Water does not just tint an underwater photograph; it rewrites it. Each pixel mixes attenuated scene radiance, backscattered light, and scene light scattered forward on its way to the camera, and many combinations of scene and water explain the same pixels. Most Gaussian-splatting methods for underwater scenes model only the first two effects and remove the water by dividing by a transmission that decays exponentially with range, so their inversion gain grows without bound. We present NIRJALA (Sanskrit for "without water"), a 3D Gaussian Splatting framework that recovers water-free appearance and geometry together. TRIVENI, a three-path image formation model, adds a range-saturating forward-scattering term to direct transmission and backscatter; we prove that this keeps the inversion gain bounded at every range and bound how errors in the estimated backscatter and attenuation propagate into the recovered colours. PRATIBIMBA, a reversible physics anchor, ties a flexible neural water model to a ten-parameter physical model in both directions, so scene content cannot hide inside the water. SAMARASA collects the physics-informed objectives, including a range-binned colour prior. To score water removal against ground truth, we built SAJALA ("with water"), a 30-scene suite path-traced in Blender Cycles with water-free images, depth, and normals under controlled water colour and turbidity. On SAJALA, NIRJALA raises PSNR from 24.34 to 24.76 dB and lowers depth AbsRel from 0.109 to 0.091 against the strongest baseline; on real SeaThru-NeRF scenes it lowers LPIPS from 0.194 to 0.175 while rendering at 44 FPS. We also discuss where appearance metrics and geometry disagree underwater, and the limits of our evaluation.
Figures & tables
SAJALA (synthetic, 30 scenes)
SeaThru-NeRF (real, 4 scenes)
Method
Venue
SSIM ↑
PSNR ↑
LPIPS ↓
RMSE ↓
AbsRel ↓
δ<1.25↑
SSIM ↑
PSNR ↑
LPIPS ↓
FPS ↑
Train ↓
NeRF [ 4 ]
ECCV’20
0.565
22.069
0.556
12.652
0.539
0.012
0.525
16.327
0.655
0.01
13 h
SeaThru-NeRF [ 9 ]
CVPR’23
0.448
18.105
0.688
12.975
0.685
0.030
0.860
27.394
0.215
0.04
6 h
3DGS [ 5 ]
TOG’23
0.653
22.682
0.421
9.806
0.464
0.399
0.868
25.918
0.221
128.65
20 min
PGSR [ 20 ]
TVCG’25
0.661
23.241
0.405
8.316
0.110
0.782
0.879
27.548
0.238
137.10
64 min
RecGS [ 41 ]
RA-L’25
0.680
22.332
0.320
9.340
0.411
0.714
0.863
25.553
0.221
109.98
48 min
Table 1: Main comparison. Left: underwater rendering and depth on all 30 SAJALA scenes. Right: underwater rendering on the four real SeaThru-NeRF scenes, with efficiency. Bold : best; underline : second best (quality metrics only). RMSE is in the units of the ground-truth depth.
Panama
IUI3-RedSea
J.G. RedSea
Curaçao
Method
SSIM ↑
PSNR ↑
LPIPS ↓
SSIM ↑
PSNR ↑
LPIPS ↓
SSIM ↑
PSNR ↑
LPIPS ↓
SSIM ↑
PSNR ↑
LPIPS ↓
NeRF [ 4 ]
.545
17.08
.650
.503
15.70
.648
.467
14.43
.639
.584
18.11
.681
SeaThru-NeRF [ 9 ]
.881
28.75
.162
.824
26.75
.281
.844
23.21
.193
.891
30.87
.222
3DGS [ 5 ]
.887
28.99
.208
.862
23.56
.241
.828
21.23
.229
.893
29.90
.204
PGSR [ 20 ]
.889
29.13
.222
.864
27.56
.279
.870
23.44
.215
.891
30.07
.236
RecGS [ 41 ]
.877
27.72
.208
.860
23.75
.244
.824
21.09
.230
.891
29.65
.202
Table 2: Per-scene results on SeaThru-NeRF. Bold : best; underline : second best.
Underwater image
Water-free image
Depth
Scene
Method
SSIM ↑
PSNR ↑
LPIPS ↓
SSIM ↑
PSNR ↑
LPIPS ↓
RMSE ↓
AbsRel ↓
δ<1.25↑
Ruin
SeaThru-NeRF [ 9 ]
0.332
16.335
0.803
0.277
13.466
0.878
9.245
0.504
0.043
3D-UIR [ 12 ]
0.607
23.541
0.327
0.341
15.020
0.565
5.135
0.032
0.978
NIRJALA
0.619
23.991
0.325
0.352
15.480
0.563
4.662
0.029
0.979
Sunken Ship
SeaThru-NeRF [ 9 ]
0.483
17.424
0.664
0.183
10.265
0.952
14.042
0.723
0.017
3D-UIR [ 12 ]
0.740
24.143
0.263
0.324
12.681
0.512
7.472
0.103
0.753
Table 3: Underwater rendering, water-free appearance, and depth on four SAJALA scenes against the two strongest medium-aware baselines. NIRJALA is best in every column.
Underwater
Water-free
Depth
Water
SSIM ↑
PSNR ↑
LPIPS ↓
SSIM ↑
PSNR ↑
LPIPS ↓
RMSE ↓
AbsRel ↓
δ<1.25↑
Blue 1
0.744
25.127
0.274
0.326
15.184
0.430
1.851
0.034
0.965
Blue 2
0.766
25.831
0.270
0.320
14.417
0.465
1.771
0.034
0.965
Blue 3
0.787
26.451
0.265
0.308
12.984
0.511
2.062
0.039
0.954
Blue 4
0.806
27.035
0.258
0.300
11.179
0.563
1.721
0.030
0.977
Green 1
0.753
25.530
0.279
0.319
14.975
0.434
1.689
0.041
0.956
Table 4: Effect of water colour and concentration (levels 1–4, increasing) on Ruin V.
Variant
SSIM ↑
PSNR ↑
LPIPS ↓
NIRJALA (full)
0.898
28.43
0.175
w/o TRIVENI
0.891
27.87
0.181
w/o PRATIBIMBA
0.893
28.06
0.189
w/o TRIVENI and PRATIBIMBA
0.878
26.98
0.216
w/o Ωir
0.894
28.23
0.184
w/o Ωreg
0.893
27.94
0.181
Table 5: Ablations on SeaThru-NeRF (mean of four scenes). PRATIBIMBA comprises Ωir and Ωreg ; “prior” is Lprior from SAMARASA.
Underwater
Water-free
Depth
Setting
SSIM ↑
PSNR ↑
LPIPS ↓
SSIM ↑
PSNR ↑
LPIPS ↓
RMSE ↓
AbsRel ↓
δ<1.25↑
Non-uniform
0.624
25.069
0.433
0.130
11.105
0.773
5.155
0.241
0.586
Sparse
0.590
25.131
0.444
0.146
12.377
0.747
0.981
0.034
0.987
Pose noise
0.446
20.823
0.567
0.078
10.726
0.820
2.679
0.116
0.857
Table 6: Robustness on SAJALA. Sparse: 8 training views. Pose noise: 1∘ rotation and 0.5% translation.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Group
Item
Value
Optimisation
Iterations
30k, Adam
Position learning rate
1.6×10−4→1.6×10−6
Appearance / opacity learning rate
2.5×10−3 / 2.5×10−2
Scale / rotation learning rate
5×10−3 / 10−3
Optical-module learning rates
10−3 to 10−2
Densification
every 100 it., from 500 to 15k
Appendix
Table 7: Hyperparameters of NIRJALA.
Symbol
Meaning
I , Igt
observed underwater image; training target
J , Jc
water-free radiance rendered by the Gaussians; its colour-balanced version (output)
ρ , r , rˉ
rendered inverse depth; range r=(ρ+10−5)−1 ; detached, normalised range
TD , TB , TS
direct, backscatter, and forward-scattering transmissions
βD , βB , βf
attenuation, backscatter, and forward-scattering coefficients
Underwater 3D reconstruction recovers scene geometry and appearance from camera, sonar, and robot-sensor data despite refraction, scattering, and motion.