Authors: Ningjing Fan, Yiqun Wang, Dong-Ming Yan, Peter Wonka
Organizations: Chongqing University, China · MAIS, Institute of Automation, Chinese Academy of Sciences and UCAS, China · King Abdullah University of Science and Technology (KAUST), Saudi Arabia
The reflective appearance, especially strong and typically near-field specular reflections, poses a fundamental challenge for accurate surface reconstruction and novel view synthesis. Existing Gaussian splatting methods either fail to model near-field specular reflections or rely on explicit ray tracing at substantial computational cost. We present Ref-DGS, a reflective dual Gaussian splatting framework that addresses this trade-off by decoupling surface reconstruction from specular reflection within an efficient rasterization-based pipeline. Ref-DGS introduces a dual Gaussian scene representation consisting of geometry Gaussians and complementary local reflection Gaussians that capture near-field specular interactions without explicit ray tracing, along with a global environment reflection field for modeling far-field specular reflections. To predict specular radiance, we further propose a lightweight, physically-aware specular adaptive mixing shader that fuses global and local specular features. Experiments demonstrate that Ref-DGS achieves state-of-the-art performance on reflective scenes while training substantially faster than ray-based Gaussian methods.
Figures & tables
Figure 1. Left: Surface reconstruction results on the RefReal dataset for two challenging scenes, gardenspheres (top) and toycar (bottom), both with strong near-field specular reflections. The bottom row reports the average training time. Right: Results on the GlossyReal dataset for the vase scene, showing surface reconstruction and rendering quality. Our method produces more accurate geometry and higher-quality reflections, as highlighted in the close-up regions.
Figure 2. An overview of our Ref-DGS framework.
Figure 3. Left: Input image. Middle: Normal map of the local reflection Gaussians Glocal . Right: Illustration of specular reflection, where the surface normal n is perpendicular to the surface, and the image distance ( id ) equals the object distance ( od ).
ball
car
coffee
helmet
teapot
toaster
mean
time ↓
Ref-NeRF
1.55
14.93
12.24
29.48
9.23
42.87
18.38
–
ENVIDR
0.74
7.10
9.23
1.66
2.47
6.45
4.61
–
UniSDF
0.45
6.88
8.00
1.72
2.80
6.45
8.71
–
PGSR
66.93
4.62
2.91
6.01
1.01
15.31
16.13
27.5m
MILo
61.30
4.47
2.36
4.87
0.62
16.49
15.02
18.9m
GaussianShader
7.03
14.05
14.93
9.33
7.17
13.08
10.93
1.10h
Table 1. Quantitative surface reconstruction comparison on the ShinySynthetic dataset (normal MAE and training time). Best results are highlighted in red and second-best in yellow.
angel
bell
cat
horse
luyu
potion
tbell
teapot
mean
CD ↓
PGSR
0.77
3.08
3.39
0.90
1.16
3.99
4.59
4.57
2.81
MILo
0.77
16.52
2.84
1.02
44.64
3.39
5.08
4.40
9.83
GaussianShader
0.85
1.10
2.56
0.73
1.07
4.74
5.74
3.40
2.53
Ref-Gaussian
0.45
0.70
1.68
0.64
0.88
0.81
0.59
1.01
0.95
Ref-GS
0.41
0.74
1.73
0.47
0.89
1.05
0.52
0.88
0.84
Table 2. Quantitative surface reconstruction comparison on the GlossySynthetic dataset (CD ×102 and normal MAE). Best results are highlighted in red and second-best in yellow.
Datasets
ShinySynthetic
GlossySynthetic
RefReal
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Ref-NeRF
33.32
0.956
0.110
25.65
0.905
0.112
23.62
0.646
0.239
ENVIDR
32.88
0.969
0.072
29.06
0.947
0.060
23.80
0.606
0.332
PGSR
26.82
0.921
0.119
25.59
0.913
0.091
24.43
0.696
0.220
MILo
27.28
0.921
0.124
25.69
0.914
0.096
23.87
0.643
0.291
GaussianShader
30.42
0.951
0.082
27.11
0.922
0.082
23.46
0.647
0.257
Table 3. Quantitative NVS comparison. Best results are highlighted in red and second-best in yellow.
Figure 4. Qualitative surface reconstruction results (meshes and normals) on the ShinySynthetic (coffee, toaster) and GlossySynthetic (cat) datasets.
Figure 5. Qualitative rendering quality and surface reconstruction on the GlossyReal (bear, vase) dataset.
Figure 6. Qualitative NVS results on the ShinySynthetic (coffee, toaster) and GlossySynthetic (cat, tbell, teapot) datasets.
Figure 7. Ablation study on surface reconstruction.
Figure 8. Ablation study on surface normal estimation. Angular error maps (red: high, blue: low) show that the full model significantly reduces geometric artifacts compared to ablated variants. Chamfer distances (CD) are reported at the bottom.
Figure 9. Ablation study on novel view synthesis.
Figure 10. Ablation study on the diffuse leakage issue, visualizing the predicted diffuse color.
EnvGS
MaterialRefGS
Ref-GS
Ours
FPS ↑
26.22
30.07
67.63
76.34
Table 4. Rendering speed (FPS) on the ShinySynthetic dataset.
PSNR ↑
SSIM ↑
LPIPS ↓
MAE ↓
CD ↓
MAE ↓
w/o Flocal w/o Glocal
34.36
0.971
0.057
1.53
0.82
1.99
w/o Glocal
34.70
0.973
0.057
1.46
0.73
1.93
w/o Adap Mix
34.98
0.974
0.054
1.46
0.64
1.92
w/o Phy-Aware
35.02
0.975
0.053
1.45
0.62
1.90
Full model
35.21
0.975
0.053
1.43
0.62
1.88
Table 5. Ablation study on the ShinySynthetic (PSNR, SSIM, LPIPS, and normal MAE) and GlossySynthetic (CD ×102 and normal MAE) datasets.