GLINT: Modeling Scene-Scale Transparency via Gaussian Radiance Transport

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.

• The paper introduces GLINT, a hybrid rasterization and ray-tracing framework that separates interface, transmission, and reflection Gaussians to resolve the transparency–depth ambiguity in scene-scale reconstruction.
• The method bootstraps transparency masks from depth discrepancies and low diffuse albedo, avoiding segmentation annotations while using geometric priors to improve transparent-surface localization and reconstruction.
• GLINT achieves 34.50 PSNR, 7.96° normal error, and 0.04 depth AbsRel on 3D-FRONT-T, but remains limited to first-order transport and optically thin glass without refraction.

Problem statement and motivation

Transparent surfaces such as architectural glass, display cases, and windows violate the single-surface visibility assumption underlying most radiance field reconstruction methods. In standard 3D Gaussian Splatting (3DGS), a single pixel receives a superposition of reflected and transmitted radiance originating from distinct physical depths. The monolithic $\alpha$-blending formulation of 3DGS conflates these contributions into entangled Gaussian primitives, creating a transparency–depth dilemma: appearance optimization drives glass opacity toward zero (so the surface vanishes under adaptive pruning), while physically grounded glass geometry requires high opacity with sharp boundaries. The result is typically ghost-like reflections or missing transparent geometry.

Existing remedies are largely object-centric. TransparentGS and TSGS require segmentation masks or assume a single first surface, and reflection-focused methods such as GaussianShader, EnvGS, and DeferredGS model only opaque reflection paths, providing no mechanism for transmitted radiance. GLINT addresses scene-scale, optically thin transparency without segmentation masks by explicitly decomposing the radiance field.

Method

GLINT builds on 2DGS, whose perspective-correct splatting on local tangent planes yields geometrically accurate surfaces. The core representation partitions scene Gaussians into three functional sets:

• Interface ($\mathcal{G}_{\text{intr}}$): the primary first-visible surface, covering both opaque and transparent material boundaries; it produces a G-buffer $\mathcal{B} = \{z, \mathbf{n}, t, s\}$ encoding depth, normal, transparency, and specularity.
• Transmission ($\mathcal{G}_{\text{trans}}$): background geometry visible through transparent surfaces.
• Reflection ($\mathcal{G}_{\text{refl}}$): environment radiance reflected at opaque and transparent interfaces.

Rendering is hybrid: rasterization handles the primary interface, while an OptiX-based ray tracer (with Gaussians converted to triangle pairs for BVH traversal, following EnvGS) resolves secondary transmission and reflection queries. Outgoing radiance is a transparency-gated interpolation $$L_o = (1-t)\,L_{\text{opaque}} + t\,L_{\text{transparent}}$$, where each branch blends base/interface radiance with traced environment or transmitted radiance using a Fresnel-weighted specular factor $k_s = s + (1-s)F(\omega_o)$ with Schlick's approximation. The transparent branch assumes optically thin glass ($\omega_t \approx \omega_o$), so refraction bending is neglected — an assumption that restricts the method to thin glass rather than refractive media.

A notable contribution is bootstrapped transparency localization without segmentation masks. The method exploits the inter-component depth discrepancy $\Delta z = |z_{\text{intr}} - z_{\text{trans}}|$ that emerges during optimization where multiple depth layers coexist, combined with low diffuse-albedo values from a pre-trained video relighting model (DiffusionRenderer) indicating specular-dominant transport. Thresholding these cues yields a binary transparency mask supervising the learned transparency buffer. The authors show that Grounded-SAM-2 masks are noisy, fragmented, or missing for scene-scale glass, whereas the bootstrapped masks are spatially coherent and boundary-sharp.

Optimization uses a photometric loss (L1 + SSIM + LPIPS) plus geometric regularization (scale-and-shift-invariant depth loss and a thresholded cosine normal loss) from DiffusionRenderer's frame-consistent video-based priors. A multi-stage schedule warms up the interface component for 5k iterations, jointly optimizes all components, and freezes the interface between 40k–60k iterations while refining transmission and reflection.

Evaluation and results

The authors introduce 3D-FRONT-T, a synthetic benchmark built on 3D-FRONT with randomly placed thin glass containers rendered via Blender Cycles (4096 spp, 200 bounces), providing ground-truth depth and normals — the first synthetic benchmark for scene-scale transparency. Real-world evaluation uses 8 DL3DV-10K scenes with prominent transparent surfaces (photometric evaluation only, since ground-truth geometry is unavailable).

On 3D-FRONT-T, GLINT outperforms 2DGS, PGSR, Ref-GS, EnvGS, and TSGS across all geometry metrics: normal MAE of 7.96° (vs. 9.89° for TSGS and 14.37° for EnvGS), depth AbsRel of 0.04 (half of TSGS's 0.08), $\delta<1.25$ of 98.32, Chamfer Distance of 0.34 dm, and F1 of 0.836. It also achieves the best photometric results on both benchmarks (PSNR 34.50 on 3D-FRONT-T and 30.21 on DL3DV-10K). Notably, TSGS — the closest transparent-surface baseline — degrades substantially in rendering quality (28.80 PSNR on the synthetic set), indicating that its first-surface formulation recovers glass geometry but fails to model transmitted radiance. Ablations attribute the largest performance drop to removing the transmission component (PSNR falls to 32.26), confirming that separating transmitted background radiance is the dominant source of improvement; removing all geometric priors degrades normal MAE from 7.96° to 24.69° and AbsRel from 0.035 to 0.126.

Two supplementary analyses strengthen the case for explicit decomposition. First, retraining EnvGS with the same DiffusionRenderer normal priors paradoxically corrupts its rendering: without a transmission component, EnvGS normally compensates by baking background textures onto incorrect depths, and enforcing geometry blocks this mechanism, forcing multi-layer radiance onto a single interface with severe blurring. Second, the runtime comparison shows a deliberate cost trade-off: GLINT renders at 51 FPS and trains in ~2.5 hours, versus 80 FPS / ~1 h for EnvGS and 159 FPS / ~40 min for rasterization-only TSGS.

Limitations and open questions

The paper concedes several boundaries. The radiance transport formulation covers only first-order interactions at the interface; multi-bounce phenomena such as nested transparency (a glass vase inside a glass cabinet) are not modeled, and the authors identify recursive ray tracing as future work at increased computational cost. The decomposition implicitly relies on multi-view consistency: under sparse viewpoints or limited parallax, assigning radiance between reflection and transmission becomes ill-posed, and the authors suggest semantic priors (e.g., vision-LLMs) as a possible remedy. The optically thin assumption excludes refractive transparency, and the real-world evaluation lacks ground-truth geometry, so geometric claims on DL3DV-10K rest on qualitative evidence. The synthetic benchmark is also small (5 scenes).

Conclusion

GLINT demonstrates that explicitly decomposing Gaussian primitives into interface, transmission, and reflection components, combined with hybrid rasterization/ray-tracing transport and video-relighting priors, resolves the transparency–depth ambiguity that monolithic $\mathcal{G}_{\text{intr}}$0-blending cannot. The bootstrapped transparency localization removes the dependency on segmentation masks that limits prior transparent-surface methods, and the 3D-FRONT-T benchmark enables quantitative geometric evaluation of transparent reconstruction. The framework's restriction to first-order, optically thin transport defines the clear scope of its applicability.