Research map / 3D vision / NeRF and radiance fields
Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field 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 SVRaster, a neural-free and Gaussian-free radiance field renderer that rasterizes adaptive multi-level sparse voxels, using direction-dependent Morton ordering to avoid popping artifacts.
Problem
Real-time high-fidelity novel-view synthesis and surface reconstruction with grid-based representations is slow due to ray casting, and naive voxel grids lack adaptive levels of detail; 3DGS suffers from popping artifacts and ill-defined volume density.
Approach
Allocate adaptive multi-level sparse voxels (Octree leaves stored in a flat array, each with trilinear corner densities and SH colors) and render them with a CUDA rasterizer that sorts voxels by ray-direction-dependent Morton order, giving correct depth order without popping.
Key results
- On Mip-NeRF360, Ours reaches 27.33 PSNR, 0.185 LPIPS and 121 FPS, versus 3DGS at 27.45 PSNR, 0.216 LPIPS and 131 FPS (re-evaluated).
- On Deep Blending, Ours reaches 29.84 PSNR and 366 FPS, versus 3DGS at 29.60 PSNR and 140 FPS.
- On Tanks&Temples mesh reconstruction, Ours achieves F-score 0.40 in 11 minutes, compared with 3DGS at 0.19 in 14 minutes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SVRaster instead of 3DGS | Mip-NeRF360 · Mip-NeRF360 (4 indoor, 5 outdoor scenes) | LPIPS: 0.216 -> 0.185 | Helped |
| use SVRaster instead of 3DGS | Mip-NeRF360 · Mip-NeRF360 (4 indoor, 5 outdoor scenes) | PSNR: 27.45 -> 27.33 | Hurt |
| use SVRaster instead of 3DGS | Mip-NeRF360 · Mip-NeRF360 (4 indoor, 5 outdoor scenes) | FPS: 131 -> 121 | Hurt |
| use SVRaster instead of Zip-NeRF | Mip-NeRF360 · Mip-NeRF360 (4 indoor, 5 outdoor scenes) | LPIPS: 0.187 -> 0.185 | Helped |
| use SVRaster instead of 3DGS | Tanks&Temples · Tanks&Temples | LPIPS: 0.176 -> 0.144 | Helped |
| use SVRaster instead of 3DGS | Tanks&Temples · Tanks&Temples | PSNR: 23.75 -> 23.04 | Hurt |
| use SVRaster instead of 3DGS | Tanks&Temples · Tanks&Temples | FPS: 180 -> 125 | Hurt |
| use SVRaster instead of 3DGS | Deep Blending · Deep Blending | LPIPS: 0.244 -> 0.228 | Helped |
| use SVRaster instead of 3DGS | Deep Blending · Deep Blending | FPS: 140 -> 366 | Helped |
| use adaptive voxel sizes instead of uniform 512³ voxels (ablation) | SVRaster · Mip-NeRF360 bonsai and bicycle | LPIPS: 0.326 -> 0.200 | Helped |
| use adaptive voxel sizes instead of uniform 512³ voxels (ablation) | SVRaster · Mip-NeRF360 bonsai and bicycle | PSNR: 25.37 -> 28.01 | Helped |
| use adaptive voxel sizes instead of uniform 512³ voxels (ablation) | SVRaster · Mip-NeRF360 bonsai and bicycle | FPS: 190 -> 171 | Hurt |
Surprising findings
- Uniform voxel grids at 512^3 resolution perform much worse than adaptive voxels at similar FPS, showing adaptive level allocation matters more than raw resolution.
- On Tanks&Temples the method has worse PSNR and FPS than 3DGS, attributed to exposure variation in training views causing floaters.
Limitations
- FPS varies significantly across scenes relative to 3DGS.
- Uses much more primitives than Gaussians, giving larger model size and higher GPU memory.
- Mesh extraction produces unnecessary geometric bumps for texture details.
Experimental setup
Datasets: Mip-NeRF360 Tanks&Temples Deep Blending DTU
Benchmarks: Mip-NeRF360 Tanks&Temples Deep Blending DTU
Metrics: PSNR SSIM LPIPS FPS F-score Chamfer distance training time
Baselines: 3DGS Zip-NeRF Plenoxels INGP Mip-NeRF360 NeuS Neuralangelo 2DGS SuGaR Voxurf
Builds on: 3D Gaussian Splatting Plenoxels 2DGS NeRF TSDF Fusion
Closest papers by meaning
- CVT-xRF: Contrastive In-Voxel Transformer for 3D Consistent Radiance Fields from Sparse Inputs (CVPR 2024 · similarity 0.90)
- Learning with Unreliability: Fast Few-shot Voxel Radiance Fields with Relative Geometric Consistency (CVPR 2024 · similarity 0.89)
- GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction (NeurIPS 2025 · similarity 0.89)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.88)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.88)
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.88)
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.