GLINT: Modeling Scene-Scale Transparency via Gaussian Radiance Transport

Youngju Na; Jaeseong Yun; Soohyun Ryu; Hyunsu Kim; Sung-Eui Yoon; Suyong Yeon

cs.CVMar 27, 2026

GLINT: Modeling Scene-Scale Transparency via Gaussian Radiance Transport

Authors: Youngju Na, Jaeseong Yun, Soohyun Ryu, Hyunsu Kim, Sung-Eui Yoon, Suyong Yeon

Organizations: KAIST · NAVER LABS

Abstract

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 $\mathcal{G}_{\text{intr}}$ captures primary first-surface, while secondary components $\mathcal{G}_{\text{trans}}$ and $\mathcal{G}_{\text{refl}}$ separately model transmission and reflection. The output color $L_{o}$ is obtained through hybrid rendering under transparency-aware radiance transport and supervised by photometric loss $\mathcal{L}_{\text{photo}}$ . DiffusionRenderer [ 26 ] provides priors that regularize the G-buffers via $\mathcal{L}_{\text{geo}}$ .

Panel of Figure 3 — Figure 3: Depth decomposition. (Left) Rendered image, (Middle) interface depth, and (Right) transmission depth.

Panel of Figure 4 — Figure 4: Obtained transparency maps. GT images (first row) and learned transparency maps (second row).

Method	Normal			Depth			Mesh
Method	MAE $\downarrow$	$11.25^{\circ}\uparrow$	$22.5^{\circ}\uparrow$	AbsRel $\downarrow$	RMSE $\downarrow$	$\delta<1.25\uparrow$	CD $\downarrow$	F1 $\uparrow$
2DGS [ 15 ]	25.97	52.19	64.55	0.20	0.24	76.71	0.85	0.688
PGSR [ 5 ]	25.39	56.83	65.58	0.22	0.25	76.68	0.52	0.807
Ref-GS [ 46 ]	41.55	18.61	36.79	0.28	0.36	57.52	1.29	0.408
EnvGS [ 39 ]	14.37	68.22	80.23	0.13	0.16	86.10	0.87	0.640
TSGS [ 24 ]	9.89	86.29	92.24	0.08	0.12	95.56	0.52	0.798

Table 1: Quantitative evaluation of geometry on the synthetic 3D-FRONT-T dataset. We report normal metrics (MAE, and accuracy thresholds of

11.25^{\circ}

22.5^{\circ}

) and depth metrics (AbsRel, RMSE,

\delta<1.25

), along with mesh metrics (CD, F1-score).

Panel of Figure 5 — 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).

Panel of Figure 6 — Figure 6: Mesh visualization comparison. The meshes are obtained from TSDF fusion following baselines.

Panel of Figure 7 — 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
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
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.

Panel of Figure 8 — Figure 8: Qualitative Ablation study of G-buffer guidance.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MAE $\downarrow$	AbsRel $\downarrow$
w/o $\mathcal{G}_{\text{trans}}$	32.26	0.93	0.085	8.11	0.038
w/o $\mathcal{G}_{\text{refl}}$	32.70	0.94	0.067	8.78	0.038
w/o $\mathcal{L}_{\text{trans}}$	33.57	0.94	0.066	8.07	0.037
w/o $\mathcal{L}_{\text{normal}}$	33.92	0.95	0.060	12.21	0.043
w/o $\mathcal{L}_{\text{depth}}$	34.05	0.95	0.058	8.54	0.061
w/o $\mathcal{L}_{\text{geo}}$	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

Panel of Figure 9 — 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.

Panel of Figure 10 — Figure 10: Qualitative ablation study on representation components. Visual comparison between the full model and ablated variants.

Panel of Figure 11 — Figure 11: Effect of geometric losses. Ablating geometric supervision leads to degraded geometry reconstruction. Removing $\mathcal{L}_{\text{depth}}$ produces inaccurate interface depth, removing $\mathcal{L}_{\text{normal}}$ results in unstable surface orientation, and removing all geometric losses ( $\mathcal{L}_{\text{geo}}$ ) severely distorts both depth and normals.

Panel of Figure 12 — Figure 12: Effect of the transparency loss $\mathcal{L}_{\text{trans}}$ . With the proposed transparency loss $\mathcal{L}_{\text{trans}}$ , the learned transparency maps align well with the true transmitted content, while removing this loss leads to noisy or inconsistent transparency estimation.

Panel of Figure 13 — Figure 13: Radiance decomposition comparison. Each $2\times 2$ 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.

Panel of Figure 14 — 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.

Panel of Figure 15 — 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 ( $\uparrow$ )	Training Time ( $\downarrow$ )	PSNR ( $\uparrow$ )	MAE ( $\downarrow$ )	AbsRel ( $\downarrow$ )
TSGS [ 24 ]	159	$\sim$ 40 mins	28.80	9.89	0.08
EnvGS [ 39 ]	80	$\sim$ 1 h	33.71	14.37	0.13
Ours	51	$\sim$ 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.

Panel of Figure 16 — 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.

Panel of Figure 17 — 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.

3D Reconstruction

3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.

See all 955 papers

RT-GS: Gaussian Splatting with Reflection and Transmittance Primitives
Apr 1, 2026Kunnong Zeng, Chensheng Peng, Yichen Xie +2Differentiable Ray Tracing 3D Gaussian Splatting
University of Utah · University of California, Berkeley
RT-Splatting: Joint Reflection-Transmission Modeling with Gaussian Splatting
May 18, 2026Ji Shi, Xianghua Ying, Bowei Xing +23D Gaussian Splatting Novel View Synthesis
State Key Laboratory of General Artificial Intelligence School of Intelligence Science and Technology Peking University
TransmissiveGS: Residual-Guided Disentangled Gaussian Splatting for Transmissive Scene Reconstruction and Rendering
May 11, 2026Zhenyu Liang, Xiao Zhang, Tianchao Li +23D Gaussian Splatting 3D Reconstruction
HKUST

Apr 1, 2026J/K move · Enter open · S save

RT-GS: Gaussian Splatting with Reflection and Transmittance Primitives

Kunnong Zeng, Chensheng Peng, Yichen Xie, Masayoshi Tomizuka+1

University of Utah · University of California, Berkeley

Differentiable Ray Tracing 3D Gaussian Splatting Physically Based Rendering Differentiable Rendering

Open paper AskPDF arXiv

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.

Abstract

3D Reconstruction

Explore similar work

RT-GS: Gaussian Splatting with Reflection and Transmittance Primitives

RT-Splatting: Joint Reflection-Transmission Modeling with Gaussian Splatting

TransmissiveGS: Residual-Guided Disentangled Gaussian Splatting for Transmissive Scene Reconstruction and Rendering