Casual Flash Lighting for Gaussian Splat Inverse Rendering

Jiamin Xu; Dongheng Wei; Jiarong Zhao; Qi Wang; James Tompkin; Weiwei Xu; Gang Xu

cs.CVOct 5, 2026

Casual Flash Lighting for Gaussian Splat Inverse Rendering

Authors: Jiamin Xu, Dongheng Wei, Jiarong Zhao, Qi Wang, James Tompkin, Weiwei Xu, Gang Xu

Organizations: Hangzhou Dianzi University, China · North China Electric Power University, China · Brown University, USA · State Key Laboratory of CAD&CG, Zhejiang University, China

Abstract

Recovering geometry, materials, and lighting from photographs is highly ambiguous when only static illumination is available. Active-lighting setups reduce the ambiguity but require dark rooms or specialized hardware. Instead, we synergize both static and flash lighting from casual indoor capture, with the flash on or off, each from independent viewpoints. The flash residual constrains albedo and the BRDF, while static lighting captures grazing-angle specular highlights that flash misses. With a 2DGS reconstruction framing, our key contribution is a GS-anchored diffuse field: a hash-encoded MLP is queried at the rasterized 2DGS depth. As it depends only on world position, it is view consistent in 3D and allows the flash residual to drive material decomposition instead of being absorbed by alpha-blending drift across views. At the same time, we render static lighting with deferred shading such that it can also supervise material decomposition. On five synthetic and three real indoor scenes, our method outperforms six recent baselines on diffuse color, albedo and roughness material parameters, and in relighting where PSNR improves by 4.17 dB over the next-best baseline.

Figures & tables

Fig. 1: Our method synergizes both static and active lighting information for inverse rendering. By capturing a set of flash-on and flash-off images from different viewpoints, our approach achieves high-quality diffuse–specular decomposition and material estimation, and outperforms current state-of-the-art inverse rendering methods in relighting quality.

Fig. 2: Our pipeline. Given a set of relightable 2DGS with material properties, we use deferred rendering to synthesize diffuse and specular colors under static illumination, as well as under a flashlight’s direct and indirect shading components. The rendered results under static and static+flash lighting are then used for inverse rendering by computing a photometric loss. To facilitate diffuse–specular decomposition, we introduce an additional GS-anchored diffuse radiance field (GADF) to regularize the diffuse color in 2DGS. The dashed arrows indicate constraints enforced via loss terms.

Fig. 3: Our GS-anchored diffuse field constrains the diffuse color in 2DGS, reducing view-dependent effects such as highlights (red arrow) and improving the decomposition of diffuse and specular components.

Method	Diffuse Color			Relighting			Albedo			Roughness	Normal	Runtime
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MAE $\downarrow$	MAE $\downarrow$	Runtime
Ours	27.49	0.852	0.151	25.89	0.916	0.147	24.14	0.856	0.178	0.1554	0.0211	40 min
Ours-SAM2	26.85	0.838	0.185	24.38	0.895	0.166	22.96	0.843	0.174	0.1942	0.0258	40 min
FIPT [ 4 ]	–	–	–	20.96	0.888	0.180	19.65	0.780	0.207	0.2794	–	95 min
IRIS [ 5 ]	–	–	–	21.72	0.877	0.173	20.63	0.799	0.221	0.1848	–	95 min
WildLight [ 15 ]	–	–	–	20.11	0.794	0.331	16.96	0.640	0.507	0.2536	0.0758	480 min

TABLE I: Quantitative comparison. Blank entries indicate unavailable results.

Fig. 4: Comparison of synthetic scenes (1/2). Qualitative comparison of material decomposition and relighting. For the FIPT and IRIS experiments, which utilize the ground-truth mesh, we only present the ground-truth normals.

Fig. 5: Comparison of synthetic scenes (2/2). Continued from Fig. 4 .

Fig. 6: Comparison of real-world scenes. For the relighting results, we report both the color and type of lighting: PL for point light, AL for area light, and SL for spotlight. Note that the original light ball may appear white, but its radiance is not truly white due to overexposure when observing the light source directly. Please refer to the video for additional visual comparisons.

Models	Albedo $\uparrow$	Diffuse Color $\uparrow$	Roughness	Normal
Models	PSNR $\uparrow$	PSNR $\uparrow$	MAE $\downarrow$	MAE $\downarrow$
w/o flash	22.23	27.56	0.2568	0.0219
w/o GADF	23.97	15.64	0.1799	0.0207
w/o MultiEnvMap	24.00	27.55	0.1647	0.0212
Albedo&roughness NGP	23.36	27.26	0.2829	0.0213
Cam&flash baseline	23.69	27.13	0.1683	0.0216

TABLE II: Quantitative ablation on synthetic scenes, averaged over all scenes. We report PSNR for albedo and diffuse color, and MAE for roughness and surface normals.

Fig. 7: Qualitative comparison of the ablation studies.

Explore similar work

DR-GS: Physically-Based Deformable and Relightable 2D Gaussians
Jun 28, 2026Jiaxin Li, Tong Wu, Yi Wei +2Inverse Rendering 3D Gaussian Splatting
Shanghai Innovation Institute, China · Zhejiang University, China · Department of Artificial Intelligence, Westlake University, China +2
GLOW: Global Illumination-Aware Inverse Rendering of Indoor Scenes Captured with Dynamic Co-Located Light & Camera
Nov 28, 2025Jiaye Wu, Saeed Hadadan, Geng Lin +5Inverse Rendering
University of Maryland · University of North Carolina
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

Jun 28, 2026, cs.CVJ/K move · Enter open · S save

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

Jiaxin Li, Tong Wu, Yi Wei, Tailin Wu+1

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

Inverse Rendering 3D Gaussian Splatting 3D Reconstruction Dynamic Scene Reconstruction Deformable Shape Reconstruction Physically Based Rendering

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

GLOW: Global Illumination-Aware Inverse Rendering of Indoor Scenes Captured with Dynamic Co-Located Light & Camera

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