Research map / 3D vision / Gaussian splatting
Subsurface Scattering for 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.
Extends 3D Gaussian Splatting with PBR materials, deferred shading, and a small MLP that predicts subsurface scattering residuals and incident light, trained on a new OLAT multi-view SSS dataset for real-time relighting.
Problem
Reconstructing and relighting translucent objects with subsurface scattering from multi-view OLAT data is expensive with NeRF-based methods, and 3D Gaussian Splatting cannot capture volumetric SSS effects.
Approach
Combine explicit 3D Gaussians with a spatially varying PBR BRDF, an implicit small-MLP SSS residual and jointly predicted incident light (with ray-traced visibility), and shade in image space via deferred shading to get sharp speculars at real-time rates.
Key results
- On the synthetic dataset at 800x800, the method reaches 37.35 PSNR and 0.986 SSIM at 161 FPS with about 2 hours of training on one RTX 4090.
- At KiloOSF's 256x256, 500-image setting, the method reaches 35.01 PSNR on synthetic data versus 25.91 for KiloOSF, with under 1 hour training versus over 20 hours.
- On real-world data at 256x256, the method reaches 26.61 PSNR and 94.5 FPS versus KiloOSF's 23.24 PSNR and 14.4 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SSS GS instead of KiloOSF | Novel-view synthesis · Synthetic test views, 500 images, 256² | PSNR: 25.91 ±1.88 -> 35.01 ±1.01 | Helped |
| use SSS GS instead of KiloOSF | Novel-view synthesis · Real-world test views, 500 images, 256² | PSNR: 23.24 ±1.58 -> 26.61 ±0.09 | Helped |
| add deferred shading (ablation) | SSS GS · Component ablation, two synthetic and two real-world scenes | PSNR: 35.61 ±2.28 -> 36.16 ±3.76 | Helped |
| add the PBR shading model (ablation) | SSS GS · Component ablation, two synthetic and two real-world scenes | PSNR: 35.00 ±1.98 -> 36.16 ±3.76 | Helped |
| use a joint MLP for residual and incident light prediction (ablation) | SSS GS · Component ablation, two synthetic and two real-world scenes | PSNR: 31.68 ±0.66 -> 36.16 ±3.76 | Helped |
| add the SSS residual (ablation) | SSS GS · Component ablation, two synthetic and two real-world scenes | PSNR: 30.23 ±2.28 -> 36.16 ±3.76 | Helped |
Surprising findings
- Removing the joint MLP for SSS and incident light degrades PSNR substantially (36.16 to 31.68), suggesting the shared network aids learning of global light transport.
Limitations
- Intricate heterogeneous or angularly dependent SSS may only be roughly approximated by the MLP.
- Changing geometry or BSSRDF requires retraining the SSS representation.
- Screen-space shading limits shadow resolution and complicates transparent refractive surfaces.
Experimental setup
Datasets: Blender Cycles synthetic OLAT renders (5 objects) real-world light-stage OLAT captures (15 objects)
Benchmarks: SSS OLAT synthetic test set SSS OLAT real-world test set
Metrics: PSNR SSIM LPIPS FPS training time
Baselines: Relightable 3D Gaussians KiloOSF
Builds on: 3D Gaussian Splatting Relightable 3D Gaussians Object-Centric Neural Scattering Functions OpenPBR Disney BRDF
Closest papers by meaning
- MaterialRefGS: Reflective Gaussian Splatting with Multi-view Consistent Material Inference (NeurIPS 2025 · similarity 0.91)
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.91)
- OMG: Opacity Matters in Material Modeling with Gaussian Splatting (ICLR 2025 · similarity 0.90)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.89)
- Geometry Field Splatting with Gaussian Surfels (CVPR 2025 · similarity 0.89)
- Ref-GS: Directional Factorization for 2D Gaussian Splatting (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.