Research map / 3D vision / NeRF and radiance fields
HybridNeRF: Efficient Neural Rendering via Adaptive Volumetric Surfaces
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 HybridNeRF, a hybrid surface-volume neural radiance field with spatially adaptive surfaceness (per-voxel beta), distance-adjusted Eikonal loss and sphere-tracing/texture-based real-time rendering at 2K resolution.
Problem
Neural radiance fields render slowly because volume rendering needs many samples and model queries per ray, while real-time VR-resolution rendering needs high fidelity including thin, semi-transparent, and view-dependent structures.
Approach
Replace the global SDF-to-density temperature beta with a spatially varying surfaceness field beta(x) learned via an Eikonal-loss criterion, so most of the scene renders as a surface with few samples while fine/translucent regions remain volumetric; add distance-weighted Eikonal loss for backgrounds and render with sphere tracing and hardware texture interpolation.
Key results
- On Eyeful Tower, HybridNeRF reaches 31.57 PSNR overall at 45.78 FPS at 2K x 2K on an RTX 4090, the only method besides 3DGS and MERF to reach the 36 FPS VR target with a >1.5 dB gain over those.
- On ScanNet++, HybridNeRF achieves 24.64 PSNR at 41.90 FPS, near Zip-NeRF (24.79) while rendering over 400x faster.
- On MipNeRF-360, HybridNeRF achieves 27.52 PSNR overall, comparable to the best real-time methods such as 3D Gaussian splatting (27.16).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use spatially adaptive β(x) instead of a learned global β (ablation) | HybridNeRF · Eyeful Tower | FPS: 28.79 -> 45.78 | Helped |
| use spatially adaptive β(x) instead of global β = 2000 (ablation) | HybridNeRF · Eyeful Tower | PSNR: 27.16 -> 31.57 | Helped |
| add distance-adjusted Eikonal loss (ablation) | HybridNeRF · Eyeful Tower | PSNR: 29.97 -> 31.57 | Helped |
| add MLP distillation (ablation) | HybridNeRF · Eyeful Tower | PSNR: 31.65 -> 31.57 | No clear effect |
| add MLP distillation (ablation) | HybridNeRF · Eyeful Tower | FPS: 35.25 -> 45.78 | Helped |
| use CUDA textures for feature storage (ablation) | HybridNeRF · Eyeful Tower | FPS: 28.48 -> 45.78 | Helped |
| use CUDA textures for feature storage (ablation) | HybridNeRF · Eyeful Tower | PSNR: 31.62 -> 31.57 | No clear effect |
| use HybridNeRF instead of Zip-NeRF | Eyeful Tower overall · Eyeful Tower overall | LPIPS: 0.216 -> 0.198 | Helped |
| use HybridNeRF instead of Zip-NeRF | Mip-NeRF 360 overall · Mip-NeRF 360 overall | PSNR: 28.49 -> 27.52 | Hurt |
| use HybridNeRF instead of 3D Gaussian splatting | Mip-NeRF 360 overall · Mip-NeRF 360 overall | LPIPS: 0.153 -> 0.167 | Hurt |
| use HybridNeRF instead of Zip-NeRF | ScanNet++ · ScanNet++ | PSNR: 24.79 -> 24.64 | Hurt |
Surprising findings
- A global learned beta gives the highest quality but slows rendering, while a fixed beta of 2000 renders faster but loses substantial quality, motivating spatially adaptive surfaceness.
- Directly optimizing beta via gradient descent overly relaxes SDF correctness, so the authors instead use Eikonal loss as the indicator for increasing beta.
Limitations
- Dense 3D grids and triplanes consume significantly more memory than hash tables, and training is memory-intensive.
- Training is about 2x slower than iNGP due to extra backpropagation for Eikonal regularization.
- A significant speed gap remains relative to Gaussian splatting approaches.
Experimental setup
Datasets: Eyeful Tower MipNeRF-360 ScanNet++
Benchmarks: Eyeful Tower MipNeRF-360 ScanNet++
Metrics: PSNR SSIM LPIPS FPS
Baselines: iNGP VR-NeRF Zip-NeRF MERF VolSDF 3D Gaussian splatting BakedSDF Mip-NeRF 360 NeRF NeRF++
Builds on: VolSDF Instant-NGP MERF Zip-NeRF BakedSDF
Closest papers by meaning
- GeoNeRF: Generalizing NeRF With Geometry Priors (CVPR 2022 · similarity 0.89)
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.89)
- ProbeSDF: Light Field Probes For Neural Surface Reconstruction (CVPR 2025 · similarity 0.88)
- Few-Shot Neural Radiance Fields under Unconstrained Illumination (AAAI 2024 · similarity 0.87)
- FrugalNeRF: Fast Convergence for Extreme Few-shot Novel View Synthesis without Learned Priors (CVPR 2025 · similarity 0.87)
- DySurface: Consistent 4D Surface Reconstruction via Bridging Explicit Gaussians and Implicit Functions (NeurIPS 2026 · similarity 0.87)
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.