VolS-GS: Relightable Gaussian Splatting with Volumetric Subsurface Scattering

Junyeong Ahn; Jaegul Choo

cs.CVOct 2, 2026

VolS-GS: Relightable Gaussian Splatting with Volumetric Subsurface Scattering

Authors: Junyeong Ahn, Jaegul Choo

Organizations: KAIST AI

Abstract

We present VolS-GS, a relightable Gaussian splatting framework that reconstructs objects from one-light-at-a-time (OLAT) captures and renders them under novel lighting and viewpoints. Relightable Gaussian Splatting methods typically model appearance independently at each primitive, which makes non-local effects difficult to represent. This limitation is particularly apparent for subsurface scattering, where light entering the object at one location can emerge at another. Rather than modeling this effect solely with a neural network or a local kernel at each primitive, we use the spatial support of the Gaussian scene as the domain of a differentiable finite-volume transport solver, so that light can propagate through the object's interior. A small network predicts scattering and absorption coefficients for each Gaussian, and the solve redistributes incident light through the resulting field. The coefficients are fit to images rather than measured, so the solve supplies a transport-shaped path for aggregating per-primitive appearance, not a measurement of the material. To keep the learned shadow and specular terms from taking over the other components, our shadow term is predicted from visibility together with the transmittances and the scattering the solve produces, and a regularizer suppresses specular highlights in regions the shadow term predicts to be unlit. Experiments on three OLAT benchmarks show that VolS-GS consistently improves relighting quality on held-out lights and views.

Figures & tables

Figure 1: VolS-GS proposes a high-quality relighting pipeline with three different modalities. Center: a rendered image can be decomposed into normal, diffuse, specular, subsurface scattering (SSS) and shadow. Left: VolS-GS faithfully synthesizes novel view renders. Right: relighting.

Figure 2: Per-primitive kernels versus transport through the medium. Top: where each method computes its subsurface response. Bottom: a C-shaped medium lit at the opening. The dipole depends on Euclidean distance, so its support crosses the gap, while our solve aggregates the subsurface response through the learned medium.

Figure 3: Overview of VolS-GS. (A) A render is the sum of a shadowed surface response and a subsurface term, compared against the captured frame. (B) The shadow is predicted from two cues gathered along the ray to the light, the fraction the Gaussians block and the fraction the medium transmits. (C) The subsurface term is propagated rather than predicted per primitive. The Gaussians define a domain $\Omega$ , the light entering it forms a source $b$ , one linear solve $F\phi=b$ spreads that source through the medium, and each Gaussian reads the result back as outgoing radiance, so light can enter in one place and leave in another.

	Method	Bunny	Dragon	Candle	Soap	Statue	Crystal	JoyfulYell	Physalis	NRHints	GS 3
										dataset mean
PSNR $\uparrow$	GS 3	34.7109	36.1753	38.3707	38.3713	35.3291	35.2767	31.5467	28.9163	31.94	31.11
	SSS-GS	36.0035	34.7397	38.3432	36.4006	35.5485	34.6201	31.4973	28.9331	25.16	28.23
	RNG	35.3300	34.6134	40.3420	38.4962	29.2148	33.6849	30.0986	28.6492	30.74	28.64
	SSD-GS	38.9147	37.7832	41.9214	38.2959	35.9034	35.2096	31.8938	29.5474	32.05	31.01
	VolS-GS (Ours)	40.1453	38.0579	43.6366	41.0098	38.3515	35.6278	32.1856	30.2210	32.37	31.59

Table 1: Main results. Left: the eight subsurface scenes, per scene, every method at the

60

k protocol iteration. Right: the remaining two datasets as test-split means over their

7

and

6

scenes (per-scene numbers trained at

100

k in Appendix D.1 ). Best red , second orange .

Figure 4: Qualitative relighting comparison under novel lighting across different scenes. Translucent and Lego are synthetic scenes from GS 3 , while Pixiu , Cat , and Pikachu are real captures from NRHints. Per-frame PSNR/SSIM/LPIPS are printed on each panel, best in bold and second underlined within a row. The inset is the error map $\lvert\text{render}-\text{GT}\rvert$ averaged over color channels and clipped at $0.15$ , on one ramp shared by every panel.

Figure 5: Material decomposition. Our own channels on two real captures, Pixiu (top) and Crystal (bottom), one row per scene. The specular column is amplified only where it is already bright ( $\times 2.5$ above its own 95th percentile), identically in both rows.

	Bunny			Statue
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Full model	40.15	0.9920	0.0112	38.35	0.9899	0.0130
$-$ Volumetric SSS	34.67	0.9812	0.0283	36.19	0.9852	0.0173
$-$ Transport cues on shadow	39.17	0.9910	0.0126	35.09	0.9818	0.0211
Shadow term multiplied onto SSS	39.02	0.9901	0.0132	37.66	0.9885	0.0144
$-$ Shadow term	36.30	0.9859	0.0179	34.41	0.9794	0.0230

Table 2: Ablations on Bunny and Statue where SSS and shadow are prevalent. Each row retrains the final configuration with one component changed.

Figure 6: Ablation study. Material decomposition on Bunny (left) and Statue (right) from the SSS-GS dataset. Rows are the full model and the four single-change variants of Table 2 .

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

dataset	res.	scenes	FPS
SSS-GS synthetic	$256^{2}$	5	$64.5\pm 4.9$
SSS-GS real	$512^{2}$	3	$66.0\pm 10.2$
NRHints real	$512^{2}$	4	$46.0\pm 2.5$
	$1024^{2}$	3	$54.8\pm 11.7$
GS 3 synthetic	$512^{2}$	6	$32.9\pm 11.3$

Appendix

Table 3: Render cost. A single run on one A6000, reported per dataset and per resolution.

	Method	Cat	CatSmall	CupFabric	Fish	FurScene	Pikachu	Pixiu
PSNR $\uparrow$	3DGS	14.5326	22.5727	25.0488	22.8411	18.4838	19.6310	18.5501
	GI-GS	13.9988	22.3222	24.3821	22.7500	17.8520	19.1867	18.1064
	GS 3	27.7597	34.6288	36.3707	31.1387	30.5467	32.6480	30.4776
	SSS-GS	16.6533	31.4044	32.9904	22.8345	20.4048	28.4206	23.4290
	RNG	26.7680	34.2488	37.1390	29.2007	27.8608	31.1881	28.7688
	SSD-GS	27.9527	34.6987	36.9159	31.0531	30.6964	32.1040	30.9603

Appendix

Table 4: Real captures, per scene (NRHints). The 3DGS and GI-GS rows are quoted from Zheng et al. (2026) . Every other row is our own retraining to the

100

k protocol iteration. Best red , second orange among retrained rows.

Figure 7: Additional qualitative comparisons. Five further test frames, one per row: Soap and JoyfulYell from the subsurface set, Drums from the GS 3 synthetic set, FurScene and Fish real captures. Per-frame PSNR/SSIM/LPIPS are printed on each panel, best bold and second underlined; the inset is $\lvert\text{render}-\text{GT}\rvert$ clipped at $0.15$ on one shared ramp.

	Method	Translucent	AnisoMetal	Drums	FurBall	Hotdog	Lego
PSNR $\uparrow$	3DGS	16.4899	17.1009	24.5093	20.1206	16.9535	15.9886
	GI-GS	16.0766	15.9567	24.6177	19.5295	16.8118	16.4229
	GS 3	29.7079	28.5958	32.9349	34.0920	32.2632	29.0746
	SSS-GS	28.2425	24.2328	28.4225	31.0156	30.8217	26.6436
	RNG	27.1813	26.1441	29.7962	33.0337	29.8325	25.8664
	SSD-GS	29.9792	28.2052	33.0411	34.1532	31.6721	28.9808

Appendix

Table 5: GS 3 synthetic scenes, per scene trained at

100

K iterations. Notation as in Table 4 .

Figure 8: Relighting under a continuous light sweep. Eight unseen light positions per scene, one per column, with the disc below each column giving its light direction.

Figure 9: Relighting under an environment map. One row per map, shown as a mirror ball at the left; two frames of a full rotation of the map for each of four scenes. Nothing is retrained.

	Statue	Bunny	JoyfulYell	Translucent	Drums	CupFabric	Pikachu
Full model	38.35	40.15	32.19	31.64	33.21	37.02	32.96
$-$ Volumetric SSS	36.19	34.67	31.54	30.88	33.11	35.78	31.43
$-$ Transport cues on shadow	35.09	39.17	32.08	31.01	32.74	36.41	32.19
Shadow term multiplied onto SSS	37.66	39.02	31.87	31.15	32.93	36.82	32.46
$-$ Shadow term	34.41	36.30	29.24	28.42	32.47	34.07	28.81

Appendix

Table 6: Ablations on further scenes (test PSNR

\uparrow

). The same five rows as Table 2 , whose Statue and Bunny PSNR columns are repeated here, extended to five further scenes. Every arm is read at the protocol iteration for its family,

60

k on the SSS-GS scenes and

100

k elsewhere.

Panel of Figure 10 — Figure 10: Two reads of the subsurface slot. (a) Candle and Pikachu with the specular leakage regularizer on (Ours) and off, same frame and channels. Without it the specular lobe paints a ghost of the object into shadowed regions and the diffuse term washes out. The Leakage column marks in red the pixels where the model’s own decay says the light is blocked ( $D_{i}<0.3$ ) while the specular lobe still fires above $0.05$ . (b) Drums and Bunny with our solve (Ours) and with SSD-GS’s per-primitive dipole in the same slot, everything else held fixed. Drums shows the channel taking up surface residual, Bunny the transported light misplaced. The SSS column of (b) is shown on its own black ground and at a $2.5\times$ gain so that both rows are legible.

	VolS-GS			Dipole SSS
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Bunny	40.15	0.9920	0.0112	39.52	0.9911	0.0119
Statue	38.35	0.9899	0.0130	36.53	0.9862	0.0165
JoyfulYell	32.19	0.9690	0.0385	31.68	0.9674	0.0408
Translucent	31.64	0.9783	0.0246	31.13	0.9778	0.0251
CupFabric	37.02	0.9823	0.0516	34.61	0.9744	0.0550

Appendix

Table 7: The dipole kernel in place of the transport solve. Only the subsurface slot changes: SSD-GS’s own scattering module and its

\sum_{\ell}1/d_{\ell}^{2}

carrier replace our solve, while the transport cues on the shadow branch, the specular regularizer, the exit term and the per-primitive amplitude are left as in the full model.

Figure 11: Decomposition across methods. Pixiu (left) and Statue (right), each row one method’s own decomposition on the same frame. For SSS-GS the SSS column shows its residual network’s output, its only channel with appreciable energy.

Figure 12: Failure cases. Bunny with the ear folded back and seen edge-on, where the tip renders darker than the capture, and Hotdog , where the plate under the bun renders darker. Under each render the magnified square marks the failing region at the same image coordinates in every column, beside the error map $\lvert\text{render}-\text{GT}\rvert$ averaged over color channels and clipped at $0.15$ on one shared ramp. Per-frame PSNR/SSIM/LPIPS are printed on each panel, best in bold and second underlined within a row.

Explore similar work

May 9, 2026cs.CV

Relightable Gaussian Splatting for Virtual Production Using Image-Based Illumination

Virtual production (VP) use LED walls to provide both background imagery and image-based lighting. While this enables on-set compositing, it couples lighting to background and scene appearance, limiting flexibility for downstream editing. In addition, inverse rendering conventionally relies on physically-based rendering to estimates 3D geometry and lighting, using environment maps. However, these maps are typically low-resolution and assume far-field lighting. In VP, with near-field and high-resolution image-based lighting, this can lead to inaccuracies and introduce complexities when editing. Addressing this, we propose a VP-specific framework for 3D reconstruction and relighting using Gaussian Splatting. This uses the known background imagery to condition the relighting process. This avoids relying on environment maps and reduces compositing to a background-image editing task. To realize our framework, we introduce a process (and associated dataset) that captures real VP scenes under varying background content and illumination conditions. This data is used to decompose a 3D scene into fixed appearance and variable lighting components. The variable lighting process simulates light transport by parameterizing each primitive with a UV coordinate, intensity value and resolution modifier. Using mipmaps, these directly sample the background texture in image space - implicitly capturing reflections and refractions without physically-based rendering. Combined with the fixed appearance component, this allows us to render relit scenes using a Gaussian Splatting rasterizer. Compared to baselines, our approach achieves higher-quality 3D reconstruction and controllable relighting. The method is efficient (<3 GB RAM, <5 GB VRAM, <2 hours training, ~35 FPS) and supports rendering useful arbitrary output variables including depth, lighting intensity, lighting color, and unlit renders.

Adrian Azzarelli, Nantheera Anantrasirichai, James Pollock +1

University of Bristol, UK · Lux Aeterna, Bristol, UK

Inverse Rendering Illumination Variations Gaussian Splatting Relighting

Ask PDF

Apr 18, 2026cs.CV

Instant Colorization of Gaussian Splats

Gaussian Splatting has recently become one of the most popular frameworks for photorealistic 3D scene reconstruction and rendering. While current rasterizers allow for efficient mappings of 3D Gaussian splats onto 2D camera views, this work focuses on mapping 2D image information (e.g. color, neural features or segmentation masks) efficiently back onto an existing scene of Gaussian splats. This 'opposite' direction enables applications ranging from scene relighting and stylization to 3D semantic segmentation, but also introduces challenges, such as view-dependent colorization and occlusion handling. Our approach tackles these challenges using the normal equation to solve a visibility-weighted least squares problem for every Gaussian and can be implemented efficiently with existing differentiable rasterizers. We demonstrate the effectiveness of our approach on scene relighting, feature enrichment and 3D semantic segmentation tasks, achieving up to an order of magnitude speedup compared to gradient descent-based baselines.

Daniel Lieber, Alexander Mock, Nils Wandel

University of Osnabrück

Gaussian Splatting Differentiable Rendering Image Relighting

Ask PDF

Jun 28, 2026cs.CV

DR-GS: Physically-Based Deformable and Relightable 2D Gaussians

Gaussian splatting (GS) has garnered significant attention in VR/AR and digital content creation due to its explicit parameterization and efficient rendering capabilities. However, existing GS-based methods for deformable objects face two key limitations: (i) illumination is erroneously baked into textures, causing physically inconsistent responses under dynamic deformations and lighting changes; (ii) snapshot-based reconstruction restricts post-reconstruction material editing. To address these challenges, we propose Deformable and Relightable GS (DR-GS), a unified Gaussian framework that integrates physically-based inverse rendering, relighting, and deformation-aware manipulation. Through explicitly disentangling geometry, illumination, and material representations, DR-GS overcomes the limitations of static snapshots, resolving unrealistic appearance under varying conditions while enabling post-reconstruction parameter editing. Extensive experiments show that DR-GS achieves leading visual quality across static reconstruction, dynamic deformation, and relighting, reliably preserving reflections and specular highlights on glossy surfaces. It further establishes a fully decoupled geometry-illumination-material pipeline, enabling high-quality 3D asset creation and comprehensive post-editing.

Jiaxin Li, Tong Wu, Yi Wei +2

Shanghai Innovation Institute, China · Zhejiang University, China · Department of Artificial Intelligence, Westlake University, China +2

Inverse Rendering Gaussian Splatting 3D Gaussian Primitives Deformable Cnn Deformable 3D Assets

Ask PDF