GeoCR: Learning a Generalist Cloud Removal Prior from Heterogeneous Observations

Cloud removal methods are typically specialized to individual datasets and input configurations, limiting reuse across sensors, spectral bands, and observation settings. We introduce GeoCR, a generalist model that unifies RGB-only-based CR and multispectral-based CR from single- or multi-temporal cloudy observations, with optional SAR guidance, within a single network.

To accommodate different spectral and sensing domains, compact input and output stems extend a pretrained RGB autoencoder while keeping its encoder and decoder trunks frozen. This shared latent interface enables a single flow transformer to jointly model clean RGB and non-RGB latents, conditioned on separate cloudy-observation streams and optional SAR tokens. Through joint pretraining on the training splits of ten datasets comprising 883,331 cloud-free target images, GeoCR learns a shared cloud removal prior across these heterogeneous configurations.

The same pretrained checkpoint supports direct inference without dataset-specific fine-tuning and efficient adaptation through low-rank adaptation (LoRA). We evaluate GeoCR against general image restoration and cloud removal methods on test splits of the contributing datasets under full-band and RGB-only settings. GeoCR achieves the best FID and DISTS on full-band SEN12MS-CR and Sen2_MTC_New and RGB-only CUHK-CR2, outperforming existing models and demonstrating the effectiveness of a reusable generative model across diverse settings.

One model, ten datasets

A single network removes clouds from RGB or multispectral, single- or multi-temporal inputs, with or without SAR.

Grid of five cloud-removal benchmarks, one per row (SEN12MS-CR, Sen2_MTC_New, CUHK-CR2, WHUS2-CRv, T-CLOUD). Columns: input cloudy image, UnCRtainTS, EMRDM, GACR, GeoCR without fine-tuning, GeoCR with LoRA, and the cloud-free ground truth. — **Qualitative comparison on cloud-removal benchmarks.** (e) GeoCR without fine-tuning and (f) with LoRA adaptation.

All methods, same scenes

Retrained baselines and both GeoCR variants on the same cloudy input, 12 scenes per setting. Select a tile to compare it with the ground truth.

Input cloudy GeoCR (ours) Ground truth Baseline (retrained)

Input Cloudy, SEN12MS-CR scene 1 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 1 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 1 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 1 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 1 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 1 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 1 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 2 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 2 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 2 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 2 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 2 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 2 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 2 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 3 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 3 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 3 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 3 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 3 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 3 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 3 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 4 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 4 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 4 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 4 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 4 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 4 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 4 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 5 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 5 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 5 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 5 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 5 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 5 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 5 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 6 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 6 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 6 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 6 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 6 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 6 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 6 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 7 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 7 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 7 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 7 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 7 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 7 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 7 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 8 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 8 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 8 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 8 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 8 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 8 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 8 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 9 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 9 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 9 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 9 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 9 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 9 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 9 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 10 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 10 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 10 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 10 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 10 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 10 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 10 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 11 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 11 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 11 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 11 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 11 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 11 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 11 — GACRECCV'26

Input Cloudy, SEN12MS-CR scene 12 — Input Cloudyinput

GeoCR (w/o FT), SEN12MS-CR scene 12 — GeoCR (w/o FT)Ours

GeoCR (LoRA), SEN12MS-CR scene 12 — GeoCR (LoRA)Ours

GT Clean, SEN12MS-CR scene 12 — GT Cleanreference

UnCRtainTS CVPRW'23, SEN12MS-CR scene 12 — UnCRtainTSCVPRW'23

EMRDM CVPR'25, SEN12MS-CR scene 12 — EMRDMCVPR'25

GACR ECCV'26, SEN12MS-CR scene 12 — GACRECCV'26

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

Paper figures

Grid of five RGB-only settings, one per row (SEN12MS-CR-RGB, Sen2_MTC_New-RGB, CUHK-CR2, WHUS2-CRv-RGB, T-CLOUD). Columns: input cloudy, pix2pix, pix2pixHD, BBDM, HI-Diff, GeoCR without fine-tuning, GeoCR with LoRA, and the ground truth. — **RGB-only settings.** Comparison with general restoration methods: pix2pix, pix2pixHD, BBDM and HI-Diff.

Quantitative results

GeoCR achieves the best FID and DISTS on full-band Sen2_MTC_New and SEN12MS-CR and RGB-only CUHK-CR2.

Radar chart of normalized FID and DISTS on CUHK-CR2, WHUS2-CRv, T-CLOUD, SEN12MS-CR and Sen2_MTC_New for UnCRtainTS, EMRDM, GACR, GeoCR (w/o FT) and GeoCR (LoRA); larger is better. — **Cross-dataset comparison.** FID and DISTS on five benchmarks, normalized as 100 × best / value (larger is better).

Table 2. Quantitative comparison on CUHK-CR2. RGB evaluation setting, all ten baselines.

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

Scroll for more columns

Method	Venue	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑
General image translation/restoration methods
pix2pix	CVPR’17	194.8	0.241	0.1201	0.538	0.319	0.452	19.46
pix2pixHD	CVPR’18	257.1	0.262	0.2130	0.379	0.304	0.401	19.34
BBDM	CVPR’23	514.6	0.542	0.5827	0.131	0.737	0.326	17.68
HI-Diff	NeurIPS’23	165.7	0.192	0.0769	0.664	0.253	0.638	23.54
Cloud removal methods
UnCRtainTS	CVPRW’23	165.6	0.265	0.0761	0.529	0.345	0.586	22.12
DiffCR	TGRS’24	245.2	0.285	0.1874	0.615	0.353	0.582	22.85
IDF-CR	TGRS’24	167.0	0.214	0.0873	0.645	0.270	0.641	23.18
ThiefCloud	TCSVT’25	136.3	0.226	0.0593	0.668	0.232	0.638	23.88
EMRDM	CVPR’25	104.4	0.167	0.0259	0.727	0.208	0.654	23.61
GACR	ECCV’26	125.0	0.178	0.0496	0.739	0.217	0.620	23.45
Ours
GeoCR (w/o FT)	–	93.6	0.153	0.0200	0.784	0.194	0.607	23.21
GeoCR (LoRA)	–	94.7	0.152	0.0194	0.779	0.196	0.601	23.11

Table 3. Quantitative comparison on Sen2_MTC_New. (a) Full-band setting. (b) RGB setting.

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

(a) Full-band (RGB + NIR)Scroll for more columns

Method	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑
UnCRtainTS	113.4	0.258	0.0597	0.471	0.428	0.588	17.15
DiffCR	95.9	0.297	0.0388	0.499	0.312	0.599	19.26
ThiefCloud	143.6	0.333	0.0767	0.310	0.508	0.481	14.92
EMRDM	96.6	0.218	0.0459	0.533	0.314	0.640	18.20
GACR	87.4	0.223	0.0358	0.535	0.320	0.626	18.90
GeoCR (w/o FT)	52.9	0.182	0.0032	0.642	0.221	0.661	20.77
GeoCR (LoRA)	52.7	0.181	0.0031	0.643	0.220	0.660	20.76

(b) RGB-onlyScroll for more columns

Method	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑
pix2pix	161.6	0.305	0.1188	0.224	0.480	0.541	20.16
pix2pixHD	153.9	0.292	0.1121	0.274	0.365	0.646	24.00
BBDM	179.4	0.392	0.1243	0.243	0.504	0.652	24.48
HI-Diff	125.8	0.302	0.0695	0.390	0.398	0.742	25.66
GeoCR (w/o FT)	79.1	0.252	0.0223	0.489	0.313	0.706	25.41
GeoCR (LoRA)	78.4	0.265	0.0313	0.520	0.333	0.712	25.07

Table 4. Quantitative comparison on SEN12MS-CR. (a) Full-band setting. (b) RGB setting.

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

(a) Full-band (13 spectral bands)Scroll for more columns

Method	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑
UnCRtainTS	80.5	0.276	0.0497	0.440	0.313	0.884	28.77
EMRDM	75.7	0.274	0.0475	0.467	0.304	0.873	28.70
GACR	96.9	0.331	0.0679	0.358	0.333	0.828	27.76
GeoCR (w/o FT)	28.4	0.202	0.0091	0.571	0.215	0.866	29.29
GeoCR (LoRA)	29.5	0.205	0.0098	0.569	0.216	0.866	29.19

(b) RGB-onlyScroll for more columns

Method	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑
pix2pix	145.9	0.317	0.1000	0.281	0.465	0.632	21.64
pix2pixHD	70.8	0.239	0.0443	0.420	0.277	0.772	27.23
BBDM	96.1	0.284	0.0536	0.331	0.367	0.710	25.44
HI-Diff	62.7	0.284	0.0342	0.438	0.320	0.837	28.28
GeoCR (w/o FT)	43.9	0.244	0.0246	0.532	0.326	0.712	22.15
GeoCR (LoRA)	52.1	0.263	0.0375	0.529	0.308	0.781	24.59

FID and DISTS are the primary metrics. T-CLOUD, CUHK-CR1 and WHUS2-CRv: Table 10.

Pretraining corpus Table 1 · ten datasets

Table 1. Overview of the GeoCR pretraining corpus. Counts refer to cloud-free target images in the training splits. Spectral bands describe the targets unless otherwise noted; cloudy conditions indicate the number of optical input frames. Tile sizes are in pixels. L1C: Level-1C; TOA/BOA: top-/bottom-of-atmosphere reflectance; RGB: red, green, and blue; NIR: near-infrared; SAR: synthetic aperture radar; GSD: ground sampling distance. B8 and B10 denote Sentinel-2 band identifiers.

Scroll for more columns

Source	Imagery source	Spectral bands	Delivered tile	GSD (m)	Cloudy cond.	SAR cond.	Clean targets
AllClear	Sentinel-2 + Sentinel-1	13 (L1C, TOA)	256²	10	1–3	✓	662,022
SEN12MS-CR	Sentinel-2 + Sentinel-1	13 (L1C, TOA)	256²	10	1	✓	107,143
WHUS2-CRv	Sentinel-2	13 (BOA; B10: L1C, TOA)	384²/192²/64²	10	1	✗	18,816
Sen2_MTC_Old	Sentinel-2	3 (RGB; +NIR for cloudy inputs)	256²	10	3	✗	88,874
Sen2_MTC_New	Sentinel-2	4 (RGB+B8, BOA)	256²	10	3	✗	2,380
T-CLOUD	Landsat-8	3 (RGB, 8-bit)	256²	30	1	✗	2,234
RICE2	Landsat-8	3 (RGB, 8-bit)	512²	30	1	✗	553
CUHK-CR1	Jilin-1 KF01B	4 (RGB+NIR, 8-bit)	512²	0.5	1	✗	508
CUHK-CR2	Jilin-1 KF01B	4 (RGB+NIR, 8-bit)	512²	0.5	1	✗	426
RICE1	Google Earth	3 (RGB, 8-bit)	512²	–	1	✗	375

Ablations Tables 5–7 · stems, temporal and SAR conditioning

Table 5. Effect of Stage 1 stem training. Reconstruction PSNR (dB) on AllClear validation samples before and after stem training, with the pretrained encoder and decoder frozen. VV and VH denote SAR polarizations.

Scroll for more columns

Recon. PSNR	RGB (w/o training)			Non-RGB (stem training)			SAR (stem training)
Recon. PSNR	All	Clear	Cloudy	All	Clear	Cloudy	All	VV	VH
Before	45.78	46.77	45.11	18.84	20.60	18.20	18.40	18.57	18.24
After	45.78	46.77	45.11	40.08	41.41	38.34	29.14	28.54	30.21
Gain	+0.00	+0.00	+0.00	+21.24	+20.81	+20.14	+10.74	+9.97	+11.98

Reconstruction grid for the RGB, single-band NIR, 10-band TOA, 10-band BOA and SAR routes (clear and cloudy observations of the same place; SAR clear only): input, reconstruction before stem training (trained routes only), and reconstruction with the final stem, labelled with PSNR in dB. — **Shared autoencoder.** Reconstruction PSNR (dB) through the frozen trunks. The RGB and initialized NIR routes stay fixed; the ten-band optical and SAR stems are trained in Stage 1.

Table 6. Temporal conditioning on Sen2_MTC_New. GeoCR uses the shared checkpoint without fine-tuning.

Scroll for more columns

Method	Cloudy	PSNR↑	SSIM↑
GeoCR	1	16.390	0.4595
GeoCR	2	19.319	0.5963
GeoCR	3	20.522	0.6568

Fixed test subset of Sen2_MTC_New; GeoCR (w/o FT).

Table 7. SAR conditioning on SEN12MS-CR. GeoCR uses the shared checkpoint without fine-tuning.

Scroll for more columns

Method	SAR	PSNR↑	SSIM↑
GeoCR	✗	28.596	0.8572
GeoCR	✓	29.442	0.8703

Fixed test subset of SEN12MS-CR; GeoCR (w/o FT).

Appendix tables Tables 9–12 · more benchmarks and settings

Table 9. Reconstruction by configuration. PSNR (dB) is computed from pooled reconstruction error on 64 training images per source, without clipping. These diagnostics differ from the held-out before/after comparison in the main paper and do not measure cloud removal.

Scroll for more columns

Route	Stem parameters	Trained	Reconstruction PSNR
10-band TOA	23,178	Yes	AllClear 38.23; SEN12MS-CR 37.69
10-band BOA	23,178	Yes	WHUS2-CRv 37.89
Single-band NIR	2,433	No	Sen2_MTC_New 42.06; CUHK-CR 40.41; Sen2_MTC_Old 49.51
SAR VV/VH	4,738	Yes	AllClear 28.51; SEN12MS-CR 31.19

Table 10. Additional quantitative comparisons. Feature metrics use RGB views; PSNR/SSIM follow Table 12.

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

Scroll for more columns

Method	FID↓	DISTS↓	KID↓	DINO↑	LPIPS↓	SSIM↑	PSNR↑
(a) T-CLOUD: RGB display imagery
pix2pix	107.6	0.254	0.0509	0.450	0.330	0.635	20.13
pix2pixHD	60.8	0.169	0.0173	0.616	0.177	0.741	24.75
BBDM	186.6	0.422	0.1179	0.259	0.546	0.541	22.69
HI-Diff	40.2	0.124	0.0070	0.726	0.114	0.874	30.54
UnCRtainTS	63.3	0.180	0.0188	0.640	0.178	0.809	26.28
DiffCR	139.6	0.358	0.0732	0.376	0.482	0.402	20.67
IDF-CR	84.4	0.220	0.0304	0.467	0.236	0.787	25.99
ThiefCloud	40.4	0.123	0.0058	0.726	0.109	0.861	29.36
EMRDM	36.5	0.121	0.0045	0.759	0.110	0.867	28.25
GACR	39.2	0.120	0.0058	0.747	0.113	0.858	29.56
GeoCR (w/o FT)	36.5	0.126	0.0024	0.763	0.122	0.785	27.15
GeoCR (LoRA)	36.3	0.125	0.0022	0.751	0.122	0.786	27.23
(b) CUHK-CR1: RGB+NIR setting
UnCRtainTS	134.8	0.204	0.0283	0.712	0.311	0.679	23.82
DiffCR	237.6	0.291	0.1360	0.649	0.318	0.573	22.75
EMRDM	77.8	0.115	-0.0006	0.845	0.146	0.760	25.64
GACR	97.1	0.133	0.0099	0.825	0.160	0.726	25.02
GeoCR (w/o FT)	80.4	0.125	-0.0025	0.825	0.165	0.680	23.88
GeoCR (LoRA)	81.5	0.125	-0.0027	0.821	0.167	0.677	23.83
(c) WHUS2-CRv: native multispectral setting
UnCRtainTS	26.9	0.142	0.0062	0.808	0.139	0.926	31.14
IDF-CR	40.0	0.179	0.0115	0.691	0.188	0.858	29.00
EMRDM	16.9	0.101	0.0009	0.875	0.097	0.937	32.55
GACR	22.6	0.230	0.0033	0.789	0.138	0.882	31.09
GeoCR (w/o FT)	18.9	0.111	0.0027	0.850	0.106	0.905	32.29
GeoCR (LoRA)	18.5	0.107	0.0025	0.858	0.103	0.903	32.15
(d) WHUS2-CRv: RGB-only setting
pix2pix	95.1	0.238	0.0523	0.432	0.274	0.746	23.75
pix2pixHD	27.9	0.148	0.0057	0.751	0.145	0.828	26.77
BBDM	50.4	0.213	0.0171	0.413	0.291	0.639	25.19
HI-Diff	17.8	0.107	0.0018	0.853	0.096	0.894	29.91
GeoCR (w/o FT)	21.6	0.137	0.0032	0.786	0.133	0.833	27.16
GeoCR (LoRA)	20.7	0.138	0.0039	0.811	0.134	0.804	25.35

Table 11. Dataset representations and pretraining mixture. Bands describe optical targets; K is the number of cloudy observations. Mix denotes the source sampling probability.

Scroll for more columns

Source	Optical representation	SAR	K	Mix (%)	Training targets
AllClear	13 bands, L1C TOA	VV/VH	1–3	53.14	662,022
SEN12MS-CR	13 bands, L1C TOA	VV/VH	1	32.93	107,143
WHUS2-CRv	13 bands, BOA; B10 TOA	–	1	8.81	18,816
Sen2_MTC_Old	RGB; NIR in cloudy inputs only	–	3	2.00	88,874
Sen2_MTC_New	RGB+B8, BOA	–	3	1.10	2,380
T-CLOUD	RGB, 8-bit	–	1	1.10	2,234
RICE2	RGB, 8-bit	–	1	0.27	553
CUHK-CR1	RGB+NIR, 8-bit	–	1	0.25	508
CUHK-CR2	RGB+NIR, 8-bit	–	1	0.21	426
RICE1	RGB, 8-bit	–	1	0.19	375
Total				100.00	883,331

Table 12. Evaluation settings. N_test denotes the size of each evaluated test split. CUHK-CR1 counts reflect the duplicate exclusions described in the Appendix (Datasets and Preprocessing).

Scroll for more columns

Setting	PSNR/SSIM bands	N_test
SEN12MS-CR	13 bands	7,899
Sen2_MTC_New	RGB+B8	687
WHUS2-CRv	13 bands	3,746
T-CLOUD	RGB	588
CUHK-CR1	RGB+NIR	98
CUHK-CR2	RGB	111
SEN12MS-CR, RGB-only	RGB	7,899
Sen2_MTC_New, RGB-only	RGB	687
WHUS2-CRv, RGB-only	RGB	3,746

One latent interface, one shared prior

GeoCR framework diagram. Top: the multi-source cloud removal pretraining corpus, with cloudy and clear examples from ten datasets. Left, Stage 1: input and output stems (3x3 convolutions) around the shared frozen RGB encoder and decoder trunks; the RGB stems stay frozen, while the non-RGB and SAR stems are trained with reconstruction losses. Right, Stage 2: the frozen stems encode clean and cloudy RGB, non-RGB and SAR inputs into latents, and a DiT initialized from a pretrained generator, with multi-stream and single-stream blocks, is trained with a flow-matching loss on the joint target latent. — **GeoCR framework.** A shared latent interface enables joint cloud removal across heterogeneous spectral, temporal, and SAR configurations.

01Shared latent interface. Compact stems map non-RGB and SAR inputs into a frozen pretrained RGB autoencoder.
02One generalist prior. A single flow transformer is jointly pretrained on 10 datasets (883,331 cloud-free targets).
03Two modes. The same checkpoint is used directly (w/o FT) or adapted with LoRA.

BibTeX

@article{do2026geocr,
  title={GeoCR: Learning a Generalist Cloud Removal Prior from Heterogeneous Observations},
  author={Do, Jeonghyeok and Kim, Munchurl},
  journal={arXiv preprint arXiv:2609.32510},
  year={2026}
}

Main comparison tables