Research map / 3D vision / Gaussian splatting
GI-GS: Global Illumination Decomposition on 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 GI-GS, a 3DGS-based inverse rendering framework that computes indirect lighting and occlusion via screen-space path tracing on a deferred-shading G-buffer, improving novel view synthesis and relighting over 3DGS-based inverse rendering baselines.
Problem
Existing 3DGS-based inverse rendering methods model indirect lighting as static per-Gaussian attributes or baked volumes, so they cannot faithfully model indirect illumination during relighting.
Approach
Render a G-buffer (depth, normal, BRDF) from 3D Gaussians, apply PBR only for direct lighting with a learnable environment map, then run tile-based adaptive path tracing on the G-buffer depth to compute occlusion and indirect lighting from the first-pass image; extend to a cubemap for world-space indirect lighting.
Key results
- On the TensoIR dataset, GI-GS reaches 36.75 NVS PSNR and 31.97 albedo PSNR, surpassing all baselines, with a 28-minute training time versus 33 minutes for GS-IR.
- Relighting PSNR on TensoIR is 24.70, ranking second behind TensoIR (28.58) and above GS-IR (24.37).
- On Mip-NeRF 360, GI-GS achieves 24.01 outdoor and 29.29 indoor PSNR, outperforming the 3DGS-based inverse rendering methods GaussianShader and GS-IR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add path-tracing-based occlusion modeling (ablation) | GI-GS · TensoIR | PSNR: 36.37 -> 36.75 | Helped |
| add path-tracing-based occlusion modeling (ablation) | GI-GS · Mip-NeRF 360 | PSNR: 26.24 -> 26.35 | Helped |
| add indirect lighting (ablation) | GI-GS · TensoIR | PSNR: 35.60 -> 36.75 | Helped |
| add indirect lighting (ablation) | GI-GS · Mip-NeRF 360 | PSNR: 25.89 -> 26.35 | Helped |
| increase sampled rays from 16 to 64 (ablation) | GI-GS · TensoIR | PSNR: 36.25 -> 36.75 | Helped |
| increase sampled rays from 64 to 256 (ablation) | GI-GS · TensoIR | PSNR: 36.75 -> 36.82 | Helped |
| increase sampled rays from 16 to 64 (ablation) | GI-GS · Mip-NeRF 360 | PSNR: 26.19 -> 26.35 | Helped |
| increase sampled rays from 64 to 256 (ablation) | GI-GS · Mip-NeRF 360 | PSNR: 26.35 -> 26.42 | Helped |
| extend path tracing from screen space to world space with a cubemap (ablation) | GI-GS · Mip-NeRF 360 | PSNR: 26.35 -> 23.81 (Outdoor); 29.29 -> 29.07 (Indoor) | Hurt |
| use GI-GS rather than GS-IR | 3D Gaussian Splatting inverse rendering · TensoIR (NVS) | PSNR: 35.33 -> 36.75 | Helped |
Surprising findings
- Ablation shows that removing indirect lighting hurts quality more than removing occlusion (TensoIR PSNR 35.60 vs 36.37), per the authors.
- The cubemap world-space extension slightly lowers quantitative metrics on Mip-NeRF 360 (e.g., 29.07 vs 29.29 indoor PSNR) while improving smoothness of occluded regions.
Limitations
- Indirect illumination ignores the specular component because Monte Carlo sampling for it is complex.
- An environment map as direct lighting may fail to capture spatially varying lighting in complex real scenes.
- Path tracing accuracy depends on high-quality geometry reconstruction.
Experimental setup
Datasets: TensoIR synthetic dataset Mip-NeRF 360
Benchmarks: TensoIR synthetic dataset Mip-NeRF 360
Metrics: PSNR SSIM LPIPS
Baselines: NeRFactor InvRender NVDiffrec TensoIR GaussianShader GS-IR NeRF++ Instant NGP Mip-NeRF 360 3DGS
Builds on: 3D Gaussian Splatting GS-IR split-sum approximation GOF
Closest papers by meaning
- GS-IR: 3D Gaussian Splatting for Inverse Rendering (CVPR 2024 · similarity 0.94)
- SVG-IR: Spatially-Varying Gaussian Splatting for Inverse Rendering (CVPR 2025 · similarity 0.91)
- Ref-GS: Directional Factorization for 2D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- Inverse Rendering using Multi-Bounce Path Tracing and Reservoir Sampling (ICLR 2025 · similarity 0.90)
- Luminance-GS: Adapting 3D Gaussian Splatting to Challenging Lighting Conditions with View-Adaptive Curve Adjustment (CVPR 2025 · similarity 0.89)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 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.