Research map / 3D vision / Gaussian splatting
HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HDR-GS, a 3D Gaussian splatting framework that jointly renders HDR views and exposure-controllable LDR views via a dual dynamic range point cloud with an MLP tone-mapper and parallel rasterization, plus recalibrated camera poses and SfM initialization.
Problem
Existing HDR novel view synthesis methods are NeRF-based and suffer from very long training and slow inference; directly applying 3DGS to multi-exposure HDR data fails to converge and cannot render HDR or change exposure.
Approach
Model HDR color with spherical harmonics in log domain and map it to LDR color via per-channel MLP tone-mapper conditioned on exposure time (Debevec-style log CRF), then rasterize HDR and LDR colors in two parallel rasterization passes; also recalibrate NDC poses to SfM camera poses and initialize Gaussians from SfM points.
Key results
- On synthetic data, HDR-GS reaches 38.31 dB HDR PSNR vs 36.40 for HDR-NeRF (+1.91 dB), with 126 fps vs 0.122 fps and 34 min vs 542 min training.
- On LDR-OE synthetic track HDR-GS achieves 41.10 dB vs 39.07 dB for HDR-NeRF (+2.03 dB).
- On real scenes HDR-GS gets 35.47 dB on LDR-OE vs 31.63 dB for HDR-NeRF (+3.84 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HDR-GS instead of HDR-NeRF | Synthetic LDR-OE (t1,t3,t5) · Synthetic LDR-OE (t1,t3,t5) | PSNR: 39.07 -> 41.10 | Helped |
| Use HDR-GS instead of HDR-NeRF | Synthetic LDR-NE (t2,t4) · Synthetic LDR-NE (t2,t4) | PSNR: 37.53 -> 36.33 | Hurt |
| Use HDR-GS instead of HDR-NeRF | Synthetic HDR · Synthetic HDR | PSNR: 36.40 -> 38.31 | Helped |
| Use HDR-GS instead of HDR-NeRF | Real LDR-OE (t1,t3,t5) · Real LDR-OE (t1,t3,t5) | PSNR: 31.63 -> 35.47 | Helped |
| Use HDR-GS instead of HDR-NeRF | Real LDR-NE (t2,t4) · Real LDR-NE (t2,t4) | PSNR: 31.43 -> 31.66 | Helped |
| Add camera recalibration to 3DGS using original NDC coordinates (ablation) | Synthetic 3DGS baseline · Synthetic LDR-OE | PSNR: 12.35 -> 14.62 | Helped |
| Add camera recalibration to 3DGS using original NDC coordinates (ablation) | Synthetic 3DGS baseline · Synthetic LDR-NE | PSNR: 11.83 -> 14.41 | Helped |
| Initialize Gaussians with SfM points (ablation) | 3DGS with recalibrated cameras · Synthetic LDR-OE | PSNR: 14.62 -> 19.46 | Helped |
| Initialize Gaussians with SfM points (ablation) | 3DGS with recalibrated cameras · Synthetic LDR-NE | PSNR: 14.41 -> 18.97 | Helped |
| Add the DDR point-cloud model (ablation) | 3DGS with recalibrated cameras and SfM initialization · Synthetic LDR-OE | PSNR: 19.46 -> 41.10 | Helped |
| Model the CRF in the logarithmic rather than linear domain (ablation) | HDR-GS tone mapper · Synthetic HDR | PSNR: 26.18 -> 38.31 | Helped |
| Train with exposures {t1,t3,t5} rather than {t1,t5} (ablation) | HDR-GS · Synthetic HDR | PSNR: 32.06 -> 38.31 | Helped |
Surprising findings
- Modeling the CRF directly in the linear domain gave much worse results (26.18 dB HDR) than the logarithmic-domain formulation (38.31 dB).
Limitations
- Memory usage of 3DGS-based methods is non-trivial and may be unaffordable on low-RAM mobile devices.
- Real datasets lack HDR ground truth, so HDR evaluation is only on synthetic scenes.
Experimental setup
Datasets: multi-view HDR datasets (HDR-NeRF data, 4 real and 8 Blender synthetic scenes)
Benchmarks: multi-view HDR synthetic scenes (HDR-NeRF dataset) multi-view HDR real scenes
Metrics: PSNR SSIM LPIPS training time (min) inference speed (fps)
Baselines: NeRF 3DGS NeRF-W HDR-NeRF
Builds on: 3D Gaussian Splatting HDR-NeRF Debevec and Malik CRF calibration
Closest papers by meaning
- GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping (CVPR 2025 · similarity 0.93)
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.92)
- EvHDR-GS: Event-guided HDR Video Reconstruction with 3D Gaussian Splatting (AAAI 2025 · similarity 0.91)
- Towards Real-World HDR Video Reconstruction: A Large-Scale Benchmark Dataset and A Two-Stage Alignment Network (CVPR 2024 · similarity 0.88)
- Wild-GS: Real-Time Novel View Synthesis from Unconstrained Photo Collections (NeurIPS 2024 · similarity 0.87)
- From Chaos to Clarity: 3DGS in the Dark (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.