Research map / 3D vision / NeRF and radiance fields

FrugalNeRF: Fast Convergence for Extreme Few-shot Novel View Synthesis without Learned Priors

CVPR 2025 · 3D vision › NeRF and radiance fields · approach: Sparse-input NeRF regularization · 8 citations (OpenAlex)

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 FrugalNeRF, a TensoRF-based few-shot NeRF using weight-sharing multi-scale voxels and cross-scale geometric adaptation that picks per-ray pseudo-GT depth by reprojection error, reaching strong quality with ~10-minute training and no learned priors.

Problem

Few-shot novel view synthesis with NeRF suffers from overfitting with extremely few input views (2-4) and long training times, and existing fixes rely on frequency scheduling or pre-trained priors.

Approach

Weight-sharing density and appearance voxels downsampled across multiple scales; for each ray, the scale with the lowest cross-view reprojection error (warping rendered color to the nearest training view, including sampled novel views) provides pseudo-GT depth that supervises all scales, removing voxel upsampling schedules and learned priors.

Key results

• On LLFF with 2 views, FrugalNeRF reaches 18.07 PSNR / 0.35 LPIPS in 10 minutes without priors, versus FreeNeRF's 17.55 PSNR / 1.5 hours.
• On DTU with 2 views, FrugalNeRF reaches 19.72 PSNR / 0.16 LPIPS in 6 minutes without priors, outperforming FreeNeRF's 18.05 PSNR.
• With monocular depth regularization, FrugalNeRF reaches 18.26 PSNR on LLFF (2 views) and 20.77 PSNR on DTU (2 views) while training in 11 and 7 minutes.

Measured effects reported in the paper

Surprising findings

• Without the geometric adaptation loss, FrugalNeRF falls short of FreeNeRF, indicating the voxel representation alone is insufficient.
• Monocular-depth-supervised dense depth loss alone (Ld) initially performs well but later suffers higher depth error from scale mismatches than the cross-scale geometric loss.

Limitations

• Relies on accurate camera poses for training.
• Struggles with large viewpoint changes or very sparse views such as 360-degree scenes.
• Novel-view losses do not fully resolve unseen regions in extreme few-shot settings.

Experimental setup

Datasets: LLFF DTU RealEstate-10K

Benchmarks: LLFF DTU RealEstate-10K

Metrics: PSNR SSIM LPIPS training time

Baselines: TensoRF DS-NeRF FreeNeRF ViP-NeRF SimpleNeRF RegNeRF DDP-NeRF GeCoNeRF SparseNeRF FSGS ZeroRF

Builds on: TensoRF FreeNeRF SparseNeRF VGOS

Closest papers by meaning

• Few-Shot Neural Radiance Fields under Unconstrained Illumination (AAAI 2024 · similarity 0.90)
• AltNeRF: Learning Robust Neural Radiance Field via Alternating Depth-Pose Optimization (AAAI 2024 · similarity 0.90)
• GoLF-NRT: Integrating Global Context and Local Geometry for Few-Shot View Synthesis (CVPR 2025 · similarity 0.90)
• ZeroRF: Fast Sparse View 360deg Reconstruction with Zero Pretraining (CVPR 2024 · similarity 0.89)
• CF-NeRF: Camera Parameter Free Neural Radiance Fields with Incremental Learning (AAAI 2024 · similarity 0.89)
• ProvNeRF: Modeling per Point Provenance in NeRFs as a Stochastic Field (NeurIPS 2024 · similarity 0.89)