Research map / 3D vision / Gaussian splatting
SVG-IR: Spatially-Varying Gaussian Splatting for Inverse Rendering
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 Spatially-varying Gaussians (SVG) with per-vertex material and normal attributes, rendered via vertex/fragment-style shading, plus a physically-based one-bounce indirect lighting model for Gaussian inverse rendering and relighting.
Problem
Gaussian-splatting-based inverse rendering yields relighting artifacts and unnatural indirect illumination because each Gaussian has constant material and normal, and indirect lighting is modeled without physical constraints.
Approach
Generalize constant per-Gaussian material to spatially-varying material by attaching albedo, roughness and normal offsets to multiple Gaussian vertices in the tangent space, interpolate colors across the Gaussian (vertex then fragment shading), and compute indirect light via ray tracing in the radiance field with a physically-based rendering equation and one-bounce relighting.
Key results
- Relighting on TensoIR Synthetic mean PSNR 31.10 vs 27.60 for RelightGS and 28.53 for TensoIR (NeRF-based).
- NVS on ADT achieves PSNR 41.48 vs 39.72 for TensoIR.
- Ablation on TensoIR Synthetic: SVG raises relighting PSNR from 28.61 to 29.45; adding physically-based illumination raises it to 31.09.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace constant Gaussians with spatially-varying Gaussians (ablation) | SVG-IR · TensoIR Synthetic relighting | PSNR: 28.614 -> 29.447 | Helped |
| replace constant Gaussians with spatially-varying Gaussians (ablation) | SVG-IR · TensoIR Synthetic NVS | PSNR: 34.640 -> 35.794 | Helped |
| add physically-based illumination (ablation) | SVG-IR with spatially-varying Gaussians · TensoIR Synthetic relighting | PSNR: 29.447 -> 31.087 | Helped |
| add physically-based illumination (ablation) | SVG-IR with spatially-varying Gaussians · TensoIR Synthetic NVS | PSNR: 35.794 -> 36.709 | Helped |
| add environment visibility (ablation) | SVG-IR lighting model without indirect illumination · TensoIR Synthetic relighting | PSNR: 29.447 -> 30.253 | Helped |
| add environment visibility (ablation) | SVG-IR lighting model without indirect illumination · TensoIR Synthetic NVS | PSNR: 35.794 -> 36.059 | Helped |
| add radiance-field indirect illumination (ablation) | SVG-IR lighting model with environment visibility · TensoIR Synthetic relighting | PSNR: 30.253 -> 31.087 | Helped |
| add radiance-field indirect illumination (ablation) | SVG-IR lighting model with environment visibility · TensoIR Synthetic NVS | PSNR: 36.059 -> 36.709 | Helped |
| use SVG-IR instead of TensoIR | TensoIR Synthetic relighting · TensoIR Synthetic | PSNR: 28.53 -> 31.10 | Helped |
| use SVG-IR instead of RelightGS | ADT relighting · ADT | PSNR: 32.84 -> 34.69 | Helped |
Limitations
- No geometric prior, so recovery of highly specular objects is unsatisfactory.
- Roughly 50%-80% more GPU memory than 3DGS/2DGS and 10%-20% more than Relightable 3DGS due to Gaussian vertices.
- Real-time rendering maintained but training time longer than some baselines is not fully discussed.
Experimental setup
Datasets: TensoIR Synthetic Aria Digital Twin (ADT) DTU MipNeRF360
Benchmarks: TensoIR Synthetic Aria Digital Twin (ADT) DTU MipNeRF360
Metrics: PSNR SSIM LPIPS training time
Baselines: MII TensoIR GaussianShader GS-IR RelightGS
Builds on: 2D Gaussian Splatting Relightable 3DGS Disney Principled BRDF GS-IR
Closest papers by meaning
- GS-IR: 3D Gaussian Splatting for Inverse Rendering (CVPR 2024 · similarity 0.91)
- GI-GS: Global Illumination Decomposition on Gaussian Splatting for Inverse Rendering (ICLR 2025 · similarity 0.91)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · similarity 0.90)
- RNG: Relightable Neural Gaussians (CVPR 2025 · similarity 0.89)
- Decoupling Appearance Variations with 3D Consistent Features in Gaussian Splatting (AAAI 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.