HDR-GS: Efficient High Dynamic Range Novel
View Synthesis at 1000× Speed via Gaussian Splatting

Advances in Neural Information Processing Systems (NeurIPS), 2024

Yuanhao Cai1, Zihao Xiao1, Yixun Liang2, Minghan Qin3, Yulun Zhang4,†, Xiaokang Yang4, Yaoyao Liu5, Alan Yuille1
1 Johns Hopkins University   2 HKUST   3 Tsinghua University
4 Shanghai Jiao Tong University   5 University of Illinois Urbana-Champaign

† Corresponding author

LDR videos sweep brightness while HDR videos visualize radiance with fixed Reinhard tone mapping. Each pair shares the same continuous spiral-forward-and-backward camera path. Explore the corresponding LDR and HDR Gaussian point clouds below each video.

Abstract

HDR-GS reconstructs high dynamic range scenes from photographs captured at different exposures. A shared Gaussian representation models scene radiance and learns how a camera maps that radiance to LDR colors. Parallel rasterization produces HDR novel views and exposure-controlled LDR images, enabling efficient training and real-time rendering.

1000×inference speed vs. HDR-NeRF
6.3%of HDR-NeRF’s training time
+3.84 / +1.91 dBLDR / HDR novel-view PSNR

Reported in the paper relative to HDR-NeRF; measurements depend on the experimental setting.

Method Overview

HDR-GS shares Gaussian geometry across HDR and LDR rendering. A learned tone mapper introduces exposure control before parallel rasterization.

HDR-GS pipeline: dual dynamic range Gaussians, learned tone mapping, and parallel differentiable rasterization for HDR and LDR outputs.
The HDR-GS framework: one scene representation, two dynamic ranges.

HDR radiance

Spherical harmonics describe directional log radiance; exponentiation yields HDR color.

si(d)= ∑ℓ=0L ∑m=−ℓℓ ki,ℓm Yℓm(d) cih(d)= esi(d)

Exposure control

Three channel-wise MLPs map log radiance and exposure to LDR.

cil(d,Δt)= gθ(si(d)+log(Δt)+b)

Parallel rasterization

Depth-ordered alpha blending renders both ranges using shared geometry and opacity.

Ir(p)= ∑j∈N(p) cjrσj ∏k<j(1−σk) ,r∈{h,l}

Here, d is view direction, Δt is exposure time, and σ is projected opacity. Adapted from Eqs. (7), (8), and (13) in the paper.

Comparisons

Qualitative and quantitative results from the paper

Synthetic Scenes

Synthetic-scene visual comparisons between HDR-GS and baseline methods.
Novel-view synthesis on synthetic scenes.
View quantitative results · Synthetic scenesSynthetic-scene PSNR, SSIM, LPIPS, training time, and inference speed comparison from the official HDR-GS paper.

Real Scenes

Real-scene visual comparisons between HDR-GS and baseline methods.
Novel-view synthesis on real captured scenes.
View quantitative results · Real scenesReal-scene quantitative comparison from the official HDR-GS paper.

BibTeX

@inproceedings{hdr_gs,
  title={HDR-GS: Efficient High Dynamic Range Novel View Synthesis
         at 1000x Speed via Gaussian Splatting},
  author={Yuanhao Cai and Zihao Xiao and Yixun Liang and Minghan Qin
          and Yulun Zhang and Xiaokang Yang and Yaoyao Liu and Alan Yuille},
  booktitle={NeurIPS},
  year={2024}
}