δ	Before	After	Δ	EO
GUSO	27.47	29.18	+1.71	23.54	26.96	+3.42	31.02
TerraMesh	20.11	32.04	+11.93	19.12	30.48	+11.37	33.95
SARLO-80	17.39	19.21	+1.82	16.71	18.49	+1.78	27.81
SAR-1M	29.95	31.25	+1.30	22.01	26.99	+4.98	33.59
3MOS	35.15	36.07	+0.92	27.30	31.92	+4.61	34.62
Average	26.01	29.55	+3.54	21.74	26.97	+5.23	32.20

(b) Downstream Adaptation Efficiency

GeoSET	Full FT	LoRA
Trainable parameters	3.852B	23.10M
Trainable fraction	100%	0.5997%
Per-dataset storage	30.8 GB	185 MB
Peak memory	96.9 GB	49.1 GB
Batch size	16	16
Throughput	67 img/s	94 img/s
Wall time/dataset	1.34 h	0.98 h

(a) Encoder training raises average SAR PSNR by +3.54 dB on original and +5.23 dB on perturbed inputs. (b) LoRA trains 23.10M parameters (0.5997%), stores 185 MB instead of 30.8 GB and takes 0.98 h per dataset.

Reconstructions from the adapted SAR encoder and the frozen EO encoder on six pretraining streams (GUSO, TerraMesh, SARLO-80, SAR-1M, 3MOS-MR, 3MOS-HR). Top row: SAR and EO reconstructions with PSNR values. Bottom row: speckle-perturbed SAR input at L = 4 and its reconstruction, with PSNR values. — **SAR and EO reconstruction across pretraining sources.** Top: reconstructions of original SAR and EO; bottom: speckle-perturbed SAR (L = 4, left) and its reconstruction (right). Values are PSNR (dB) relative to the original observations; values in parentheses are for the perturbed inputs.

Table 7. SAR conditioning intervention on pretraining sources. True denotes the corresponding SAR observation, shuffled uses a mismatched SAR observation, and null uses an all-zero SAR latent. The null condition is included during training through conditioning dropout.

Bold best↓ lower is better · ↑ higher is better

Scroll for more columns

	PSNR↑			LPIPS↓
Source	True	Shuffled	Null	True	Shuffled	Null
GUSO	13.36	10.24	10.58	0.548	0.678	0.690
SARLO-80	15.88	11.96	11.65	0.514	0.678	0.693
TerraMesh	13.69	10.94	10.80	0.491	0.670	0.695

True SAR conditioning is best on PSNR and LPIPS for all three pretraining sources.

Three stages, one parent

GeoSET framework. Top: multi-source SAR-EO pretraining corpus (GUSO, TerraMesh, SARLO-80, SAR-1M, 3MOS) and the downstream SET benchmarks. Stage 1: speckle-robust SAR representation pretraining with a frozen EO decoder. Stage 2: generalist SAR-to-EO pretraining of a DiT with flow matching, conditioned on the SAR latent. Stage 3: downstream adaptation with LoRA or full fine-tuning. — **GeoSET framework.** Stages 1–2 build one generalist parent; Stage 3 adapts it to each benchmark.

Stage 1Speckle-robust SAR encoder. Reconstructs SAR from speckle-perturbed input through a frozen decoder.
Stage 2Generalist FLUX.2 pretraining. The generator, conditioned on SAR instead of text, is pretrained on over 3M curated SAR–EO pairs.
Stage 3Per-benchmark adaptation. LoRA (0.60% of generator parameters, ≈1 h per dataset) or full fine-tuning.

BibTeX

@article{do2026geoset,
  title={GeoSET: Generalist Foundation Model for SAR-to-EO Image Translation},
  author={Do, Jeonghyeok and Kim, Munchurl},
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2026}
}

Prior work: C-DiffSET (IEEE TCSVT 2026)

@article{do2026cdiffset,
  title={C-diffset: Leveraging latent diffusion for sar-to-eo image translation with confidence-guided reliable object generation},
  author={Do, Jeonghyeok and Lee, Jaehyup and Lee, Seungchul and Kim, Munchurl},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2026},
  publisher={IEEE}
}