acceptodds
Under review as a conference paper at ICLR 2027

High Dynamic Range 3D Gaussian Splatting via Luminance–Chromaticity Decomposition

Abstract

High Dynamic Range (HDR) 3D reconstruction is pivotal for professional content creation in filmmaking and virtual production. Existing methods typically rely on multi-exposure Low Dynamic Range (LDR) supervision to constrain the learning process within vast brightness spaces, resulting in complex, dual-branch architectures. This work explores the feasibility of learning HDR 3D models exclusively in the HDR data space to simplify model design. By analyzing 3D Gaussian Splatting (3DGS) for HDR imagery, we reveal that its failure stems from the limited capacity of Spherical Harmonics (SHs) to capture extreme radiance variations across views, often biasing towards high-radiance observations and underfitting. While increasing the maximum SH degree improves training fitting, it leads to severe overfitting and excessive parameter overhead. To address this, we propose Luminance–Chromaticity Decomposition Gaussian Splatting (LCD-GS). By decoupling luminance and chromaticity into independent parameters, LCD-GS significantly enhances learning flexibility with minimal parameter increase (e.g., one extra scalar per primitive). Notably, LCD-GS maintains the original training and inference pipeline, requiring only a change in color representation. This explicit decomposition naturally enables primitive-level luminance editing during inference. Extensive experiments on synthetic and real datasets demonstrate that LCD-GS consistently outperforms state-of-the-art methods in reconstruction fidelity and dynamic-range preservation even with a simpler, more efficient architecture, providing an elegant paradigm for professional-grade HDR 3D modeling. Code and datasets will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
high dynamic range, novel view synthesis, 3D Gaussian Splatting, neural rendering, 3D reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_XFwyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LCD-GS, a 3DGS variant for HDR novel view synthesis trained directly on native HDR images. It argues that vanilla 3DGS's spherical harmonics entangle luminance with chromaticity, and proposes a luminance-chromaticity decomposition. Experiments on synthetic HDR-Syn-8S and a new real-world iHDR-4S dataset show gains over 3DGS and HDR-NVS baselines, plus luminance editing.

Critical strengths

The paper identifies that vanilla 3DGS's spherical harmonics entangle luminance with chromaticity, causing underfitting under HDR supervision. The proposed decomposition adds only one scalar per primitive and integrates without architectural change, with consistent gains over 3DGS. This matters because native HDR 3D reconstruction is increasingly relevant.

Critical weaknesses
  1. The paper operates in a linear HDR color space but does not compare against methods [1,2,3] addressing HDR in linear RAW color space. While the paper's input differs from Bayer RAW in data form and capture pipeline, both carry linear HDR radiance. These RAW-based solutions may also be effective for the paper's input setting. Specifically, LE3D [1] explicitly states that Spherical Harmonics (SH) is unsuitable for RAW linear color space and replaces it with a tiny Color MLP. The absence of any comparison to these methods, and of any discussion of whether their solutions transfer to linear HDR color space, leaves it unclear whether the proposed luminance-chromaticity decomposition is necessary or merely one of several viable fixes to the same SH problem.

  2. HDR-GS [4] is trained with HDR supervision on synthetic data in its original design, but it is unclear from the paper whether Table 2 evaluates it with HDR supervision or under the LDR-only protocol. This ambiguity should be clarified. For GaussHDR [5], it reconstructs HDR without native HDR ground truth. The paper evaluates GaussHDR without providing HDR supervision in Table 2. A fair comparison would train HDR-GS and GaussHDR with HDR supervision (in μ\mu-law). Additionally, if HDR supervision is available, Table 3 (iHDR-4S) could include both HDR-GS and GaussHDR for a more complete comparison, which the paper currently does not provide.

[1] Xin Jin, Pengyi Jiao, Zheng-Peng Duan, Xingchao Yang, Chongyi Li, Chun-Le Guo, and Bo Ren. Lighting Every Darkness with 3DGS: Fast Training and Real-Time Rendering for HDR View Synthesis. NeurIPS 2024.

[2] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. NeRF in the Dark: High Dynamic Range View Synthesis from Noisy Raw Images. CVPR 2022.

[3] Shreyas Singh, Aryan Garg, and Kaushik Mitra. HDRSplat: Gaussian Splatting for High Dynamic Range 3D Scene Reconstruction from Raw Images. BMVC 2024.

[4] Yuanhao Cai, Zihao Xiao, Yixun Liang, Minghan Qin, Yulun Zhang, Xiaokang Yang, Yaoyao Liu, and Alan L. Yuille. HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting. NeurIPS 2024.

[5] Jinfeng Liu, Lingtong Kong, Bo Li, and Dan Xu. GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping. CVPR 2025.

Questions for authors

How to conduct chromaticity-level editing such as white balance adjustment?

Reviewer_HDcFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Luminance–Chromaticity Decomposition Gaussian Splatting (LCD-GS) for reconstructing 3D scenes directly from high-dynamic-range (HDR) images. The authors argue that standard 3D Gaussian Splatting (3DGS) struggles with large changes in radiance because its spherical harmonics (SH) color representation handles brightness and view-dependent color together. LCD-GS gives each Gaussian a view-independent positive luminance value and uses SH to model view-dependent chromaticity. The method also uses μ-law compression during training and keeps most of the original 3DGS pipeline. Experiments on eight synthetic scenes and four real scenes show better reconstruction scores than the compared methods, along with fast rendering. The paper also demonstrates simple luminance editing.

Critical strengths
  1. Practical problem. Reconstructing 3D scenes directly from HDR images is useful because HDR images preserve bright and dark details that may be lost in LDR images.

  2. Simple method with a clear purpose. LCD-GS separates luminance from view-dependent chromaticity with only a small change to 3DGS. The explicit luminance value also makes brightness editing straightforward.

  3. Strong results and useful ablations. LCD-GS reaches 41.84 dB PSNR on HDR-Syn-8S, compared with 31.53 dB for the default 3DGS setting (Table 2). Table 5 also tests both methods with and without μ-law compression. LCD-GS improves PSNR in both settings, which supports the value of the decomposition itself. The paper further reports image quality and rendering speed.

Critical weaknesses
  1. Baseline convergence needs a stronger test. The paper shows that raising the SH degree gives only a small gain for standard 3DGS. This is useful, but it does not show whether the baselines have reached their best results. Longer training or better-tuned learning rates may change the gap. Learning curves and results at matched or larger training budgets would make the comparison more convincing.

  2. The real-world evaluation covers a narrow set of scenes. The paper uses eight synthetic scenes and four real scenes. The real examples mostly show small indoor arrangements. This makes it hard to judge performance on outdoor scenes, strong light sources, reflections, and other HDR conditions. More views of the same scenes cannot replace more varied scenes.

  3. The view-independent luminance choice needs further study. LCD-GS assigns one luminance value to each Gaussian, although brightness can change with view direction on shiny or reflective surfaces. Tests on these materials would help show where the representation works well and where it may fail.

Questions for authors
  1. Could the authors show PSNR or validation curves as training progresses for standard 3DGS and the higher-SH baselines? How does the comparison change when these methods train longer or use tuned learning-rate schedules?

  2. Table 5 provides a useful comparison with and without μ-law compression on HDR-Syn-8S. Could the authors also run this ablation on iHDR-4S and report results for each real scene?

  3. Could the authors test a wider range of real HDR scenes, such as outdoor scenes, strong backlighting, and reflective objects? Per-scene results would also help readers see whether the improvement is consistent.

  4. How does LCD-GS handle brightness that changes with viewing direction, such as specular highlights? Could the authors add an example or a failure-case analysis?

  5. Could the authors clarify whether the reported results are stable across training runs or different initializations?

Reviewer_SbsiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes LCD-GS, a 3DGS-based HDR 3D reconstruction method that is directly supervised with HDR images instead of multi-exposure LDR images. LCD-GS decomposes the SH-based color of each primitive into a view-independent luminance scalar and an SH-based chromaticity term. The method adds only one parameter per primitive and keeps the original 3DGS pipeline. The authors also introduce iHDR-4S, a real-world multi-view HDR dataset captured with an iPhone in ProRAW format.

Critical strengths
  1. The proposed LCD-GS is simple and lightweight. It only changes the color parameterization and can be easily integrated into other 3DGS variants.
  2. The proposed iHDR-4S dataset is a real-world multi-view HDR dataset , which would benefit the community.
Critical weaknesses
  1. In Eq.3, since the chromaticity term is normalized, the total radiance of each primitive equals to L and is view-independent. Would this assumption cause some issue when modeling specular highlights, whose radiance changes substantially as view direction changes? More more experiments or discussion on this issue would be helpful.

  2. Although the authors distinguishes RGB-domain HDR from RAW in related works, learning HDR 3D scenes directly from linear HDR observations has already been studied by RawNeRF, HDRSplat and LE3D. This weakens the claim of being the first study in this direction. Moreover, their setting aligns better with LCD-GS than the current baselines, which aim to recover HDR from LDR images. It would be better if the authors could compare with RawNeRF, HDRSplat and LE3D.

  3. The results of Mono-HDR-GS and Mono-HDR-NeRF in Table.3 is surprisingly low. Do these two method use GT HDR images as supervision during optimization? Authors should clarify this in the paper.

Questions for authors
  1. How does LCD-GS perform in regions with strong view-dependent effects, such as specular highlights or reflections of light sources on glossy surfaces?
  2. It would be helpful if authors can provide more discussion and comparison with RawNeRF, HDRSplat and LE3D.
  3. Were Mono-HDR-GS and Mono-HDR-NeRF trained with GT HDR supervision in Tables 2 and 3?

Comments on the paper

1 comment

Thanks for the responses. I still think a fair comparison against GaussHDR should be using HDR supervision. Since its color representation is HDR color + HDR-to-LDR tone-mapping, you can remove the tone-mapping, render HDR images from HDR Gaussians, and directly supervise the HDR Gaussians with HDR GT images. Similarly, you can train GaussHDR and HDR-GS on iHDR-4S dataset in above manner. Besides, I want to know why the performance differences between different activation functions are so large in Table R9 and Table 5, especially for HDR-Syn-8S dataset.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.