Research map / 3D vision / NeRF and radiance fields
LookCloser: Frequency-aware Radiance Field for Tiny-Detail Scene
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 FA-NeRF, a frequency-aware radiance field that quantifies per-point 3D frequency via progressive patch-based image regression and uses a frequency grid, feature re-weighting, frequency-averaged sampling, and adaptive ray marching to reconstruct both scene structure and tiny details.
Problem
NeRF methods either capture high-frequency local detail or broad scene structure but struggle to model scenes whose frequency content varies by orders of magnitude across views and regions.
Approach
Estimate a 3D frequency for scene points by progressively increasing encoded frequency until image patches reach SSIM threshold, project 2D frequency to 3D using focal length and depth, store it in a frequency grid, and use it to re-weight hash-grid features, guide sampling density and ray marching interval.
Key results
- On the Multi-frequency dataset detail view (4032x3024), FA-NeRF reaches 26.29 PSNR / 0.843 SSIM versus 24.29 PSNR / 0.802 SSIM for 3D-GS.
- On the structure view (600x600) of the Multi-frequency dataset, FA-NeRF achieves 32.44 PSNR versus 30.97 for iNGP-Big.
- Renders at 20 FPS on a single RTX 4090 GPU while rendering high-frequency details.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FA-NeRF instead of iNGP-Big | Multi-frequency dataset · Structure View (600× 600) | PSNR: 30.97 -> 32.44 | Helped |
| Use FA-NeRF instead of 3D-GS | Multi-frequency dataset · Detail View (4032× 3024) | PSNR: 24.29 -> 26.29 | Helped |
| Use FA-NeRF instead of Mip-NeRF360 | MipNeRF-360 dataset · MipNeRF-360 dataset | PSNR: 31.49 -> 31.20 | Hurt |
| Use FA-NeRF instead of 3D-GS | MipNeRF-360 dataset · MipNeRF-360 dataset | SSIM: 0.926 -> 0.931 | Helped |
| Use FA-NeRF instead of 3D-GS | Tanks&Temples dataset · Tanks&Temples dataset | SSIM: 0.831 -> 0.821 | Hurt |
| Add the Frequency Grid instead of using a uniform frequency value (ablation) | FA-NeRF · Music Room, normal-res(600×600) | PSNR: 31.95 -> 33.52 | Helped |
| Add the Frequency Grid instead of using a uniform frequency value (ablation) | FA-NeRF · Music Room, high-res(4032× 3024) | PSNR: 24.90 -> 26.97 | Helped |
| Add feature re-weighting (ablation) | FA-NeRF · Music Room, normal-res(600×600) | PSNR: 33.58 -> 33.52 | Hurt |
| Add feature re-weighting (ablation) | FA-NeRF · Music Room, high-res(4032× 3024) | PSNR: 26.73 -> 26.97 | Helped |
| Add frequency-averaged sampling (ablation) | FA-NeRF · Music Room, high-res(4032× 3024) | PSNR: 25.84 -> 26.97 | Helped |
| Add adaptive ray marching (ablation) | FA-NeRF · Music Room, high-res(4032× 3024) | PSNR: 25.42 -> 26.97 | Helped |
Surprising findings
- Disabling feature re-weighting improves normal-resolution results while degrading high-resolution performance, suggesting low-frequency signals overshadow high-frequency ones under limited capacity.
- BungeeNeRF adapted to multi-frequency scenes performs similarly to its Mip-NeRF baseline, indicating divide-and-conquer multi-scale strategies do not solve frequency variation.
Limitations
- Rendering speed of 20 FPS leaves room for further acceleration, as the authors note.
- Frequency estimation relies on a predefined SSIM threshold and patch-based progressive regression, which adds preprocessing.
- Evaluated mainly on indoor scenes of its own Multi-frequency dataset with limited scene count (five).
Experimental setup
Datasets: Multi-frequency dataset MipNeRF-360 Tanks&Temples
Benchmarks: Multi-frequency dataset MipNeRF-360 Tanks&Temples
Metrics: PSNR SSIM LPIPS (VGG)
Baselines: 3D-GS Mip-NeRF360 Instant-NGP TensoRF BungeeNeRF
Builds on: Instant-NGP Mip-NeRF 360 NeRF 3D Gaussian Splatting
Closest papers by meaning
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.87)
- BLiRF: Bandlimited Radiance Fields for Dynamic Scene Modeling (AAAI 2024 · similarity 0.87)
- FrugalNeRF: Fast Convergence for Extreme Few-shot Novel View Synthesis without Learned Priors (CVPR 2025 · similarity 0.86)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.86)
- NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods (NeurIPS 2025 · similarity 0.85)
- ExtraNeRF: Visibility-Aware View Extrapolation of Neural Radiance Fields with Diffusion Models (CVPR 2024 · similarity 0.85)
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.