Research map / 3D vision / Gaussian splatting
From Chaos to Clarity: 3DGS in the Dark
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 a self-supervised framework that jointly extracts noise and reconstructs HDR 3D Gaussian Splatting from limited noisy raw images using a noise-robust reconstruction loss with a physical noise prior.
Problem
3DGS trained directly on noisy raw images overfits noise, producing elongated Gaussians that degrade reconstruction quality and rendering speed, especially with few training views.
Approach
A noise extractor network predicts noise from raw images, constrained by a heteroscedastic Gaussian plus fixed-pattern noise model via NLL and covariance losses, while the 3DGS branch is optimized toward pseudo noise-free raw images using a RawNeRF-style scaling loss on denoised targets.
Key results
- Full-view training: Ours reaches 23.53 RGB PSNR and 80 FPS vs HDR Scaffold-GS 22.69 PSNR and 73 FPS.
- Limited views: about 3 dB PSNR gain over HDR Scaffold-GS and about 4 dB over LDR Scaffold-GS, with about 4x faster FPS at 4 views.
- Outperforms two-stage denoiser+3DGS pipelines by about 1 dB PSNR across training view counts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the noise-robust reconstruction loss instead of RawNeRF loss on raw images | Scaffold-GS on RawNeRF limited-view settings · RawNeRF limited views (4, 8, 12, 16, 20 views) | PSNR: approximately 3 dB | Helped |
| Train on raw images rather than ISP-processed RGB images | Scaffold-GS on RawNeRF limited-view settings · RawNeRF limited views (4, 8, 12, 16, 20 views) | PSNR: approximately a 4 dB improvement | Helped |
| Use the noise-robust reconstruction loss instead of RawNeRF loss on raw images | Scaffold-GS on RawNeRF limited 4-view setting · RawNeRF limited 4 views | FPS: about 4 times faster | Helped |
| Use the proposed method instead of two-stage denoiser + 3DGS pipelines | Scaffold-GS on RawNeRF limited-view settings · RawNeRF limited views | PSNR: 1 dB | Helped |
| Add lens distortion correction (ablation) | Ours on RawNeRF full views · RawNeRF full views | Raw PSNR: 59.33 -> 59.49 | Helped |
| Set λnd=0.5 and λcov=20 rather than λnd=5 and λcov=20 (ablation) | Ours on RawNeRF full views · RawNeRF full views | Raw PSNR: 59.49 -> 56.36 | Hurt |
| Set λnd=10 and λcov=20 rather than λnd=5 and λcov=20 (ablation) | Ours on RawNeRF full views · RawNeRF full views | Raw PSNR: 59.49 -> 56.84 | Hurt |
| Set λnd=5 and λcov=0 rather than λnd=5 and λcov=20 (ablation) | Ours on RawNeRF full views · RawNeRF full views | Raw PSNR: 59.49 -> 55.38 | Hurt |
| Set λnd=5 and λcov=2 rather than λnd=5 and λcov=20 (ablation) | Ours on RawNeRF full views · RawNeRF full views | Raw PSNR: 59.49 -> 57.10 | Hurt |
| Set λnd=0.5 and λcov=50 rather than λnd=5 and λcov=20 (ablation) | Ours on RawNeRF full views · RawNeRF full views | Raw PSNR: 59.49 -> 56.72 | Hurt |
Surprising findings
- Noise in raw images causes 3DGS to produce many elongated Gaussians, which reduces FPS as well as quality.
- Some high-frequency signals resemble noise, causing over-smoothing in full-view training.
Limitations
- Noise extractor and covariance loss add training time compared to standard Scaffold-GS.
- High-frequency signals resembling noise get over-smoothed, especially with full views.
- Evaluated only on the RawNeRF iPhone X dataset with a noise model calibrated per camera.
Experimental setup
Datasets: RawNeRF dataset SID dataset
Benchmarks: RawNeRF dataset
Metrics: PSNR SSIM LPIPS FPS Raw PSNR
Baselines: LDR Scaffold-GS HDR Scaffold-GS BM3D ELD PMN RViDeNet Neighbor2Neighbor LGBPN
Builds on: 3D Gaussian Splatting Scaffold-GS RawNeRF ELD PMN
Closest papers by meaning
- GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping (CVPR 2025 · similarity 0.90)
- 3DGS-Enhancer: Enhancing Unbounded 3D Gaussian Splatting with View-consistent 2D Diffusion Priors (NeurIPS 2024 · similarity 0.90)
- EvHDR-GS: Event-guided HDR Video Reconstruction with 3D Gaussian Splatting (AAAI 2025 · similarity 0.90)
- DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos (NeurIPS 2025 · similarity 0.89)
- DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus (NeurIPS 2024 · similarity 0.89)
- Robust Neural Rendering in the Wild with Asymmetric Dual 3D Gaussian Splatting (NeurIPS 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.