Research map / 3D vision / NeRF and radiance fields
Spatial Annealing for Efficient Few-shot Neural 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 SANeRF, a spatial annealing regularizer for pre-filtered hybrid NeRF representations (TriMipRF) that exponentially shrinks the sampling sphere during training, enabling efficient few-shot neural rendering.
Problem
Hybrid grid-based NeRFs overfit or converge abnormally with sparse input views, and existing few-shot regularizers add rendering cost or complex pipelines; FreeNeRF's frequency annealing relies on frequency positional encoding and is incompatible with hybrid representations.
Approach
Derives the frequency band limit of a pre-filtered tri-plane hybrid representation and shows that annealing the spatial sampling sphere radius exponentially is equivalent to FreeNeRF-style frequency annealing, implemented by changing the sampling kernel size with one line of code, plus spherical harmonic degree truncation.
Key results
- On Blender with 8 views, SANeRF reaches 24.53 PSNR in 2.5K iterations (35 s), versus TriMipRF 21.63 PSNR and FreeNeRF 24.26 PSNR (7.5 h).
- With 5K iterations SANeRF reaches 24.81 PSNR at 70 s, about 350x faster than FreeNeRF.
- On LLFF and DTU with 3 views (JAX MipNeRF), SANeRF gives 19.67 and 18.15 PSNR versus FreeNeRF 19.63 and 18.02.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add spatial annealing to TriMipRF (ablation) | TriMipRF · Blender, 8 input views | PSNR: 21.63 -> 22.83 | Helped |
| Add the spherical harmonic mask to TriMipRF (ablation) | TriMipRF · Blender, 8 input views | PSNR: 21.63 -> 22.27 | Helped |
| Combine spatial annealing with the spherical harmonic mask (ablation) | TriMipRF · Blender, 8 input views | PSNR: 21.63 -> 24.53 | Helped |
| Increase training iterations from 2.5K to 5K | SANeRF · Blender, 8 input views | PSNR: 24.53 -> 24.81 | Helped |
| Use spatial annealing instead of FreeNeRF frequency regularization | Mip-NeRF · LLFF, 3 input views | PSNR: 19.63 -> 19.67; 0.05 dB | Helped |
| Use spatial annealing instead of FreeNeRF frequency regularization | Mip-NeRF · DTU, 3 input views | PSNR: 18.02 -> 18.15; 0.1 dB | Helped |
| Apply spatial annealing to TriMipRF | TriMipRF · Blender, 8 input views | PSNR: nearly 3 dB | Helped |
| Use SANeRF instead of FreeNeRF | Few-shot neural rendering on Blender · Blender, 8 input views | PSNR: 0.27 dB | Helped |
| Use SANeRF instead of FreeNeRF | Few-shot neural rendering on Blender · Blender, 8 input views | Training speed: 700$\times$ | Helped |
| Use SANeRF instead of TriMipRF as input view count varies | TriMipRF · Blender, 8 input views | PSNR: 2.50 dB | Helped |
| Use SANeRF instead of TriMipRF as input view count varies | TriMipRF · Blender, 20 input views | PSNR: 1.22 dB | Helped |
Surprising findings
- Spatial annealing alone raises PSNR on Blender by 1.2 dB but does not fix color distortion, which the spherical harmonic degree truncation addresses.
Limitations
- Gains on LLFF and DTU over FreeNeRF are marginal (up to 0.05 dB and 0.1 dB PSNR); DTU SSIM is lower than FreeNeRF.
- Spatial annealing hyperparameters (initial sphere size, decay rate, steps, stop point) are tuned per setup and differ between TriMipRF and JAX MipNeRF implementations.
- The band-limit analysis approximates the bandwidth of the learned field with an error bound rather than an exact formula.
Experimental setup
Datasets: Blender LLFF DTU
Benchmarks: Blender LLFF DTU
Metrics: PSNR SSIM LPIPS training time depth MSE
Baselines: NeRF MipNeRF TriMipRF DietNeRF InfoNeRF MixNeRF FreeNeRF VGOS FSGS DNGaussian RegNeRF
Builds on: TriMipRF FreeNeRF MipNeRF grid tangent kernel
Closest papers by meaning
- FrugalNeRF: Fast Convergence for Extreme Few-shot Novel View Synthesis without Learned Priors (CVPR 2025 · similarity 0.88)
- Few-Shot Neural Radiance Fields under Unconstrained Illumination (AAAI 2024 · similarity 0.88)
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.87)
- Ced-NeRF: A Compact and Efficient Method for Dynamic Neural Radiance Fields (AAAI 2024 · similarity 0.87)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.87)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · 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.