Underwater 3D reconstruction and appearance restoration remain challenging due to the complex optical properties of water, such as wavelength-dependent attenuation and scattering. Existing Neural Radiance Fields (NeRF)-based approaches often suffer from slow rendering and limited practicality, while vanilla 3D Gaussian Splatting (3DGS) lacks an effective mechanism to account for underwater degradation. To address this, we propose WaterClear-GS, a physics-informed reformulation of underwater Gaussian splatting that models degradation as intrinsic Gaussian attributes rather than external medium fields. Furthermore, we adopt a dual-branch optimization strategy that preserves underwater photometric consistency while encouraging a practical separation between restoration-oriented latent clean appearance and degradation. This strategy is enhanced by depth-guided geometry regularization and perception-driven image supervision, together with exposure constraints, spatially adaptive regularization, and physics-informed spectral regularization, which collectively promote spatial coherence and plausible visual appearance. Extensive experiments on standard benchmarks and our collected dataset demonstrate that WaterClear-GS achieves strong performance on both novel view synthesis (NVS) and underwater image restoration (UIR) tasks, while maintaining 160+ FPS real-time rendering. The code will be available at https://sherii888.github.io/WaterClear-GS/.
Figures & tables
Figure 1 : Illustration of underwater imaging and our results. Top : Light from objects is affected by significant attenuation and scattering, which intensify with distance. Bottom : Our method enables high-quality rendering at over 160 FPS and restores plausible clean appearance.
Figure 2 : WaterClear-GS pipeline. Our method augments each Gaussian with wavelength-aware optical proxy parameters for attenuation, backscatter, and veiling light. A shared Gaussian splatting core supports dual-branch rendering: (a) the underwater branch fits the input observations, and (b) the clear branch learns a restoration-oriented latent clean appearance. Depth-guided supervision stabilizes geometry, and exposure, spatial, and spectral regularizations further improve restoration quality.
Figure 3 : Visualization of the OAGM decomposition. The underwater observation is represented using latent clear radiance together with direct-attenuation and backscatter proxies. These components are rendering-oriented variables rather than calibrated medium measurements.
Method
SeaThru-NeRF
Submerged3D
Water3D
Avg.
Avg.
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Time ↓
FPS ↑
SeaThru-NeRF [ 22 ]
26.47
0.819
0.282
17.22
0.492
0.598
24.37
0.722
0.460
> 10h
< 0.1
3DGS [ 21 ]
26.71
0.901
0.204
22.71
0.767
0.315
29.19
0.858
0.307
5.37min
234.53
UW-GS [ 49 ]
27.15
0.899
0.202
21.98
0.731
0.356
26.55
0.849
0.315
30.79min
31.72
SeaSplat [ 53 ]
27.84
0.909
0.178
23.14
0.769
0.319
28.86
0.861
0.293
29.38min
60.28
WaterSplatting [ 23 ]
29.57
0.925
0.150
24.55
0.768
0.297
29.14
0.853
0.303
10.11min
75.10
Table 1: Average NVS results on SeaThru-NeRF [ 22 ] , Submerged3D [ 18 ] , and Water3D. Time and FPS are averaged across all scenes. Top-three results are marked by first , second , and third .
Figure 4 : Qualitative results of the NVS task. We present rendered underwater images and their depth maps. The pseudo-depth maps are used as a reference. Zoomed-in regions (highlighted with red bounding boxes) illustrate detailed differences. Our method consistently preserves geometric structures and fine-level details in most cases.
Figure 5 : Qualitative comparison for the UIR task. Across the shown scenes, WaterClear-GS generally produces more balanced colors and clearer local details, while some baselines exhibit underexposure, residual haze, or color shifts.
Method
ΔE00↓
ψˉ↓
NIQE ↓
UIQM ↑
UCIQE ↑
SeaThru-NeRF [ 22 ]
24.73
24.82
5.31
2.91
0.26
UW-GS [ 49 ]
19.13
21.44
4.11
4.16
0.28
SeaSplat [ 53 ]
20.55
21.68
4.13
3.65
0.34
WaterSplatting [ 23 ]
24.45
23.99
4.75
4.08
0.24
RUSplatting [ 18 ]
22.44
22.46
3.99
4.02
0.30
3D-UIR [ 57 ]
22.37
23.40
4.08
3.79
0.35
Table 2: Average UIR quality comparison.
Figure 6 : Results of single-channel and RGB optical modeling.
Figure 7 : Module-level visual ablation. OAGM enables the latent clean branch, reconstruction optimization improves structural fidelity, and restoration optimization further refines the recovered appearance.
Figure 8 : Restoration under controlled synthetic turbidity. The model is optimized from degraded observations without using the clean reference as supervision.
Base
OAGM
Rec. Opt.
Res. Opt.
PSNR ↑
SSIM ↑
LPIPS ↓
✓
✗
✗
✗
28.57
0.902
0.235
✓
✗
✓
✗
29.10
0.924
0.161
✓
✓
✗
✗
28.71
0.906
0.224
✓
✓
✗
✓
28.83
0.905
0.224
✓
✓
✓
✗
29.79
0.929
0.151
✓
✓
✓
✓
29.93
0.931
0.149
Table 3 : Ablation study of the main components on the SeaThru-NeRF dataset. Base denotes our 3DGS backbone trained with the same hyperparameters as WaterClear-GS.
Setting
PSNR ↑
SSIM ↑
LPIPS ↓
ΔE00↓
ψˉ↓
w/o Le
29.80
0.930
0.150
16.26
15.05
w/o Ls
30.00
0.931
0.149
15.84
15.77
w/o Lp
29.97
0.930
0.151
16.49
15.97
Full model
29.93
0.931
0.149
14.56
14.74
Table 4: Restoration-loss ablation on SeaThru-NeRF. ΔE00 and ψˉ measure latent clear-view color fidelity.