Research map / 3D vision / Gaussian splatting
HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis
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 HyRF, a hybrid representation combining a small set of explicit 8-parameter Gaussians with hash-grid neural fields (decoupled geometry and appearance) plus a neural background sphere, cutting 3DGS model size by over 20x while matching or exceeding its quality.
Problem
3D Gaussian Splatting requires large memory due to per-Gaussian view-dependent and anisotropic parameters, while neural-field compression of Gaussians struggles to capture high-frequency detail and distant backgrounds.
Approach
Decompose the scene into a compact set of explicit Gaussians storing only positions, isotropic scale, opacity and diffuse color, plus hash-grid neural fields split into a geometry field and a view-dependent appearance field, with residual aggregation with explicit parameters and a neural background sphere composited by alpha blending.
Key results
- On Mip-NeRF360 HyRF reaches 27.78 PSNR / 0.816 SSIM / 0.211 LPIPS at 102 FPS and 49 MB, versus 3DGS at 27.21 PSNR and 734 MB.
- On Deep Blending HyRF achieves 30.37 PSNR at 34 MB versus 3DGS 29.41 PSNR at 676 MB, and on NeRF Synthetic 33.72 PSNR at 13 MB.
- With compression, HyRF reaches 18.04 MB on Mip-NeRF360 with 27.66 PSNR, outperforming five 3DGS compression methods in size and quality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add decoupled geometry and appearance neural fields (ablation) | HyRF · Tanks & Temples | PSNR: 23.78 -> 24.07 | Helped |
| add explicit Gaussian properties (ablation) | HyRF · Tanks & Temples | PSNR: 23.45 -> 24.07 | Helped |
| add neural fields to predict Gaussian properties (ablation) | HyRF · Tanks & Temples | PSNR: 22.22 -> 24.07 | Helped |
| add background rendering (ablation) | HyRF · Tanks & Temples | PSNR: 23.43 -> 24.07 | Helped |
| add visibility pre-culling (ablation) | HyRF · Tanks & Temples | FPS: 27 -> 106 | Helped |
| add explicit Gaussian color (ablation) | HyRF · Deep Blending | PSNR: 29.18 -> 30.37 | Helped |
| add explicit Gaussian scale (ablation) | HyRF · Deep Blending | PSNR: 28.74 -> 30.37 | Helped |
| add explicit Gaussian opacity (ablation) | HyRF · Deep Blending | PSNR: 30.21 -> 30.37 | Helped |
| use hybrid MLP appearance modeling instead of SH coefficients (ablation) | HyRF · Deep Blending | PSNR: 30.12 -> 30.37 | Helped |
| use hybrid MLP appearance modeling instead of SH coefficients (ablation) | HyRF · Deep Blending | Size (MB): 267 -> 34 | Helped |
| add background rendering (ablation) | HyRF · KITTI 2011_09_26_drive_0002 | PSNR: 19.42 -> 19.56 | Helped |
Surprising findings
- Removing explicit scale causes training instability and the largest quality drop among explicit-property ablations (PSNR 28.74 vs 30.37 on Deep Blending).
- Disabling the neural background actually reduces size slightly but drops PSNR by about 0.6 dB on Tanks & Temples while only marginally changing FPS.
Limitations
- Does not address aliasing inherited from 3DGS.
- Sometimes produces inaccurate surface reconstruction.
- Neural field components rely on high-end GPUs for high rendering speed; not efficient on web or integrated graphics.
Experimental setup
Datasets: MipNeRF360 Tanks & Temples Deep Blending NeRF Synthetic Mill19 Urbanscene3D KITTI
Benchmarks: MipNeRF360 Tanks & Temples Deep Blending NeRF Synthetic Mill19 Urbanscene3D KITTI Spec-GS dataset
Metrics: PSNR SSIM LPIPS FPS model size (MB) training time peak GPU memory
Baselines: Mip-NeRF360 Plenoxels Instant-NGP 3DGS Scaffold-GS MipNeRF MegaNeRF SwitchNeRF GOF Spec-GS Mini-Splatting2 DashGaussian
Builds on: 3D Gaussian Splatting Instant-NGP Mip-NeRF360 Scaffold-GS
Closest papers by meaning
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.93)
- Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · similarity 0.90)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.90)
- 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.90)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.90)
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.