Physically Inspired Gaussian Splatting for HDR Novel View Synthesis

Huimin Zeng; Yue Bai; Hailing Wang; Yun Fu

cs.CVMar 30, 2026

Physically Inspired Gaussian Splatting for HDR Novel View Synthesis

Authors: Huimin Zeng, Yue Bai, Hailing Wang, Yun Fu

Organizations: Department of Electrical and Computer Engineering, Northeastern University · Khoury College of Computer Science, Northeastern University

Abstract

High dynamic range novel view synthesis (HDR-NVS) reconstructs scenes with dynamic details by fusing multi-exposure low dynamic range (LDR) views, yet it struggles to capture ambient illumination-dependent appearance. Implicitly supervising HDR content by constraining tone-mapped results fails in correcting abnormal HDR values, and results in limited gradients for Gaussians in under/over-exposed regions. To this end, we introduce PhysHDR-GS, a physically inspired HDR-NVS framework that models scene appearance via intrinsic reflectance and adjustable ambient illumination. PhysHDR-GS employs a complementary image-exposure (IE) branch and Gaussian-illumination (GI) branch to faithfully reproduce standard camera observations and capture illumination-dependent appearance changes, respectively. During training, the proposed cross-branch HDR consistency loss provides explicit supervision for HDR content, while an illumination-guided gradient scaling strategy mitigates exposure-biased gradient starvation and reduces under-densified representations. Experimental results across realistic and synthetic datasets demonstrate our superiority in reconstructing HDR details (e.g., a PSNR gain of 2.04 dB over HDR-GS), while maintaining real-time rendering speed (up to 76 FPS). Code and models are available at https://huimin-zeng.github.io/PhysHDR-GS/.

Figures & tables

Figure 1: Variation of camera exposure $\Delta t$ and ambient illumination $\Delta L_{a}$ scales the HDR signal in different ways: $\Delta t$ causes a global change $\Delta I_{HDR}$ , while $\Delta L_{a}$ induces local changes $\Delta\hat{I}_{HDR}$ ( e.g . , nameplate of the luckycat) via lighting-conditioned radiance variation $\Delta L_{o}$ . Their different response patterns reveal complementary ways of modeling dynamic-range details.

Figure 2: Overview of the proposed PhysHDR-GS, where Gaussian color is modeled from intrinsic reflectance and ambient illumination. The image–exposure (IE) branch modulates exposure $t$ on 2D images, while the Gaussian–illumination (GI) branch modulates ambient illumination $L_{a}$ on 3D Gaussians, yielding complementary dynamic-range details. Tone mapper $f$ performs tone mapping and dual-branch fusion for final LDR results. During training, a cross-branch HDR consistency loss $\mathcal{L}_{cons}$ enables explicit HDR self-supervision. Illumination-Guided Gradient Scaling rescales per-Gaussian gradients with $s_{a}$ to mitigate under-splitting in extreme exposure regions.

Figure 3: Illustration of the tone mapper $f$ . Given inputs $I_{HDR}\times t$ and $\hat{I}_{HDR}$ , the tone-mapping MLP $f_{tm}$ first predicts global and local LDR outputs. The fusion MLP $f_{mix}$ then cross-fuses these global-local pairs to produce the final LDR result $I_{LDR}$ .

Figure 4: Gradient and illumination deviation analysis, where over/under-exposed pixels lie in flat regions of the tone mapping curve and yield a small Gaussian gradient. The gradient shows positive correlation with reciprocal illumination deviation $1/\Delta L_{a}$ .

Method		HDR-NeRF-Real						HDR-Plenoxels-Real
		LDR-OE ( $t_{1},t_{3},t_{5}$ )			LDR-NE ( $t_{2},t_{4}$ )			LDR-OE ( $t_{1},t_{3},t_{5}$ )			LDR-NE ( $t_{2},t_{4}$ )
		PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
exp3	HDR-NeRF [ 12 ]	34.27	0.9532	0.063	32.15	0.9475	0.074	-	-	-	-	-	-
	HDR-GS [ 3 ]	34.87	0.9697	0.021	31.02	0.9636	0.029	31.17	0.9509	0.040	28.60	0.9285	0.051
	GaussHDR [ 25 ]	36.05	0.9739	0.015	33.49	0.9707	0.017	31.50	0.9530	0.037	28.92	0.9326	0.044
	GaussHDR $\dagger$ [ 25 ]	36.32	0.9770	0.011	33.84	0.9738	0.014	32.87	0.9590	0.028	29.78	0.9391	0.037

Table 1: Quantitative results on realistic HDR-NeRF-Real [ 12 ] and HDR-Plenoxels-Real [ 15 ] datasets, where the best and second-best results are highlighted in red and yellow , respectively.

\dagger

indicates variants built on Scaffold-GS [ 27 ] . The proposed method achieves the overall best performance, demonstrating its effectiveness in synthesizing high-quality novel views across different exposure levels.

Method		LDR-OE ( $t_{1},t_{3},t_{5}$ )			LDR-NE ( $t_{2},t_{4}$ )			HDR
Method		PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
exp3	HDR-NeRF [ 12 ]	38.82	0.9657	0.032	38.07	0.9641	0.034	26.63	0.9523	0.046
	HDR-GS [ 3 ]	40.28	0.9781	0.018	27.07	0.8744	0.127	17.51	0.6982	0.205
	GaussHDR [ 25 ]	42.28	0.9853	0.007	41.65	0.9850	0.007	37.78	0.9704	0.017
	GaussHDR $\dagger$ [ 25 ]	43.87	0.9899	0.004	42.74	0.9894	0.004	39.08	0.9767	0.011
	Ours	43.11	0.9848	0.008	42.48	0.9852	0.008	38.77	0.9710	0.016

Table 2: Quantitative results on the synthetic HDR-NeRF-Syn [ 12 ] dataset, where the best and second-best results are highlighted in red and yellow , respectively.

\dagger

indicates variants built on Scaffold-GS [ 27 ] . Our method consistently outperforms baselines on both LDR and HDR scenarios, demonstrating its effectiveness in reconstructing HDR details and preserving information during tonemapping.

Figure 5: Qualitative comparisons on LDR views. For each method, we show the reconstructed LDR image and the residual map w.r.t. the ground truth. Competing methods exhibit noticeable missing content in saturated regions (e.g., screen reflections in the 1st row), indicating information loss after tone mapping, whereas our method effectively preserves fine structures and details.

Figure 6: Qualitative comparisons on HDR views, where we include residual maps between the results and GT to highlight the difference. HDR-NeRF and HDR-GS struggle to reproduce correct illumination levels due to the absence of HDR supervision during training, leading to inaccurate brightness and lost details. By imposing cross-branch HDR consistency, our method faithfully estimates scene lighting and reconstructs fine structures ( e.g . , basket edges in the 1st row).

Method	Rendering (ms)	Throughput (FPS)	Training (min)	Memory (MB)
HDR-NeRF	4189	0.24	500	11049
HDR-GS	9	117	10	5014
GaussHDR	19	53	28	5596
GaussHDR $\dagger$	26	38	21	6221
Ours	13	76	15	3274
Ours $\dagger$	19	53	18	3920

Table 3: Efficiency analysis measured on a single NVIDIA A6000 GPU. Rendering time and throughput are evaluated at an output resolution of

400\times 400

Method	HDR-NeRF-Real						HDR-Plenoxels-Real
	LDR-OE ( $t_{1},t_{3},t_{5}$ )			LDR-NE ( $t_{2},t_{4}$ )			LDR-OE ( $t_{1},t_{3},t_{5}$ )			LDR-NE ( $t_{2},t_{4}$ )
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
IE branch	36.18	0.9766	0.0105	33.38	0.9727	0.0142	32.73	0.9576	0.0274	29.47	0.9368	0.0358
+ GI branch	36.27	0.9770	0.0096	33.46	0.9728	0.0132	32.83	0.9570	0.0271	29.34	0.9350	0.0362
+ HDR-cons	36.43	0.9774	0.0094	33.84	0.9732	0.0129	32.97	0.9584	0.0262	29.45	0.9364	0.0354
+ I-GS	36.91	0.9777	0.0093	34.15	0.9737	0.0124	33.06	0.9592	0.0253	2 9.76	0.9389	0.0339

Table 4: Ablation studies on HDR-NeRF-Real and HDR-Plenoxels-Real dataset, with the best and second-best results highlighted in red and yellow . IE branch indicates the baseline containing only the image-exposure branch. GI branch, HDR-cons and I-GS denote Gaussian-illumination branch, self-consistent HDR learning and illumination-guided gradient scaling, respectively.

Figure 7: Visualization of ablation studies, where residual maps between LDR results and LDR GT are provided. Including GI branch effectively captures lighting-dependent appearance ( e.g . , table reflections in 1st row) and reduces color distortions. Introducing HDR-cons and I-GS further refines structural details.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Method			Box			Computer			Flower			Luckycat
Method			PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
exp3	LDR-OE	HDR-NeRF [ 12 ]	35.13	0.9612	0.055	34.28	0.9486	0.076	33.21	0.9525	0.058	34.47	0.9504	0.063
		HDR-GS [ 3 ]	36.21	0.9797	0.011	36.02	0.9720	0.017	32.44	0.9603	0.039	34.82	0.9668	0.018
		GaussHDR [ 25 ]	37.17	0.9821	0.009	36.24	0.9735	0.015	34.61	0.9681	0.020	36.19	0.9718	0.014
		GaussHDR $\dagger$ [ 25 ]	37.18	0.9834	0.008	36.37	0.9737	0.013	35.06	0.9757	0.012	36.67	0.9750	0.012
		Ours	37.94	0.9845	0.007	35.67	0.9761	0.012	33.91	0.9732	0.015	37.23	0.9779	0.010

Appendix

Table 5: Per-scene quantitative comparisons of each scene on HDR-NeRF-Real [ 12 ] dataset. For each scene, the best and second-best results are highlighted in red and yellow . LDR-OE and LDR-NE denote the LDR results with exposure

\{t_{1},t_{3},t_{5}\}

and

\{t_{2},t_{4}\}

, respectively.

Method			Character			Coffee			Desk			Plant
Method			PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
exp3	LDR-OE	HDR-GS [ 3 ]	36.08	0.9799	0.024	28.13	0.9468	0.046	30.07	0.9366	0.039	30.39	0.9403	0.049
		GaussHDR [ 25 ]	36.47	0.9813	0.024	29.10	0.9512	0.041	29.79	0.9355	0.038	30.63	0.9439	0.043
		GaussHDR $\dagger$ [ 25 ]	39.09	0.9847	0.017	29.82	0.9597	0.029	30.52	0.9417	0.031	32.05	0.9499	0.036
		Ours	39.45	0.9854	0.015	28.94	0.9538	0.035	29.79	0.9351	0.037	30.56	0.9402	0.046
		Ours $\dagger$	40.34	0.9870	0.013	29.51	0.9585	0.026	30.66	0.9428	0.028	31.72	0.9486	0.034

Appendix

Table 6: Per-scene quantitative comparisons on the HDR-Plenoxels-Real [ 15 ] dataset. HDR-NeRF [ 12 ] is not included, as it is evaluated on HDR-Plenoxels-Real. For each scene, the best and second-best results are highlighted in red and yellow . LDR-OE and LDR-NE denote LDR results with exposure sets

\{t_{1},t_{3},t_{5}\}

and

\{t_{2},t_{4}\}

, respectively.

Method		Bathroom			Bear			Chair			Desk
Method		PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
LDR-OE	HDR-NeRF [ 12 ]	33.98	0.9067	0.108	43.96	0.9917	0.007	34.36	0.9286	0.064	38.78	0.9735	0.021
	HDR-GS [ 3 ]	42.29	0.9796	0.008	38.61	0.9835	0.015	37.75	0.9734	0.013	43.62	0.9885	0.004
	GaussHDR [ 25 ]	42.38	0.9803	0.006	45.60	0.9934	0.002	37.44	0.9671	0.015	42.84	0.9888	0.004
	GaussHDR $\dagger$ [ 25 ]	42.97	0.9844	0.006	46.71	0.9945	0.001	38.48	0.9746	0.011	43.77	0.9905	0.003
	Ours	42.34	0.9790	0.006	46.76	0.9942	0.002	37.89	0.9708	0.012	43.78	0.9881	0.003

Appendix

Table 7: Per-scene quantitative comparisons on the HDR-NeRF-Syn [ 12 ] dataset under the exp3 exposure setting. For each scene, the best and second-best results are highlighted in red and yellow . LDR-OE and LDR-NE denote LDR results with exposure sets

\{t_{1},t_{3},t_{5}\}

and

\{t_{2},t_{4}\}

, respectively.

Method		Bathroom			Bear			Chair			Desk
Method		PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
LDR-OE	HDR-NeRF [ 12 ]	33.97	0.9066	0.108	43.81	0.9915	0.007	34.31	0.9284	0.064	38.26	0.9701	0.023
	HDR-GS [ 3 ]	40.62	0.9734	0.011	36.31	0.9745	0.020	35.21	0.9610	0.022	41.34	0.9847	0.006
	GaussHDR [ 25 ]	41.02	0.9731	0.009	44.02	0.9912	0.003	37.01	0.9678	0.014	42.46	0.9872	0.004
	GaussHDR $\dagger$ [ 25 ]	42.08	0.9807	0.006	45.40	0.9934	0.002	37.65	0.9712	0.012	43.03	0.9896	0.003
	Ours	40.75	0.9740	0.009	45.42	0.9933	0.002	37.36	0.9695	0.013	42.31	0.9884	0.004

Appendix

Table 8: Per-scene quantitative comparisons on the HDR-NeRF-Syn [ 12 ] dataset under the exp1 exposure setting. For each scene, the best and second-best results are highlighted in red and yellow . LDR-OE and LDR-NE denote LDR results with exposure sets

\{t_{1},t_{3},t_{5}\}

and

\{t_{2},t_{4}\}

, respectively.

Figure 8: Qualitative results on LDR views, where residual maps between the rendered results and ground truth are visualized to highlight the difference. HDR-GS struggles to balance between dynamic-range content reconstruction and fine-structure preservation( e.g . , characters in the highlight region of the 2nd row are missing), whereas our approach delivers accurate dynamic-range estimations and sharper details.

Figure 9: Qualitative results on HDR views, where residual maps between the rendered results and ground truth are visualized to highlight the differences. HDR-NeRF and HDR-GS fail to reach the target illumination level due to the absence of HDR ground truth during training, whereas our results yield accurate dynamic-range estimations and sharper reconstructions ( e.g . , leaf textures in the 4th row).

Novel View Synthesis

Novel View Synthesis creates images of a scene from camera viewpoints not present in the input, using learned geometry and appearance models.

See all 282 papers

InstantHDR: Single-forward Gaussian Splatting Initialization for HDR 3D Reconstruction
Mar 11, 2026Dingqiang Ye, Jiacong Xu, Jianglu Ping +3HDR Imaging Neural Rendering
Johns Hopkins University, USA · Shenzhen University, China
HDR-NSFF: High Dynamic Range Neural Scene Flow Fields
Mar 9, 2026Shin Dong-Yeon, Kim Jun-Seong, Kwon Byung-Ki +1HDR Video Reconstruction HDR Imaging
School of Electrical Engineering, KAIST · Dept. of Electrical Engineering, POSTECH · Grad. School of Artificial Intelligence, POSTECH +1
P2GS: Physical Prior-guided Gaussian Splatting for Photometrically Consistent Urban Reconstruction
May 16, 2026Kota Shimomura, Hidehisa Arai, Tsubasa Takahashi +2Multi-View Consistency 3D Gaussian Splatting
Chubu University1 · Turing Inc.2

Mar 11, 2026J/K move · Enter open · S save

InstantHDR: Single-forward Gaussian Splatting Initialization for HDR 3D Reconstruction

Dingqiang Ye, Jiacong Xu, Jianglu Ping, Yuxiang Guo+2

Johns Hopkins University, USA · Shenzhen University, China

HDR Imaging Neural Rendering Novel View Synthesis Feed-Forward 3D Reconstruction Gaussian Splatting

Open paper AskPDF arXiv

High dynamic range (HDR) novel view synthesis (NVS) aims to reconstruct HDR scenes from multi-exposure low dynamic range (LDR) images. Existing HDR pipelines heavily rely on known camera poses, well-initialized dense point clouds, and time-consuming per-scene optimization. Current feed-forward alternatives overlook the HDR problem by assuming exposure-invariant appearance. To bridge this gap, we propose InstantHDR, a feed-forward network that initializes 3D HDR scenes from uncalibrated multi-exposure LDR collections in a fast single forward pass. Specifically, we design a geometry-guided appearance modeling for multi-exposure fusion, and a meta-network for generalizable scene-specific tone mapping. Due to the lack of HDR scene data, we build a pre-training dataset, called HDR-Pretrain, for generalizable feed-forward HDR models, featuring 168 Blender-rendered scenes, diverse lighting types, and multiple camera response functions. Comprehensive experiments show that our InstantHDR delivers a single-forward HDR initialization at approximately 700 times the speed of SoTA optimization-based methods, and reaches comparable quality in real settings after lightweight post-optimization while remaining approximately 20 times faster. All code, models, and datasets: https://github.com/Bugjudger/InstantHDR.

Abstract

Novel View Synthesis

Explore similar work

InstantHDR: Single-forward Gaussian Splatting Initialization for HDR 3D Reconstruction

HDR-NSFF: High Dynamic Range Neural Scene Flow Fields

P2GS: Physical Prior-guided Gaussian Splatting for Photometrically Consistent Urban Reconstruction