Research map / 3D vision / Gaussian splatting
GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping
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 GaussHDR, a 3D Gaussian Splatting method for HDR novel view synthesis that unifies 3D and 2D local tone mapping via a residual context-feature tone mapper and uncertainty-weighted loss combination.
Problem
HDR novel view synthesis from multi-exposure LDR views is unstable under 3D tone mapping and degrades LDR fitting under 2D tone mapping, and global tone mappers limit modeling capacity.
Approach
Assigns each Gaussian a context feature rendered alongside colour so a shared residual MLP tone mapper (global mapping plus learned residual) can act locally in both 3D and 2D tone mapping; the two LDR losses are combined with learned uncertainties that adaptively weight them.
Key results
- On HDR-NeRF real scenes (Exp-3, LDR-OE), Ours (Scaffold-GS) reaches 36.77 PSNR vs 35.96 for HDR-Scaffold-GS.
- On HDR-NeRF synthetic scenes, Ours (Scaffold-GS) HDR PSNR is 39.02 vs 29.22 for HDR-Scaffold-GS under Exp-3.
- Ablation (Scaffold-GS, Exp-1) shows the full unified local tone mapping with uncertainty gives LDR-OE 35.37 and HDR 38.60 PSNR on real/synthetic, outperforming global and single-branch variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace 3D global tone mapping with 3D local tone mapping (ablation) | GaussHDR with Scaffold-GS · HDR-NeRF real scenes, LDR-OE | PSNR: 33.94 -> 34.41 | Helped |
| replace 2D global tone mapping with 2D local tone mapping (ablation) | GaussHDR with Scaffold-GS · HDR-NeRF real scenes, LDR-OE | PSNR: 32.45 -> 34.19 | Helped |
| combine 3D and 2D local tone mapping instead of global tone mapping (ablation) | GaussHDR with Scaffold-GS · HDR-NeRF real scenes, LDR-OE | PSNR: 33.22 -> 35.37 | Helped |
| add the residual term to the local tone mapper (ablation) | GaussHDR with Scaffold-GS · HDR-NeRF synthetic scenes, HDR | PSNR: 18.46 -> 38.60 | Helped |
| use uncertainty-based modulation and merging rather than a fixed weighted loss combination (ablation) | GaussHDR · different scenes | LDR PSNR | Helped |
Surprising findings
- Without the residual form (local tone mapper trained alone), HDR performance collapses to 18.46 PSNR on synthetic scenes, showing the global tone mapper initialization is essential.
Limitations
- Synthetic HDR results rely on unit exposure loss for scale fixing, which assumes synthetic setup.
- Reported HDR comparisons against HDR-GS are under a setting without HDR ground-truth supervision, so comparability depends on re-implementation.
Experimental setup
Datasets: HDR-NeRF synthetic dataset HDR-NeRF real dataset HDR-Plenoxels real dataset
Benchmarks: HDR-NeRF synthetic dataset HDR-NeRF real dataset HDR-Plenoxels real dataset
Metrics: PSNR SSIM LPIPS
Baselines: HDR-NeRF HDR-GS HDR-Scaffold-GS HDR-Plenoxels
Builds on: 3D Gaussian Splatting HDR-GS HDR-NeRF Scaffold-GS uncertainty learning for multi-task losses
Closest papers by meaning
- HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting (NeurIPS 2024 · similarity 0.93)
- EvHDR-GS: Event-guided HDR Video Reconstruction with 3D Gaussian Splatting (AAAI 2025 · similarity 0.91)
- From Chaos to Clarity: 3DGS in the Dark (NeurIPS 2024 · similarity 0.90)
- Zero-Shot Structure-Preserving Diffusion Model for High Dynamic Range Tone Mapping (CVPR 2024 · similarity 0.90)
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.89)
- Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene (ICLR 2025 · similarity 0.89)
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.