Research map / 3D vision / Gaussian splatting
6DGS: Enhanced Direction-Aware Gaussian Splatting for Volumetric 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 6D Gaussian Splatting (6DGS), which slices 6D spatial-angular Gaussians into view-conditioned 3D Gaussians with SH color, view-dependent opacity and SVD-based adaptive control, to achieve higher quality with fewer points than 3DGS and N-DG while rendering in real time.
Problem
Novel view synthesis of scenes with strong view-dependent effects (e.g., physically based ray-traced volumes and glossy/translucent objects) while maintaining real-time rendering; prior N-DG 6D Gaussians are sub-optimal in representation and density control.
Approach
Extends N-DG's 6D Gaussians by using spherical harmonics color, a view-direction-dependent opacity modulation with a tunable exponent lambda_opa (fixed or learnable), and SVD of the conditional covariance to recover scale and rotation for 3DGS-style densification; the 6D Gaussian is sliced into a conditional 3D Gaussian per view so the 3DGS rasterizer can be reused.
Key results
- On 6DGS-PBR, 6DGS averages 37.61 dB PSNR versus 27.53 dB for 3DGS and 36.10 dB for N-DG, with about 67K points versus 115K for 3DGS.
- On Synthetic NeRF, 6DGS matches 3DGS (33.26 vs 33.25 dB PSNR) while using about 23% fewer Gaussian points.
- 6DGS-flash (with FlashGS) reaches an average of 326.3 FPS on 6DGS-PBR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace 3DGS with 6DGS | 3DGS · 6DGS-PBR avg | PSNR: 27.53 -> 37.61; +10.08 dB | Helped |
| Replace N-DG with 6DGS | N-DG · 6DGS-PBR avg | PSNR: 36.10 -> 37.61; +1.51 dB | Helped |
| Replace 3DGS with 6DGS | 3DGS · Synthetic NeRF avg | PSNR: 33.25 -> 33.26 | No clear effect |
| Replace N-DG with 6DGS | N-DG · Synthetic NeRF avg | PSNR: 29.00 -> 33.26 | Helped |
| Add spherical harmonics for view-dependent color instead of per-Gaussian RGB (ablation) | 6DGS · 6DGS-PBR avg | PSNR: 35.85 -> 37.61 | Helped |
| Add view-dependent opacity modulation with f_cond (ablation) | 6DGS · 6DGS-PBR avg | PSNR: 33.31 -> 37.61 | Helped |
| Set λ_opa to 0.35 rather than 1 (ablation) | 6DGS · 6DGS-PBR avg | PSNR: 36.05 -> 37.49 | Helped |
| Use learnable λ_opa rather than fixed λ_opa=0.35 (ablation) | 6DGS · 6DGS-PBR avg | PSNR: 37.49 -> 37.61 | Helped |
| Raise minimum opacity threshold from 0.005 to 0.01 (ablation) | 6DGS · 6DGS-PBR avg | PSNR: 37.56 -> 37.49 | No clear effect |
| Add spherical harmonics for view-dependent color instead of per-Gaussian RGB (ablation) | 6DGS · Synthetic NeRF avg | PSNR: 32.64 -> 33.26 | Helped |
| Add view-dependent opacity modulation with f_cond (ablation) | 6DGS · Synthetic NeRF avg | PSNR: 33.12 -> 33.20 | Helped |
| Set λ_opa to 0.35 rather than 1 (ablation) | 6DGS · Synthetic NeRF avg | PSNR: 33.21 -> 33.20 | No clear effect |
Surprising findings
- On Synthetic NeRF, the N-DG baseline's point count is lower than 6DGS but its quality is much lower, and the paper notes that a larger minimum opacity threshold is needed once view-dependent opacity decay is introduced.
Limitations
- Training time about 20% higher than 3DGS and the slicing step is not yet implemented in CUDA.
- Gains on standard real-world datasets (Deep Blending, Tanks & Temples) are small or absent, and the method underperforms 3DGS on some Shiny scenes.
- Evaluated on few scenes and no dynamic scenes or integration with advanced lighting models.
Experimental setup
Datasets: 6DGS-PBR Synthetic NeRF Deep Blending Tanks & Temples Shiny
Benchmarks: 6DGS-PBR Synthetic NeRF Deep Blending Tanks & Temples Shiny
Metrics: PSNR SSIM number of Gaussian points FPS training time
Baselines: 3DGS N-DG FlashGS (6DGS-flash variant)
Builds on: 3D Gaussian splatting N-dimensional Gaussians FlashGS
Closest papers by meaning
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.91)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.91)
- Hardware-Rasterized Ray-Based Gaussian Splatting (CVPR 2025 · similarity 0.91)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.91)
- 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · similarity 0.91)
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.