While 3D Gaussian splatting has emerged as a powerful paradigm, it fundamentally fails to model transparency such as glass panels. The core challenge lies in decoupling the intertwined radiance contributions from transparent interfaces and the transmitted geometry observed through the glass. We present GLINT, a framework that models scene-scale transparency through explicit decomposed Gaussian representation. GLINT reconstructs the primary interface and models reflected and transmitted radiance separately, enabling consistent radiance transport. During optimization, GLINT bootstraps transparency localization from geometry-separation cues induced by the decomposition, together with geometry and material priors from a pre-trained video relighting model. Extensive experiments demonstrate consistent improvements over prior methods for reconstructing complex transparent scenes.
Figures & tables
Figure 2: Pipeline Overview. The interface component Gintr captures primary first-surface, while secondary components Gtrans and Grefl separately model transmission and reflection. The output color Lo is obtained through hybrid rendering under transparency-aware radiance transport and supervised by photometric loss Lphoto . DiffusionRenderer [ 26 ] provides priors that regularize the G-buffers via Lgeo .
Table 1: Quantitative evaluation of geometry on the synthetic 3D-FRONT-T dataset. We report normal metrics (MAE, and accuracy thresholds of 11.25∘ , 22.5∘ ) and depth metrics (AbsRel, RMSE, δ<1.25 ), along with mesh metrics (CD, F1-score).
Figure 5: Qualitative comparison on synthetic scenes. Each column shows results from GT, PGSR [ 5 ] , EnvGS [ 39 ] , TSGS [ 24 ] , and Ours. For each scene, rows correspond to RGB (top) and depth maps (bottom).
Figure 6: Mesh visualization comparison. The meshes are obtained from TSDF fusion following baselines.
Figure 7: Qualitative comparison on DL3DV-10K dataset. Each column shows results from GT, PGSR [ 5 ] , EnvGS [ 39 ] , TSGS [ 24 ] , and Ours. For each scene, rows correspond to RGB (top) and normal (bottom) maps.
Method
DL3DV-10K [ 27 ]
3D-FRONT-T
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
2DGS [ 15 ]
29.18
0.91
0.13
32.26
0.94
0.08
PGSR [ 5 ]
29.26
0.92
0.11
31.96
0.95
0.07
Ref-GS [ 46 ]
29.11
0.92
0.12
32.30
0.94
0.09
EnvGS [ 39 ]
29.65
0.91
0.12
33.71
0.94
0.07
TSGS [ 24 ]
25.94
0.85
0.19
28.80
0.87
0.14
Table 2: Photometric evaluation on real and synthetic datasets. GLINT achieves state-of-the-art rendering quality, quantitatively outperforming all baseline methods on both benchmarks.
Figure 8: Qualitative Ablation study of G-buffer guidance.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
MAE ↓
AbsRel ↓
w/o Gtrans
32.26
0.93
0.085
8.11
0.038
w/o Grefl
32.70
0.94
0.067
8.78
0.038
w/o Ltrans
33.57
0.94
0.066
8.07
0.037
w/o Lnormal
33.92
0.95
0.060
12.21
0.043
w/o Ldepth
34.05
0.95
0.058
8.54
0.061
w/o Lgeo
33.62
0.95
0.060
24.69
0.126
Table 3: Ablation studies on our method. We report PSNR, SSIM, LPIPS, Normal MAE, and Depth AbsRel.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: 3D-FRONT-T dataset. RGB, depth, and normal ground truth examples. As can be seen in the transparent surface (glass), the outgoing radiance for each pixel is entanglement of radiance from interface surface and the transmitted radiance.
Figure 10: Qualitative ablation study on representation components. Visual comparison between the full model and ablated variants.
Figure 11: Effect of geometric losses. Ablating geometric supervision leads to degraded geometry reconstruction. Removing Ldepth produces inaccurate interface depth, removing Lnormal results in unstable surface orientation, and removing all geometric losses ( Lgeo ) severely distorts both depth and normals.
Figure 12: Effect of the transparency loss Ltrans . With the proposed transparency loss Ltrans , the learned transparency maps align well with the true transmitted content, while removing this loss leads to noisy or inconsistent transparency estimation.
Figure 13: Radiance decomposition comparison. Each 2×2 grid is arranged in clockwise order as rendered RGB, base color, reflection and transmission. EnvGS does not provide a transmission component, so the corresponding slot is left blank.
Figure 14: Comparison with EnvGS [ 39 ] trained with a DiffusionRenderer [ 26 ] prior. With the DiffusionRenderer normal priors, EnvGS still fails to produce consistent depth and normals for transparent regions. In particular, the transmission surface remains incorrectly reconstructed, showing blurry rendering quality.
Figure 15: Comparison of transparency masks across two scenes. Each scene is shown with three viewpoints (rows), including ground-truth reference RGB (left), our predicted transparency maps (middle), and G-SAM2 [ 31 ] masks (right). GT images provide visual context, highlighting that our method consistently isolates transparent surfaces, while Grounded-SAM2 often produces noisy, incomplete or completely missing masks.
Method
FPS ( ↑ )
Training Time ( ↓ )
PSNR ( ↑ )
MAE ( ↓ )
AbsRel ( ↓ )
TSGS [ 24 ]
159
∼ 40 mins
28.80
9.89
0.08
EnvGS [ 39 ]
80
∼ 1 h
33.71
14.37
0.13
Ours
51
∼ 2.5 h
34.50
7.96
0.04
Appendix
Table 4: Comparison of computational costs and reconstruction quality on the 3D-FRONT-T dataset.
Figure 16: Qualitative comparison on synthetic scenes. Each column shows results from GT, PGSR [ 5 ] , EnvGS [ 39 ] , TSGS [ 24 ] , and Ours. For each scene, rows correspond to RGB (top), depth (middle), and normal (bottom) maps.
Figure 17: Additional qualitative comparisons on the DL3DV-10K dataset. Each column shows results from GT, PGSR [ 5 ] , EnvGS [ 39 ] , TSGS [ 24 ] , and Ours. For each scene, rows correspond to RGB (top), depth (middle), and normal (bottom) predictions.
Gaussian Splatting is a powerful tool for reconstructing diffuse scenes, but it struggles to simultaneously model specular reflections and the appearance of objects behind semi-transparent surfaces. These specular reflections and transmittance are essential for realistic novel view synthesis, and existing methods do not properly incorporate the underlying physical processes to simulate them. To address this issue, we propose RT-GS, a unified framework that integrates a microfacet material model and ray tracing to jointly model specular reflection and transmittance in Gaussian Splatting. We accomplish this by using separate Gaussian primitives for reflections and transmittance, which allow modeling distant reflections and reconstructing objects behind transparent surfaces concurrently. We utilize a differentiable ray tracing framework to obtain the specular reflection and transmittance appearance. Our experiments demonstrate that our method successfully produces reflections and recovers objects behind transparent surfaces in complex environments, achieving significant qualitative improvements over prior methods where these specular light interactions are prominent.
Kunnong Zeng, Chensheng Peng, Yichen Xie +2
University of Utah · University of California, Berkeley
3D Gaussian Splatting (3DGS) enables real-time novel view synthesis with high visual quality. However, existing methods struggle with semi-transparent specular surfaces that exhibit both complex reflections and clear transmission, often producing blurry reflections or overly occluded transmission. To address this, we present RT-Splatting, a framework that disentangles each Gaussian's geometric occupancy from its optical opacity. This factorization yields a unified surface-volume scene representation with a single set of Gaussian primitives. Our hybrid renderer interprets this representation both as a surface to capture high-frequency reflections and as a volume to preserve clear transmission. To mitigate the ambiguity in jointly optimizing reflection and transmission, we introduce Specular-Aware Gradient Gating, which suppresses misleading gradients from highly specular regions into the transmission branch, effectively reducing distracting floaters. Experiments on challenging semi-transparent scenes show that RT-Splatting achieves state-of-the-art performance, delivering high-fidelity reflections and clear transmission with real-time rendering. Moreover, our factorization naturally enables flexible scene editing. The project page is available at https://sjj118.github.io/RT-Splatting.
Ji Shi, Xianghua Ying, Bowei Xing +2
State Key Laboratory of General Artificial Intelligence School of Intelligence Science and Technology Peking University
Transmissive scenes are ubiquitous in daily life, yet reconstructing and rendering them remains highly challenging due to the inherent entanglement between near-field reflections from the surrounding environment on the transmissive surface, and the transmitted content of the scene behind it. This coupling gives rise to dual surface geometries and dual radiance components within each observation, posing ambiguities for standard methods. We present TransmissiveGS, a novel framework for disentangled reconstruction and rendering of transmissive scenes. Specifically, we model the scene with a dual-Gaussian representation and introduce a deferred shading function to jointly render the two Gaussian components. To separate reflection and transmission, we exploit the inherent multi-view inconsistency of reflections and leverage the residuals from reconstructing multi-view consistent content as cues for disentangled geometry and appearance modeling. We further propose a reflection light field that enables high-fidelity estimation of near-field reflections. During training, we introduce a high-frequency regularization to preserve fine details. We also contribute a new synthetic dataset for evaluating transmissive surface reconstruction. Experiments on both synthetic and real-world scenes demonstrate that TransmissiveGS consistently outperforms prior Gaussian Splatting-based methods in both reconstruction and rendering quality for transmissive scenes.