Reconstructing 3D Scenes in Native High Dynamic Range (2511.12895v1)

Abstract: High Dynamic Range (HDR) imaging is essential for professional digital media creation, e.g., filmmaking, virtual production, and photorealistic rendering. However, 3D scene reconstruction has primarily focused on Low Dynamic Range (LDR) data, limiting its applicability to professional workflows. Existing approaches that reconstruct HDR scenes from LDR observations rely on multi-exposure fusion or inverse tone-mapping, which increase capture complexity and depend on synthetic supervision. With the recent emergence of cameras that directly capture native HDR data in a single exposure, we present the first method for 3D scene reconstruction that directly models native HDR observations. We propose {\bf Native High dynamic range 3D Gaussian Splatting (NH-3DGS)}, which preserves the full dynamic range throughout the reconstruction pipeline. Our key technical contribution is a novel luminance-chromaticity decomposition of the color representation that enables direct optimization from native HDR camera data. We demonstrate on both synthetic and real multi-view HDR datasets that NH-3DGS significantly outperforms existing methods in reconstruction quality and dynamic range preservation, enabling professional-grade 3D reconstruction directly from native HDR captures. Code and datasets will be made available.

• The paper introduces NH-3DGS, a novel framework that decouples luminance and chromaticity to directly optimize native HDR radiance fields.
• It employs μ-law compression and sensor-domain training to ensure gradient stability and robust radiometric fidelity in high-contrast scenes.
• Quantitative evaluations on Syn-8S and RAW-4S benchmarks show superior PSNR, SSIM, and real-time performance compared to conventional methods.

Native HDR 3D Scene Reconstruction via Luminance–Chromaticity Decomposition

Background and Motivation

High Dynamic Range (HDR) imaging has become indispensable in professional content creation, supporting applications that mandate accurate radiometric fidelity across extensive luminance ranges, including virtual production, advanced rendering, and digital cinematography. Standard 3D scene reconstruction pipelines, however, are dominantly constrained to Low Dynamic Range (LDR) imagery, a limitation that severely restricts their utility in high-contrast, artifact-prone environments. Prior HDR view synthesis methods typically require multi-exposure fusion or RAW sensor data, which introduces additional complexity, capture overhead, and reliance on synthetic domain translations.

Recent innovations in native HDR sensor hardware now allow acquisition of single-exposure HDR images without the drawbacks of multi-exposure stacking. This paper, "Reconstructing 3D Scenes in Native High Dynamic Range" (2511.12895), directly addresses the challenge of reconstructing radiance fields from these native HDR signals. The authors present NH-3DGS, a novel framework that fundamentally redesigns color representation to support direct HDR optimization while mitigating the entanglement problems endemic to existing spherical harmonics (SH)-based approaches.

Analysis of LDR-Specific Limitations in 3DGS and NeRF

Standard 3D Gaussian Splatting (3DGS) encodes appearance at each Gaussian via low-order SH coefficients that simultaneously model both luminance and chromaticity. While adequate for LDR inputs, this joint representation fails under native HDR conditions where radiance can vary by several orders of magnitude across views (see Figure 1).

Figure 1: Examples of underexposure and overexposure in high-contrast HDR scenes. Standard 3DGS yields blurring artifacts in dark regions when given HDR supervision.

Direct application of vanilla 3DGS or LDR-centric NeRF architectures to HDR data leads to optimization bias—the network gradients become dominated by high-luminance views, resulting in perceptual artifacts, blurred shadows, and hue inconsistencies. Attempts to mitigate this by elevating SH order—e.g., increasing degree from 3 to 5—provide marginal improvements in quantitative metrics but introduce excessive parameter overhead and perceptual artifacts (Figure 2).

Figure 2: Elevation of SH order marginally mitigates blurring but induces additional artifacts and increases computation, confirming limitations in SH-based representation for HDR.

Native HDR Gaussian Splatting: Luminance–Chromaticity Decomposition

NH-3DGS introduces a physically principled and perceptually grounded decoupling of luminance and chromaticity within each Gaussian’s color representation. Rather than SH coefficients encoding all brightness and chromaticity, the method factorizes appearance such that luminance is a direct scalar parameter and chromaticity is encoded through view-dependent, low-order SH coefficients. This decomposition is motivated both by the radiometric structure of HDR data and human visual cortex processing.

Figure 3: NH-3DGS pipeline with HDR image sets and camera poses. Native HDR Gaussian representation is learned via luminance–chromaticity decomposition, enabling accurate synthesis for arbitrary radiance and viewpoint.

Mathematically, the rendered color for a Gaussian under direction $\mathbf{d}$ is given by

$$c(\mathbf{d}) = \text{Lum} \cdot f_\mathrm{SH}(\theta, \phi), \quad f_\mathrm{SH}(\theta, \phi) = \sum_{l=0}^L \sum_{m=-l}^{l} \mathbf{b}_l^m Y_l^m(\theta,\phi),$$

where $\text{Lum}$ is the scalar luminance parameter, and $f_{SH}$ accounts for chromaticity. This formulation enables the model to efficiently optimize HDR radiance fields with enhanced gradient stability and perceptual fidelity, especially in shadow and highlight regions.

The training pipeline applies a $\mu$-law compression to facilitate radiometric consistency, particularly emphasizing gradient signal in low-luminance regions—a critical aspect for scene reconstruction accuracy in HDR domains.

Evaluation on Synthetic and Real HDR Datasets

NH-3DGS is assessed on two benchmarks: Syn-8S (synthetic Blender-generated HDR scenes with paired multi-exposure LDRs) and RAW-4S (captured multi-view RAW HDR dataset with complex radiometric and geometric properties). The evaluations employ PSNR, SSIM, LPIPS, and frame rate metrics, with ablation on SH order and direct comparison to NeRF, 3DGS, HDR-NeRF, HDR-GS, Mono-HDR-GS, Mono-HDR-NeRF, RAW-NeRF, and LE3D.

Figure 4: HDR renderings on RAW-4S dataset show that 3DGS struggles in low-illumination zones, while NH-3DGS maintains spatial precision and color neutrality.

Quantitative results (Syn-8S):

• NeRF and 3DGS (HDR input): Substantially lower PSNR and perceptual quality, with pronounced artifacting in luminance extremes.
• 3DGS-L4 and L5: Marginal PSNR gains, but considerable slowdown and artifact introduction.
• NH-3DGS: 39.77 dB PSNR, 0.972 SSIM, 0.011 LPIPS at 233 fps—exceeding all alternatives, including multi-exposure methods, with negligible real-time performance compromise.

  Figure 5: Qualitative comparison on Syn-8S showing NH-3DGS’ faithful detail and dynamic range recovery in dark and bright regions compared to competing approaches.

On RAW-4S, NH-3DGS achieves the highest RAW-PSNR and RGB-PSNR, outperforming LE3D and RAW-NeRF, while using lower memory and enabling real-time inference. Importantly, the model is trained on raw Bayer data—prior to any demosaicing—enabling precise physical radiance recovery and implicit learning of color interpolation.

Figure 6: Comparison on RAW-4S dataset; NH-3DGS preserves fine content under extreme illumination, outperforming methods trained on demosaiced inputs.

Practical and Theoretical Implications

NH-3DGS constitutes a significant step forward in 3D radiance field modeling for HDR imagery, enabling high-fidelity scene reconstruction directly from raw sensor data. Key implications include:

• Professional-Grade HDR Reconstruction: Direct radiometric modeling from native HDR sources obviates complex exposure bracketing, multi-shot protocols, and synthetic HDR-LDR domain translation, aligning reconstruction with requirements of virtual production and photorealistic rendering workflows.
• Efficient, Robust HDR Optimization: The luminance–chromaticity factorization improves optimization stability, particularly in high-contrast scenes, and maintains computational efficiency comparable to standard 3DGS pipelines.
• Sensor-Domain Learning: Direct loss computation on Bayer-pattern raw images, as opposed to demosaiced RGB, enables learning that is intrinsically aligned with sensor physics, facilitating future developments in camera-aware reconstruction, low-light modeling, and physically consistent rendering.

Future directions likely include end-to-end HDR-aware pipelines leveraging native sensor architectures, refinement of decomposition strategies for alternative data modalities (multispectral, event-based HDR), and further integration with real-time graphics engines and AR/VR platforms.

Conclusion

The presented NH-3DGS framework delivers state-of-the-art 3D radiance field reconstruction performance for native HDR and RAW scenes. By systematically identifying and resolving the entanglement limitations of SH-based color modeling in high-dynamic-range domains, the method achieves superior radiometric fidelity, computational efficiency, and practical scalability. The results highlight the necessity and efficacy of directly modeling HDR data for professional 3D content synthesis and signal a clear trajectory for future HDR-centric research in scene representation and rendering.