Prior-Guided Single-Exposure HDR Novel View Synthesis via Base-Gain Representation

Armin Shafiee Sarvestani; Sheyang Tang; Jialu Xu; Zhou Wang

cs.CVOct 29, 2026

Prior-Guided Single-Exposure HDR Novel View Synthesis via Base-Gain Representation

Authors: Armin Shafiee Sarvestani, Sheyang Tang, Jialu Xu, Zhou Wang

Organizations: Department of Electrical and Computer Engineering, University of Waterloo

Abstract

Single-exposure high dynamic range (HDR) novel view synthesis (NVS) recovers HDR scenes from low dynamic range (LDR) views captured at a single exposure and estimates camera response function (CRF) for LDR re-exposure. Without HDR ground-truth, using single-exposure LDR views as the sole source of direct supervision underconstrains HDR reconstruction and cannot ensure its faithful recovery, motivating the need for learned LDR-to-HDR reconstruction priors. Moreover, single-exposure LDR views supervise CRF only over a limited range, necessitating a learnable prior for reliable re-exposure beyond it. We propose PriorGS, a 3D Gaussian Splatting framework that addresses both challenges by (1) guiding HDR recovery with pseudo-HDR supervision from 2D single-exposure HDR reconstruction methods while anchoring it to input LDR views through a closed reconstruction loop; and (2) regularizing CRF modeling with a learnable prior that guarantees monotonicity and promotes plausible responses. Yet cross-view inconsistencies in independently generated pseudo-HDR targets impair scene reconstruction when HDR is encoded directly as Gaussian's color attribute. We therefore introduce a 3D base-gain HDR representation that factors HDR radiance into an LDR base and an overlaid 3D gain field, promoting robustness to such inconsistencies and enabling 3D HDR appearance editing. Extensive experiments demonstrate PriorGS's substantial superiority over baselines.

Figures & tables

Figure 1: Given the same single-exposure inputs, methods lacking learned LDR-to-HDR priors reconstruct the LDR novel view faithfully but fail at HDR radiance recovery. We leverage this prior to recover radiance much closer to the GT. HDR results are tone-mapped identically for visualization.

Figure 2: ( top row ) Cross-view inconsistencies in pseudo-HDR targets ( bottom row ) produce severe artifacts w/ direct HDR encoding, while our base–gain representation promotes robustness to such inconsistencies.

Figure 3: Overview of PriorGS . Given single-exposure multi-view LDR captures of a scene, we reconstruct the 3D HDR scene as an LDR base extended by a 3D gain field. Input views supervise the LDR base, while pseudo-HDR targets generated by a 2D single-exposure HDR reconstruction model impose LDR-to-HDR prior on HDR recovery. We then re-expose the HDR view by a prior-constrained learnable CRF to match the same input view, which closes the reconstruction loop.

Method	HDR-NeRF-Real – LDR			HDR-Plenoxels-Real – LDR
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
HDR-GS	11.10	0.5212	0.370	4.70	0.2799	0.456
GaussHDR	16.28	0.7849	0.176	17.73	0.7141	0.197
Mono-HDR-GS	11.43	0.6627	0.301	9.79	0.5570	0.389
SeHDR	13.75	0.6991	0.182	14.21	0.6644	0.232
PriorGS (Ours)	24.31	0.9248	0.064	23.00	0.9021	0.092

Table 1: Quantitative results of ( top half ) re-exposed LDR-NVS on real scenes of HDR-NeRF and HDR-Plenoxels datasets and ( bottom half ) HDR-NVS and re-exposed LDR-NVS on synthetic scenes of HDR-NeRF dataset. red and yellow denote the best and second-best results, respectively.

Figure 4: Qualitative HDR-NVS results given low- ( $\Delta t_{1}$ ), medium- ( $\Delta t_{3}$ ), and high-exposure ( $\Delta t_{5}$ ) inputs across ( top half ) synthetic and ( bottom half ) real scenes. PriorGS substantially outperforms all baselines in high-fidelity HDR reconstruction across all scenes and exposure settings.

Figure 5: Qualitative re-exposed LDR-NVS results across scenes for three input $\rightarrow$ output exposure setups: ( top row ) $\Delta t_{1}\rightarrow\Delta t_{5}$ , ( middle row ) $\Delta t_{3}\rightarrow\Delta t_{1}$ , ( bottom row ) $\Delta t_{5}\rightarrow\Delta t_{3}$ . PriorGS consistently outperforms all baselines in novel-view LDR re-exposure at unseen exposures.

Figure 6: View-consistent 3D HDR appearance editing via gain field scaling with $\rho$ . We sweep $\rho$ and show an HDR view at a low-exposure ( top row ) and after a fixed tone-mapping ( bottom row ). Increasing $\rho$ expands dynamic range (DR) and enhances HDR appearance and chromatic contrast.

(ID) Description	HDR-NeRF-Real – LDR			HDR-Plenoxels-Real – LDR
(ID) Description	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
(a) Remove LDR-to-HDR prior	18.39	0.8503	0.099	15.00	0.7881	0.141
(b) Remove reconstruction loop	23.52	0.9180	0.073	22.97	0.9010	0.099
(c) Base-Gain $\rightarrow$ direct encoding	22.71	0.8981	0.105	15.05	0.6860	0.251
(d) Gain Field $\rightarrow$ $\operatorname{MLP}({\bm{c}},{\bm{\mu}},{\bm{d}})$	24.44	0.9238	0.066	22.64	0.8979	0.093
(e) Gain Field $\rightarrow$ $\operatorname{MLP}({\bm{c}})$	24.28	0.9268	0.066	22.07	0.8732	0.111

Table 2: Ablations on re-exposed LDR-NVS on real scenes of HDR-NeRF and HDR-Plenoxels, with best and second-best results denoted by red and yellow .

x\rightarrow y

denotes replacing

x

with

y

Figure 7: Ablations on HDR views for ( top left ) removing LDR-to-HDR prior; ( bottom left ) removing reconstruction loop; and replacing our base-gain HDR encoding by ( top right ) two MLP variants or ( bottom right ) direct HDR encoding. All ablations degrade HDR reconstruction quality.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 8 — Figure 8: Representational capacity of our proposed CRF model on the DoRF database. (a) the 201 DoRF response curves. (b) the independently fitted responses of our CRF. All curves are displayed using a logarithmic irradiance axis; the zero-irradiance samples are included during fitting but omitted from the visualization.

Method	HDR-NeRF-Real – LDR			HDR-Plenoxels-Real – LDR
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
GaussHDR	16.28	0.7849	0.176	17.73	0.7141	0.197
SeHDR	13.75	0.6991	0.182	14.21	0.6644	0.232
PriorGS + DeepHDR	20.38	0.8604	0.085	18.43	0.8111	0.115
PriorGS + ExpandNet	21.18	0.8771	0.094	21.23	0.8624	0.100
PriorGS + GMNet	19.36	0.8482	0.095	16.94	0.7887	0.126

Appendix

Table 3: PriorGS’s performance after integration with different 2D HDR reconstruction methods on ( top half ) re-exposed LDR-NVS on real scenes of HDR-NeRF and HDR-Plenoxels datasets and ( bottom half ) HDR-NVS and re-exposed LDR-NVS on synthetic scenes of HDR-NeRF dataset. red and yellow denote the best and second-best results, respectively.

Method	HDR-NeRF-Real – LDR			HDR-Plenoxels-Real – LDR
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
HDR-GS	35.24	0.9667	0.0198	30.43	0.9428	0.0492
GaussHDR	34.42	0.9662	0.0366	30.54	0.9447	0.0500
Mono-HDR-GS	35.65	0.9656	0.0172	31.02	0.9465	0.0464
SeHDR	30.08	0.9632	0.0346	21.77	0.9037	0.1157
PriorGS (ours)	36.10	0.9700	0.0158	31.71	0.9482	0.0439

Appendix

Table 4: Quantitative results on LDR-NVS at the training exposure on real scenes of HDR-NeRF and HDR-Plenoxels datasets. red and yellow denote the best and second-best results, respectively.

Method	HDR-NeRF-Syn – LDR
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
HDR-GS	41.56	0.9825	0.0088
GaussHDR	40.90	0.9795	0.0154
Mono-HDR-GS	41.80	0.9828	0.0083
SeHDR	32.53	0.9768	0.0199
PriorGS (ours)	41.77	0.9836	0.0087

Appendix

Table 5: Quantitative results on LDR-NVS at the training exposure on synthetic scenes of HDR-NeRF dataset. red and yellow denote the best and second-best results, respectively.

Method	HDR-NeRF-Real – LDR			HDR-Plenoxels-Real – LDR
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
GaussHDR $\dagger$	22.76	0.8739	0.128	21.03	0.8635	0.115
GaussHDR	16.28	0.7849	0.176	17.73	0.7141	0.197
PriorGS (Ours)	24.31	0.9248	0.064	23.00	0.9021	0.092
Method	HDR-NeRF-Syn – HDR			HDR-NeRF-Syn – LDR
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$

Appendix

Table 6: Quantitative comparison of PriorGS vs. the 3DGS ( Kerbl et al., 2023 ) and Scaffold-GS

\dagger

( Lu et al., 2024 ) variants of GaussHDR ( Liu et al., 2025 ) for ( top half ) re-exposed LDR-NVS on real scenes of HDR-NeRF and HDR-Plenoxels datasets and ( bottom half ) HDR-NVS and re-exposed LDR-NVS on synthetic scenes of HDR-NeRF dataset. red and yellow denote the best and second-best results, respectively.

Figure 9: Further qualitative HDR-NVS results given low- ( $\Delta t_{1}$ ), medium- ( $\Delta t_{3}$ ), and high-exposure ( $\Delta t_{5}$ ) inputs across ( top half ) synthetic and ( bottom half ) real scenes.

Figure 10: Further qualitative re-exposed LDR-NVS results. Using single-exposure LDR views for training, we render a novel view at all five exposures, including the training exposure. The training exposure is ( top half ) $\Delta t_{3}$ for a synthetic scene and ( bottom half ) $\Delta t_{2}$ for a real scene.

Figure 11: Learned MLP-based and Prior-Constrained CRF curves and novel-view LDR re-exposures. Both CRFs faithfully reproduce the LDR view at training exposure, where target intensities fall within their CRFs’ supervised range. Beyond this range, MLP-based CRF curve exhibits RGB divergence, early saturation, and nonmonotonic behavior, resulting in poor rendering at test exposure that involves mapping to CRF’s unsupervised range. Our prior-constrained CRF maintains a plausible, monotonic response in this range, producing more faithful test-exposure renderings.

Panel of Figure 12 — Figure 12: Changes in an HDR view’s log-luminance histogram by varying $\rho$ from $0.5$ to $1.5$ when $\rho$ scales (a) 3D HDR radiance and (b) the 3D gain field. Increasing $\rho$ for 3D HDR radiance scaling only shifts the histogram (equivalent to an exposure shift) and leaves the dynamic range (DR) unchanged. For 3D gain-field scaling, it instead changes the histogram and substantially expands the DR.

Figure 13: Further example of 3D HDR appearance editing via gain field scaling with $\rho$ . We sweep $\rho$ and show an HDR view at a low-exposure ( top row ) and after a fixed tone-mapping ( bottom row ). Increasing $\rho$ expands dynamic range (DR) and enhances HDR appearance and chromatic contrast. Since $\rho$ operates directly on 3D Gaussians, the edits are inherently view-consistent.

Figure 14: Localized 3D HDR appearance editing by adjusting $\rho$ for a manually selected subset of 3D Gaussians across three scenes: ( top ) brightening the region beneath a shelf, ( middle ) dimming an overly bright book, and ( bottom ) enhancing a color checker’s vividness. We show before-and-after 2D visualizations of the 3D gain field and HDR renderings from two viewpoints, displayed at a selected exposure. As $\rho$ operates directly in 3D, the edits are inherently view-consistent.

Figure 15: Further qualitative ablations on HDR views for ( left ) removing LDR-to-HDR reconstruction priors and ( right ) disabling the reconstruction loop. Removing the LDR-to-HDR reconstruction prior results in inaccurate HDR radiance recovery, while disabling the reconstruction loop degrades HDR reconstruction quality. All HDR views are tone-mapped identically for visualization.

Figure 16: Further qualitative comparison (ablation) of our base-gain HDR representation against direct HDR encoding. ( top row ) Cross-view inconsistencies in pseudo-HDR targets ( bottom row ) produce severe artifacts with direct HDR encoding, while our base–gain representation promotes robustness to such inconsistencies and achieves higher quality HDR reconstruction.

Figure 17: Further qualitative ablations on HDR views for replacing our gain field with two MLP variants: one predicts each Gaussian’s HDR radiance from its LDR-base color ${\bm{c}}$ alone; the other also uses position ${\bm{\mu}}$ and direction ${\bm{d}}$ . Our base-gain representation outperforms both MLP variants in reconstructing high-frequency details in regions with challenging lighting.

Figure 18: Qualitative ablations on LDR views for replacing our proposed CRF with MLPs in 2D and 3D and with fixed gamma CRF. All configurations reproduce the training exposure, yet our CRF achieves much higher reconstruction fidelity at the test exposure.

Figure 19: Qualitative ablation of LDR-base supervision on HDR views. Removing this supervision severely impairs scene reconstruction, causing failure in texture recovery all across the scene.

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 304 papers

Physically Inspired Gaussian Splatting for HDR Novel View Synthesis
Mar 30, 2026Huimin Zeng, Yue Bai, Hailing Wang +1Novel View Synthesis HDR Imaging
Department of Electrical and Computer Engineering, Northeastern University · Khoury College of Computer Science, Northeastern University
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
ExpoCM: Exposure-Aware One-Step Generative Single-Image HDR Reconstruction
May 4, 2026Aoyu Liu, Zhen Liu, Ziyi Wang +3Image Enhancement HDR Imaging
University of Electronic Science and Technology of China

Mar 30, 2026J/K move · Enter open · S save

Physically Inspired Gaussian Splatting for HDR Novel View Synthesis

Huimin Zeng, Yue Bai, Hailing Wang, Yun Fu

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

Novel View Synthesis HDR Imaging Neural Rendering 3D Gaussian Splatting

Open paper AskPDF arXiv

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/.

Abstract

Novel View Synthesis

Explore similar work

Physically Inspired Gaussian Splatting for HDR Novel View Synthesis

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

ExpoCM: Exposure-Aware One-Step Generative Single-Image HDR Reconstruction