Radiometrically Consistent Gaussian Surfels for Inverse Rendering

Kyu Beom Han; Jaeyoon Kim; Woo Jae Kim; Jinhwan Seo; Sung-eui Yoon

cs.CVMar 2, 2026

Radiometrically Consistent Gaussian Surfels for Inverse Rendering

Authors: Kyu Beom Han, Jaeyoon Kim, Woo Jae Kim, Jinhwan Seo, Sung-eui Yoon

Organizations: School of Computing, Korea Advanced Institute of Science and Technology, Daejeon, South Korea

Abstract

Inverse rendering with Gaussian Splatting has advanced rapidly, but accurately disentangling material properties from complex global illumination effects, particularly indirect illumination, remains a major challenge. Existing methods often query indirect radiance from Gaussian primitives pre-trained for novel-view synthesis. However, these pre-trained Gaussian primitives are supervised only towards limited training viewpoints, thus lack supervision for modeling indirect radiances from unobserved views. To address this issue, we introduce radiometric consistency, a novel physically-based constraint that provides supervision towards unobserved views by minimizing the residual between each Gaussian primitive's learned radiance and its physically-based rendered counterpart. Minimizing the residual for unobserved views establishes a self-correcting feedback loop that provides supervision from both physically-based rendering and novel-view synthesis, enabling accurate modeling of inter-reflection. We then propose Radiometrically Consistent Gaussian Surfels (RadioGS), an inverse rendering framework built upon our principle by efficiently integrating radiometric consistency by utilizing Gaussian surfels and 2D Gaussian ray tracing. We further propose a finetuning-based relighting strategy that adapts Gaussian surfel radiances to new illuminations within minutes, achieving low rendering cost (<10ms). Extensive experiments on existing inverse rendering benchmarks show that RadioGS outperforms existing Gaussian-based methods in inverse rendering, while retaining the computational efficiency.

Figures & tables

Figure 1: We introduce RadioGS, a novel inverse rendering framework that models accurate indirect illumination by providing a novel physically-based supervision on unobserved directions. (a) Compared to existing Gaussian-based methods ( Gu et al., 2024 ; Sun et al., 2025 ) , our method provides realistic inter-reflection between the red bulb and the blobs on the yellow lego surface, (b) leading to robust decomposition of scene properties. (c) Our method can also generate realistic indirect illumination on new lighting conditions for real objects from Stanford-ORB dataset ( Kuang et al., 2023 ) .

Figure 2: Overview of our RadioGS. Left: Our radiometric consistency loss $\mathcal{L}_{rad}$ provides physically-based supervision on indirect radiances from views unobserved by image reconstruction loss $\mathcal{L}_{recon}$ , by enforcing consistency between surfel radiance $L_{\mathbf{G}}$ and physically-based rendered (PBR) radiance $L_{\mathbf{G}}^{\mathbf{PBR}}$ of Gaussian surfels. Radiometric consistency is seamlessly integrated into the inverse rendering framework, guiding Gaussian surfels to obtain physically-based radiance for delivering realistic indirect radiance to other surfels. 2D Gaussian ray tracing is deployed to jointly optimize ray-traced Gaussian surfels with our radiometric consistency loss. Right: Black-dotted arrows show NVS supervision, which leaves the occluded green Gaussian unconstrained. Pink-dotted arrows show our radiometric consistency providing additional supervision on its outgoing radiance along unseen directions (e.g., towards other Gaussian surfels).

Method	Novel View Synthesis			Normal	Albedo			Relight			Training
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MAE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	hours
TensoIR	35.09	0.976	0.040	4.100	29.27	0.950	0.085	28.58	0.944	0.081	4
GS-IR	35.33	0.974	0.039	4.948	29.94	0.921	0.100	24.37	0.885	0.096	0.5
GI-GS	36.75	0.972	0.037	5.253	29.90	0.921	0.099	24.70	0.886	0.106	0.5
R3DG	33.35	0.964	0.041	5.927	29.27	0.951	0.078	27.37	0.909	0.083	1.1
IRGS	35.43	0.964	0.049	4.209	30.62	0.956	0.072	29.91	0.935	0.076	0.9

Table 1: Quantitative comparisons on TensoIR dataset ( Jin et al., 2023 ) . The results are colored in rank as 1st , 2nd , and 3rd . Our method surpasses existing Gaussian-based methods and a NeRF-based method in most metrics, while maintaining the computational efficiency with the average training time of 1 hour. We report our relighting metric using Gaussian ray tracing (Ours) and finetuning-based method (Ours*).

Figure 3: Qualitative result on the “lego” scene of TensoIR dataset . Our method provides enhanced decomposition and realistic relighting results compared to Gaussian-based methods. Specifically, our method shows noticeably robust performance on regions with high geometric complexity, such as the highlighted bucket. Best viewed in zoom.

Figure 4: Relighting results on the “armadillo” scene of TensoIR dataset.

Method	NVS	Roughness	Albedo	Relight
Method	PSNR $\uparrow$	MSE $\downarrow$	PSNR $\uparrow$	PSNR $\uparrow$
R3DG	34.10	0.010	28.65	33.12
IRGS	34.44	0.008	30.50	34.35
SVG-IR	34.14	0.009	29.06	32.59
Ours	34.98	0.011	30.69	34.87

Table 2: Quantitative comparisons on Synthetic4Relight dataset.

Figure 5: Qualitative results on the “hotdog” scene of Synthetic4Relight ( Zhang et al., 2022 ) dataset. Our method models natural inter-reflection between the sausages and the buns, showing superior reconstruction performance on highlighted regions. IRGS shows relatively bright and fluctuating indirect illumination, which led to darker albedo reconstruction. SVG-IR models relatively darker indirect illumination, returning brighter albedo reconstruction. Best viewed in zoom.

Figure 6: Ablation studies on our radiometric consistency. The left sub-figure demonstrates how our radiometric consistency loss $\mathcal{L}_{rad}$ provides guidance on radiances towards unobserved views such as the interstices, leading to enhanced albedo reconstruction (red box). Also, our method guides the generation of inter-reflections between the ketchup and the plate (yellow box). The right table contains PSNR metrics for the ablation studies.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Relight			Rendering
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	ms
IRGS	29.91	0.935	0.076	1090
SVG-IR	31.10	0.946	0.056	82.48
PBR split-sum	32.09	0.953	0.048	38.64
PBR finetuned	31.59	0.952	0.049	38.29
Surfel finetuned	31.41	0.948	0.052	5.902

Appendix

Table 3: Relighting Performance and Rendering cost during relighting on TensoIR dataset.

Figure 7: Qualitative comparison on illumination components on the “hotdog” scene of TensoIR dataset. Best viewed in zoom.

Figure 8: Qualitative comparison on illumination components on the“lego” scene of TensoIR dataset. Best viewed in zoom.

Figure 9: Qualitative comparison on NVS, albedo reconstruction, normal reconstruction, and relighting on the “armadillo” scene of the TensoIR dataset. Best viewed in zoom.

Figure 10: Qualitative comparison on NVS, albedo reconstruction, roughness reconstruction, and relighting on the “jugs” scene of the Synthetic4Relight dataset. Best viewed in zoom.

Figure 11: Qualitative comparison on NVS, albedo reconstruction, roughness reconstruction, and relighting on the “air baloons” scene of the Synthetic4Relight dataset. Best viewed in zoom.

Figure 12: Qualitative comparison on NVS, albedo reconstruction, roughness reconstruction, and relighting on the “chair” scene of the Synthetic4Relight dataset. Best viewed in zoom.

Method	NVS			Indirect Illumination			Geometry	Albedo
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Normal MAE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
IRGS	35.0982	0.9660	0.0436	24.2219	0.8792	0.1092	4.1835	30.0931	0.9521	0.0752
SVG-IR	36.9634	0.9786	0.0266	30.9747	0.9134	0.0843	4.2624	29.7354	0.9309	0.0822
Ours w/o $\mathcal{L}_{rad}$	36.3471	0.9764	0.0276	30.0954	0.9161	0.0752	3.8332	30.3425	0.9470	0.0760
Ours	37.8519	0.9822	0.0212	32.8832	0.9266	0.0726	3.6048	30.6224	0.9502	0.0744

Appendix

Table 4: Quantitative comparison against baselines (IRGS, SVG-IR) and our ablation model on our new dataset. Our method significantly outperforms all baselines in indirect illumination reconstruction.

Figure 13: Qualitative comparison on novel-view synthesis and indirect illumination on our dataset. Best viewed in zoom.

Figure 14: Qualitative comparison on indirect illumination for four different lighting conditions on the “hotdog” scene of our dataset. Best viewed in zoom.

Method	NVS			Geometry	Albedo			Relighting
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Normal MAE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Split-sum	36.659	0.9800	0.0254	4.0066	27.471	0.9339	0.0828	28.816	0.9371	0.0575
RT Precompute	35.547	0.9728	0.0332	3.7138	30.523	0.9481	0.0750	31.965	0.9516	0.0496
RT w/o Diff.	37.252	0.9789	0.0275	3.6946	30.673	0.9507	0.0732	32.050	0.9524	0.0484
Ours	37.858	0.9801	0.0266	3.6889	31.048	0.9523	0.0719	32.092	0.9533	0.0478

Appendix

Table 5: Ablation study on radiometric consistency strategies. We compare our method with baselines using different indirect illumination handling. Best results are highlighted in bold .

Train views	NVS PSNR		Indirect PSNR
Train views	Ours	Ours w/o $\mathcal{L}_{rad}$	Ours	Ours w/o $\mathcal{L}_{rad}$
100%	37.85	36.35	32.88	30.10
50%	37.54 (-0.31)	35.81 (-0.54)	32.79 (-0.09)	29.15 (-0.95)
25%	36.79 (-1.06)	34.65 (-1.70)	32.71 (-0.17)	27.89 (-2.21)

Appendix

Table 6: Ablation study on training view scarcity. We report NVS and Indirect PSNR metrics across different subsets of training views. Values in parentheses denote the performance drop relative to the 100% setting.

$N_{g}$	NVS			Geometry	Albedo			Relight			Render
$N_{g}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Normal MAE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	(ms)
1024	37.8206	0.9799	0.0268	3.6900	30.9137	0.9518	0.0730	32.0569	0.9529	0.0481	38.5
2048 (Ours)	37.8580	0.9801	0.0266	3.6889	31.0479	0.9521	0.0721	32.0920	0.9533	0.0478	38.6
4096	37.8707	0.9802	0.0264	3.6852	31.0495	0.9522	0.0721	32.1112	0.9532	0.0478	38.7
8192	37.8799	0.9802	0.0263	3.6834	31.0507	0.9523	0.0720	32.1294	0.9533	0.0477	38.5

Appendix

Table 7: Ablation study on the number of surfels (

N_{g}

) for radiometric consistency. We vary

N_{g}

while fixing

N_{s}=64

. Increasing

N_{g}

improves quality without affecting rendering cost.

$N_{s}$	NVS			Geometry	Albedo			Relight			Render
$N_{s}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Normal MAE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	(ms)
16	37.4199	0.9792	0.0275	3.7174	30.8656	0.9505	0.0735	32.1684	0.9518	0.0494	16.2
32	37.8125	0.9799	0.0268	3.6979	30.9014	0.9510	0.0730	32.1537	0.9528	0.0483	22.6
64 (Ours)	37.8707	0.9802	0.0264	3.6852	31.0479	0.9521	0.0721	32.1112	0.9532	0.0478	38.6
128	37.6548	0.9798	0.0268	3.6821	30.9563	0.9512	0.0728	31.9768	0.9526	0.0480	58.3

Appendix

Table 8: Ablation study on the number of secondary ray samples (

N_{s}

). We vary

N_{s}

while fixing

N_{g}=2048

. Performance saturates around

N_{s}=64

$N_{g}$	Average (ms)	Bonsai (ms)	Counter (ms)	Kitchen (ms)	Room (ms)
1024	81.23	77.37	72.38	86.81	88.43
2048 (Ours)	81.40	77.78	72.95	84.29	90.56
4096	81.69	78.87	72.19	85.28	90.41
8192	81.53	77.75	73.34	85.22	89.82
16384	81.86	77.75	73.24	86.08	90.38

Appendix

Table 9: Rendering cost analysis on MipNeRF360 with varying

N_{g}

. We fix

N_{s}=64

. Consistent with TensoIR,

N_{g}

does not significantly affect rendering speed.

$N_{s}$	Average (ms)	Bonsai (ms)	Counter (ms)	Kitchen (ms)	Room (ms)
16	24.57	23.49	23.86	26.07	24.85
32	43.43	42.00	39.32	47.31	45.10
64 (Ours)	81.40	77.78	72.95	84.29	90.56

Appendix

Table 10: Rendering cost analysis on MipNeRF360 with varying

N_{s}

. We fix

N_{g}=2048

. Reducing

N_{s}

significantly decreases rendering time, allowing for trade-offs between speed and quality.

Figure 15: Qualitative results on the MipNeRF360 dataset. Best viewed in zoom.

Method	NVS PSNR $\uparrow$	Albedo PSNR $\uparrow$	Relight PSNR $\uparrow$
NVS init. + $\lambda_{\text{rad}}=0$ during IR	35.70	30.31	31.51
NVS init.	37.43	30.34	31.66
Ours w/o $\mathcal{L}_{rad}$ ( $\lambda_{\text{rad}}=0$ during IR)	35.82	30.82	31.69
Ours	37.86	31.05	32.09

Appendix

Table 11: Ablation study on initialization strategies. We compare our initialization (Ours) against NVS-based initialization baselines. “NVS init.” denotes NVS pre-training followed by our standard optimization. “during IR” refers to the inverse rendering optimization stage.

Figure 16: Ablation study of our initialization method on the “hotdog” scene of TensoIR dataset.

Figure 17: Qualitative comparison on reconstructed environment map on TensoIR dataset. Best viewed in zoom.

Figure 18: Qualitative comparison on reconstructed environment map on Synthetic4Relight dataset. Best viewed in zoom.

Figure 19: Illustrative figure on how our finetuning-based relighting adapts surfel radiances for new lighting conditions.

Figure 20: Qualitative ablation study on three relighting method: 1) Gaussian ray tracing that queries indirect radiance via split-sum approximation“PBR render (Split-sum)”, 2) Gaussian ray tracing that queries indirect radiance from finetuned surfels “PBR render (Finetuned)”, and 3) direct rasterization with finetuned surfel radiances “Surfel render (Finetuned)”.

Method	Sample Count ( $N$ )	Rendering (ms) $\downarrow$	VRAM (MB) $\downarrow$
Ours (Finetuned)	-	5.90	308.2
Baseline (Ray Tracing)	64	38.64	1512.6
	128	58.31	2523.3
	256	86.99	4589.5

Appendix

Table 12: Comparison of rendering speed and memory usage. We compare our finetuned model against the baseline relighting method with varying ray tracing sample counts (

N

). Our method achieves significantly lower rendering time and memory footprint.

Inverse Rendering

Inverse rendering estimates scene geometry, materials, lighting, and other causes of images from observed visual or sensor data.

See all 69 papers

Inter-Reflective Gaussian Splatting for Robust and Efficient Inverse Rendering
Jul 24, 2026Chun Gu, Xiaofei Wei, Zixuan Zeng +2Inverse Rendering 3D Gaussian Splatting
School of Data Science, Fudan University, and Shanghai Innovation Institute · Shanghai Innovation Institute · School of Data Science, Fudan University
GAINS: Gaussian-based Inverse Rendering from Sparse Multi-View Captures
Dec 10, 2025Patrick Noras, Jun Myeong Choi, Didier Stricker +2Inverse Rendering 3D Gaussian Splatting
University of Kaiserslautern-Landau, Kaiserslautern, Germany · German Research Center for Artificial Intelligence, Kaiserslautern, Germany · University of North Carolina at Chapel Hill, NC, USA +1
VisibilitySplat: Ray-Traced Occlusion for Relighting with 3D Gaussian Splatting
Oct 14, 2026Sven Kluge, Oliver StaadtInverse Rendering 3D Gaussian Splatting
University of Rostock, Albert-Einstein-Straße 22, 18059 Rostock, Germany

Jul 24, 2026J/K move · Enter open · S save

Inter-Reflective Gaussian Splatting for Robust and Efficient Inverse Rendering

Chun Gu, Xiaofei Wei, Zixuan Zeng, Yuxuan Yao+1

School of Data Science, Fudan University, and Shanghai Innovation Institute · Shanghai Innovation Institute · School of Data Science, Fudan University

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

Open paper AskPDF arXiv

Faithful inverse rendering requires visibility and indirect radiance to explain secondary illumination and inter-reflection, yet rasterization-oriented Gaussian representations do not naturally support the secondary-ray queries needed to recover them. We present IRGS++ (Inter-Reflective Gaussian Splatting), a unified robust and efficient Gaussian inverse rendering framework. During transport-aware optimization, IRGS++ employs differentiable 2D Gaussian ray tracing on surface-oriented Gaussian primitives to query visibility and indirect radiance on the fly and evaluate the full rendering equation for inter-reflective transport. This physical core makes Gaussian inverse rendering physically grounded beyond rasterized appearance modeling. To make this backbone useful beyond low-gloss dielectric scenes, the framework incorporates metallic-aware material modeling and robust reflective initialization for glossy, specular, and metallic materials. To make it practical, multiple importance sampling and denoising stabilize finite-sample rendering, while mesh-based secondary-attribute queries reduce the cost of relighting under novel illumination. Quantitative evaluations on low-gloss and glossy benchmarks show improved decomposition and relighting quality together with favorable quality--speed trade-offs under the reported configurations, while real-world studies illustrate plausible relighting under novel illumination.

Abstract

Inverse Rendering

Explore similar work

Inter-Reflective Gaussian Splatting for Robust and Efficient Inverse Rendering

GAINS: Gaussian-based Inverse Rendering from Sparse Multi-View Captures

VisibilitySplat: Ray-Traced Occlusion for Relighting with 3D Gaussian Splatting