FourieRF: Few-Shot NeRFs via Progressive Fourier Frequency Control

Bingchen Gong; Diego Gomez; Maks Ovsjanikov

REVIEW 3 major objections 5 minor 25 references

FourieRF: Few-Shot NeRFs via Progressive Fourier Frequency Control

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FourieRF claims that band-limiting the Fourier spectrum of TensoRF feature grids during training prevents few-shot NeRF artifacts and speeds up training.

desk verdict Useful, fast few-shot NeRF baseline that overclaims SOTA; the core frequency-mask math as written is broken and the reported schedule is internally inconsistent. read the letter →

arxiv 2502.01405 v1 pith:SR2LWIIG submitted 2025-02-03 cs.CV

Diego Gomez , Bingchen Gong , Maks Ovsjanikov This is my paper · ORCID

classification cs.CV

keywords few-shotNeRFneuralradiancefieldsfrequencycurriculumFourierdomainmaskingTensosparse-viewnovelviewsynthesiscoarse-to-fineoptimization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

FourieRF is a training-time cure for the few-shot NeRF collapse: when only a handful of views are available, radiance fields overfit early, producing high-frequency floaters that later training cannot remove. The paper argues that explicitly clipping the higher Fourier coefficients of the TensoRF feature grids and linearly releasing them during training forces the model to first lock in coarse geometry and only then add detail. The result is artifact-free rendering from 3 to 9 views, on par with FreeNeRF and ZeroRF, while training in about 10 minutes at 0.93x the time of vanilla TensoRF. If correct, this shows that no pretrained modules or data priors are needed for robust few-shot reconstruction — a frequency curriculum suffices.

What carries the argument

The load-bearing mechanism is a Fourier-domain masking schedule applied to the feature grids of TensoRF. For each 1D feature vector $v$, the operation is $\hat{v} = \operatorname{IFFT}(\operatorname{FFT}(v) \odot \alpha(f_t))$, and for each 2D matrix a circular mask $\beta(f_t)$ clips coefficients outside a radius; the threshold $f_t$ increases linearly as $f_t = f_{t-1} + \Delta$. This mask is applied before gradients are accumulated, so it adds no differentiability complications. The 2D circular mask and the per-mode treatment of the VM decomposition let the method control the spatial frequency of geometry and appearance independently, which is what turns the curriculum into a stable coarse-to-fine trajectory.

What would settle it

Train FourieRF on a scene with a large untextured region or strong occlusion and measure the geometry recovered at the low-frequency stage; if the coarse shape deviates from the ground-truth surface, then the curriculum's premise fails and final quality will not beat vanilla TensoRF.

Watch

Extended reading notes

Core claim

The central discovery is that the failure mode of few-shot NeRF — catastrophic overfitting that fills the scene with floaters — can be avoided by controlling the maximum Fourier frequency of the learned feature grids. The method projects the 1D vectors and 2D matrices of TensoRF's CP/VM tensor decomposition into Fourier space, zeroes coefficients above a time-dependent threshold, and transforms back; the threshold starts very low (0.01% of coefficients on real scenes, 30% on synthetic) and grows linearly each iteration. Under this schedule, low-frequency-constrained optimization recovers a correct coarse shape even from 3 views, and the progressive release of higher frequencies refines that shape without reintroducing artifacts. On the Blender synthetic dataset (4 and 6 views) and the LLFF real dataset (3, 6, and 9 views), FourieRF achieves PSNR, SSIM, and LPIPS on par with FreeNeRF and ZeroRF, while its per-iteration Fourier pass adds virtually no cost.

Load-bearing premise

The curriculum works only if low-frequency-constrained optimization recovers a correct coarse shape; if it locks in a wrong shape, the later high-frequency stages just refine that error.

Editorial extensions

If this is right

Few-shot NeRF reconstruction can be made robust without any pretrained component: a simple frequency schedule on the feature grids reproduces the gains of learned priors.
Grid-based radiance fields, which are normally the fastest to train, can be kept fast while becoming sparse-view capable, cutting training from hours or days to roughly 10 minutes.
The prior is scene-agnostic, so the same recipe transfers from synthetic objects to real forward-facing scenes without retuning the core mechanism.
Recovering a clean coarse shape early in training can serve as a stable initialization for later stages, which is the reason the linear schedule does not drift into floaters.
Because the method applies to both CP and VM decompositions, it sets a simple baseline that other grid-based few-shot approaches can be measured against.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same band-limiting idea could be dropped into other explicit representations, such as hash grids or 3D Gaussian splatting, as a generic anti-floaters regularizer.
The linear schedule is a choice, not a necessity: an adaptive schedule driven by validation loss or by the occupancy of the density field might shorten training further or handle scenes with very different frequency content.
The paper's own limitation statement suggests a natural extension: where occlusion hides parts of the shape, the low-frequency stage cannot recover what was never observed, so combining the curriculum with a data-driven prior for inpainting is a plausible next step.
If the coarse-geometry claim transfers, then the intermediate low-frequency renders themselves could be used as cheap probes for camera registration or scene understanding tasks.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Quick read of FourieRF. The idea is simple and sensible: restrict the Fourier content of TensoRF's feature grids early in training, then relax the restriction. That gives you a few-shot prior with no pretraining and no depth, and it trains in about ten minutes. The experiments are honest enough: on Blender they're essentially tied with ZeroRF and a bit behind on PSNR in some settings, but they crush it on training time; on LLFF they're on par with FreeNeRF and more than 30x faster. If you need a fast, generic few-shot NeRF baseline, this is a genuinely useful data point. The comparison to FreeNeRF is fair, and I appreciate that they test on real scenes where ZeroRF fails.

Now the soft spots. The paper's central mechanism is described with a real flaw. Equation 2 defines a mask that keeps only the first t_alpha frequency indices, which is not Hermitian-symmetric; applied to a real feature vector, that IFFT returns complex values. The paper never says they use rfft/irfft or a symmetric mask, so the math as written does not do what the text claims. That's not a minor typo—it's the core of the method. The 2D version has the same problem: a circle centered at (d1/2, d2/2) only acts as a low-pass if the spectrum is fftshifted first, which is never mentioned.

The schedule also looks inconsistent. The main text says f0 = 0.01% and delta = (1-f0)/N, which is a sensible ramp over 10k iterations. The supplementary gives f0 = 0.3 and delta = 1/2000 for Blender, which reaches full frequency by iteration 1400 and leaves the curriculum inactive for most of training. One of these is wrong, and the supplementary's own arithmetic (1/2000 = 2e-3) is also off by a factor of four. The paper should say exactly what schedule was used, and ideally show a sensitivity curve.

Also, the "new state-of-the-art" claim in the introduction is not supported by the tables. On Blender 4-view, ZeroRF has higher PSNR (21.94 vs 21.73); on 6-view, ZeroRF is 24.73 vs 23.93. FourieRF's advantage is speed, not top-line quality. That's a fine contribution, but the sentence should be toned down.

Overall, the empirical core looks solid. The method is plausible and the results are reproducible in principle. The mathematical description needs a careful rewrite, and the schedule inconsistency needs to be resolved. I'd send this to reviewers—the idea is worth the time—but it has to go back for a major revision to fix the method definition. It's not a desk reject.

For the reading group, I'd say maybe—the idea is a useful baseline, but the paper needs a cleanup before I'd want to spend a full meeting on it.

Would I cite it? Probably yes if I work on few-shot NeRF; the speed result is meaningful. But I'd wait for the corrected version.

Referee Report

3 major / 5 minor

Summary. The paper proposes FourieRF, a few-shot NeRF method built on TensoRF that applies a progressive Fourier-domain low-pass mask to the 1D and 2D feature grids. The authors claim that starting from low Fourier frequencies and linearly increasing the cutoff during training induces a coarse-to-fine curriculum that avoids the floaters typical of few-shot grid-based NeRFs. Experiments on the Blender synthetic dataset and the LLFF real-scene dataset report results on par with FreeNeRF and ZeroRF while training in about 0.93x the time of vanilla TensoRF, with supplementary per-scene tables and an ablation over the frequency increment parameter. The paper positions the method as a fast, fully generic, training-only alternative to data-driven few-shot NeRF priors.

Significance. If the central mechanism is correctly specified, the paper would make a useful contribution: a simple, training-only frequency curriculum that achieves competitive few-shot rendering quality at a fraction of the training cost of FreeNeRF and ZeroRF. The authors provide per-scene quantitative tables, a speed comparison, a code link, and an ablation of the frequency increment, which are all valuable for reproducibility. However, the manuscript currently contains load-bearing ambiguities in the mathematical definition of the Fourier mask, an internal inconsistency in the reported frequency schedule, and a lack of ablation isolating the frequency control from the additional regularizers. These issues must be resolved before the causal claim about progressive Fourier control can be accepted.

major comments (3)

[Section 4.2, Eq. (2) and Eq. (3)] The mask described in Eq. (2) is not Hermitian-symmetric. For a real feature vector v, FFT(v) has conjugate symmetry, but the mask keeps only the low-index coefficients without mirroring them to the corresponding negative-frequency positions, so IFFT(FFT(v) ⊙ α(f_t)) is generally complex. Since TensoRF feature grids are real, the operation as written is undefined. The same issue applies to the 2D mask in Eq. (3): a circle centered at (d1/2, d2/2) is a low-pass filter only if fftshift is applied before masking and ifftshift after. Please specify whether rfft/irfft is used, whether fftshift/ifftshift is applied, or how the complex intermediate is converted back to real features. If the real part is taken, the effective operation is not an ideal band-limiting filter, and the claim that the method constrains the maximum Fourier frequency needs to be rephrased.
[Section 4.2, 'Progressive Inclusion of Coefficients'; Supplementary Material B] The reported schedule is internally inconsistent. The main text defines Δ = (1 − f0)/N, but Supplementary Material B states that for the Blender synthetic runs f0 = 0.3 and δ = 1/2000 with N = 10000 iterations. Using δ = 1/2000 and f0 = 0.3, f_t reaches 1.0 at iteration 1400, so the frequency curriculum is inactive for the remaining 8600 iterations. If instead the formula Δ = (1 − f0)/N is used, then δ should be 7 × 10^−5, not 5 × 10^−4. Please clarify which schedule was actually used, report the per-dataset values consistently, and also reconcile the main-text statement 'f0 = 0.01% of Fourier coefficients' with the supplementary values f0 = 0.3 (synthetic) and f0 = 0.01 (real).
[Section 5, Ablations; Supplementary Material B] The experiments do not isolate the contribution of the proposed Fourier curriculum. Supplementary Material B states that for synthetic scenes 'setting the weight decay to 0.2 in the optimizer is the key to removing floaters' and that real-scene experiments use FreeNeRF's occlusion regularization, a TV loss, an L1 density loss, and optional gradient scaling. Because these are strong regularizers known to combat floaters in few-shot NeRF, the quantitative gains cannot be attributed solely to progressive frequency control. Please add ablations that (a) remove the frequency control while keeping all other losses and hyperparameters, and (b) remove the additional regularizers while keeping frequency control. Without such ablations, the central claim that the Fourier curriculum is the operative mechanism is not established.

minor comments (5)

[Section 1, Introduction] The introduction says FourieRF 'establishes a new state-of-the-art,' but Tables 1 and 2 show performance that is on par with, and in several metrics below, FreeNeRF and ZeroRF (e.g., Table 2, 9-view LPIPS: Ours 0.193 vs. FreeNeRF 0.160). Please make the wording consistent with the actual results, e.g., 'on par with state-of-the-art while being substantially faster.'
[Section 4.2, Eq. (2)] The fractional value at the mask boundary, tα(ft) − floor(tα(ft)), is only defined for one index. If a full FFT is used, the corresponding negative-frequency coefficient must receive the same fractional value to preserve Hermitian symmetry; if rfft is used, the boundary handling should be stated explicitly.
[Section 4.2, 'Progressive Inclusion of Coefficients'] The threshold f_t is used in Eqs. (1)-(3) before its update rule is introduced, and the notation f_t vs. f0 and Δ vs. δ is inconsistent between the main text and the supplement. Please define all symbols at first use and use a single notation throughout.
[Table 3 and Section 5.2] The 0.93x training-time claim is reported only for the Blender dataset at 10k iterations. Please state the exact hardware and time measurement procedure, and clarify whether the LLFF experiments show the same relative speedup.
[Supplementary Material B] There is a typo: 'We build our code base in top of the TensoRF repository' should read 'on top of.' Also, the hyperparameter listing would be easier to follow if presented as a table rather than prose.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FourieRF's frequency-curriculum prior is an empirical training procedure validated against external baselines; no load-bearing self-citation, fitted-as-prediction, or definitional reduction found.

full rationale

The paper's central claim is that progressively increasing the maximal Fourier frequency of TensoRF feature grids yields artifact-free few-shot reconstruction. This is not derived from a fitted parameter or from a self-citation chain. The frequency mask and schedule in Eqs. (1)-(3) are explicit operations on the paper's own feature grids, with the schedule f_t = f_{t-1} + Delta stated independently of the reported metrics. Quantitative claims are tested on Blender and LLFF against external baselines (FreeNeRF, ZeroRF, TensoRF, DietNeRF, RegNeRF), with per-scene numbers in the supplementary material. No parameter is fitted to reproduce the target PSNR/SSIM/LPIPS values, and no success criterion is defined in terms of the method's own output. The authors cite no prior work of their own; the closest conceptual precursor, FreeNeRF's frequency masking, is cited as external prior work and explicitly distinguished as MLP-specific, with the contribution being its adaptation to grid-based TensoRF. The coarse-to-fine claims (i)-(iii) in Section 4.2 are empirical assumptions tested by experiments, not conclusions forced by construction. The skeptical observations about the non-Hermitian mask in Eq. (2) and the schedule reaching f_t=1 at iteration 1400 under the supplementary's delta=1/2000 are technical-correctness concerns about whether the described mechanism matches the implementation; they do not make the result equivalent to its inputs, so they are outside the circularity definition used here.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper contributes a training procedure rather than a physical theory. Its free parameters are dataset-level hyperparameters, hand-chosen rather than fitted to reproduce metrics. The key assumptions are the coarse-to-fine frequency prior and the transferability of one schedule per dataset. No new physical or mathematical entities are introduced.

free parameters (5)

Initial clipping threshold f0 = 0.3 (Blender synthetic), 0.01 (LLFF real)
Chosen per dataset; controls how many Fourier coefficients are active at the start of training. Main text says 0.01% but supplementary gives 0.01, an inconsistency.
Frequency increment per iteration Delta = 2e-3 (synthetic), 1e-4 (real)
Determines how fast higher frequencies are released; ablations show a peak in PSNR, highlighting sensitivity.
TV loss weight = 1.0
Applied to appearance and density features; not part of the Fourier parameterization but needed for reported results.
L1 density loss weight (real scenes only) = 1e-4
Used on LLFF instead of weight decay to reduce floaters.
Optimizer weight decay = 0.2 (synthetic), 0 (real)
Set to 0.2 for synthetic to remove floaters; authors note this is key.

assumptions (4)

standard math Standard discrete Fourier transform properties: the forward and inverse FFT are exact linear operators for fixed grid sizes.
Used in Eqs. 1 and 3 for feature parameterization.
domain assumption TensoRF's vector-matrix decomposition and volume rendering pipeline is a valid, trainable radiance field representation.
The method is built directly on TensoRF; any failure of TensoRF as a representation would propagate.
domain assumption Low Fourier frequencies of the feature grids correspond to coarse geometry and appearance, and high frequencies to fine detail.
This is the paper's core observation (Section 3 and claims in Section 4.2), justifying the curriculum.
domain assumption A single linear schedule (f0, Delta) transfers across all scenes within each dataset without per-scene tuning.
Reported results use one setting per dataset; no per-scene selection is described.

how reviews work

0 comments

Cite this review

Pith. "Pith review of FourieRF: Few-Shot NeRFs via Progressive Fourier Frequency Control." pith.science (2026). https://pith.science/paper/SR2LWIIG

@misc{pith2026250201405,
  author       = {Pith},
  title        = {Pith review of: FourieRF: Few-Shot NeRFs via Progressive Fourier Frequency Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SR2LWIIG}},
  note         = {Machine review of arXiv:2502.01405}
}

read the original abstract

In this work, we introduce FourieRF, a novel approach for achieving fast and high-quality reconstruction in the few-shot setting. Our method effectively parameterizes features through an explicit curriculum training procedure, incrementally increasing scene complexity during optimization. Experimental results show that the prior induced by our approach is both robust and adaptable across a wide variety of scenes, establishing FourieRF as a strong and versatile baseline for the few-shot rendering problem. While our approach significantly reduces artifacts, it may still lead to reconstruction errors in severely under-constrained scenarios, particularly where view occlusion leaves parts of the shape uncovered. In the future, our method could be enhanced by integrating foundation models to complete missing parts using large data-driven priors.

Figures

Figures reproduced from arXiv: 2502.01405 by the authors.

**Figure 1.** FourieRF serves as an effective and simple baseline for tackling the few-shot rendering problem. The vanilla approach often encounters high-frequency artifacts early in the optimization process. We introduce an explicit curriculum training procedure that gradually incorporates higher frequencies to mitigate this. This method ensures a stable training trajectory, eliminating major artifacts and enhancing overall rend… view at source ↗

**Figure 2.** Method illustration. From left to right. Feature vectors and matrices are initialized in the spatial space. They are projected using the FFT. The Fourier coefficients are clipped using the masking procedure. Finally, the inverse FFT is applied to retrieve the smoothed features. shot rendering using 3D Gaussian splatting [10, 24]; however, these rely on depth information, placing them outside the scope of our compa… view at source ↗

**Figure 3.** Coarse Geometry Extraction. Our method is capable of extracting correct coarse geometry from as little as 3 views. This coarse geometry remains relatively stable regardless of the number of views we input. struggle of all NeRF representations is that they are data hungry, they require several images to perform the NVS task properly. When given limited images, say 3 or 6, these models are extremely prone to overfitti… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Overfitting on the few-shot rendering problem. “Catastrophic overfitting” is a common behavior for standard NeRF representations on the few-shot rendering problem. Degenerate geometry is learned, which might result in plausible views near train inputs but does not gen…

**Figure 5.** Figure 5: Comparison on Blender Dataset. In the Lego scene, trained with 4 views, we compare the performance of FreeNeRF, ZeroRF, and our method. ZeroRF renders a compact and clean reconstruction of the scene, however, at the cost of omitting some key details. FreeNeRF fails in …

**Figure 6.** Figure 6: Comparison in LLFF Dataset. In the horns scene, we evaluated the performance of FreeNeRF, ZeroRF, and our method under a 3-view training setup. ZeroRF struggled to reconstruct coherent geometry, resulting in significant inconsistencies. FreeNeRF, while more stable, pro…

**Figure 7.** Figure 7: Choice of ∆ vs performance. We investigated the effect of varying the speed at which high-frequencies are integrated during training, using the Blender Dataset with 6 views. The baseline performance without our method is highlighted in red, while our best result is s…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 17 canonical work pages

[1]

Barron, Ben Mildenhall, Matthew Tancik, Pe- ter Hedman, Ricardo Martin-Brualla, and Pratul P

Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Pe- ter Hedman, Ricardo Martin-Brualla, and Pratul P. Srini- vasan. Mip-NeRF: A multiscale representation for anti- aliasing neural radiance fields. ICCV, 2021. 2

work page 2021
[2]

MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo

Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international con- ference on computer vision , pages 14124–14133, 2021. 2, 3

work page 2021
[3]

TensoRF: Tensorial radiance fields

Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2, 4, 5, 7, 8, 1

work page 2022
[4]

Putting nerf on a diet: Semantically consistent few-shot view synthesis

Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885–5894, 2021. 2, 3, 7

work page 2021
[5]

NeRFshop: Interactive editing of neural radiance fields

Cl ´ement Jambon, Bernhard Kerbl, Georgios Kopanas, Stavros Diolatzis, Thomas Leimk ¨uhler, and George Dret- takis. NeRFshop: Interactive editing of neural radiance fields. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 6(1), 2023. 2

work page 2023
[6]

TensoIR: Tensorial inverse rendering

Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Song- fang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2

work page 2023
[7]

3d gaussian splatting for real-time radiance field rendering

Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2

2023
[8]

InfoNeRF: Ray entropy minimization for few-shot neural volume ren- dering

Mijeong Kim, Seonguk Seo, and Bohyung Han. InfoNeRF: Ray entropy minimization for few-shot neural volume ren- dering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 12912– 12921, 2022. 2, 3

work page 2022

Show all 25 references

[9]

Adam: A method for stochastic optimization

Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

arXiv
[10]

Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normaliza- tion

Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normaliza- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 2...
[11]

Decoupled weight decay regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1

2017 arXiv
[12]

Local light field fusion: Practical view syn- thesis with prescriptive sampling guidelines

Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view syn- thesis with prescriptive sampling guidelines. ACM Trans. Graph., 38(4):1–14, 2019. 7

2019
[13]

Srinivasan, Matthew Tancik, Jonathan T

Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. In ECCV, 2020. 1, 2, 7

2020
[14]

Instant neural graphics primitives with a multires- olution hash encoding

Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a multires- olution hash encoding. ACM Trans. Graph., 41(4), 2022. 2

2022
[15]

Reg- nerf: Regularizing neural radiance fields for view synthesis from sparse inputs

Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg- nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition,...

2022
[16]

Floaters no more: Radiance field gradient scaling for improved near-camera training

Julien Philip and Valentin Deschaintre. Floaters no more: Radiance field gradient scaling for improved near-camera training. arXiv preprint arXiv:2305.02756, 2023. 1

2023 arXiv
[17]

Flipn- erf: Flipped reflection rays for few-shot novel view synthe- sis

Seunghyeon Seo, Yeonjin Chang, and Nojun Kwak. Flipn- erf: Flipped reflection rays for few-shot novel view synthe- sis. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 22883–22893, 2023. 2, 3

2023
[18]

Zerorf: Fast sparse view 360deg reconstruction with zero pretrain- ing

Ruoxi Shi, Xinyue Wei, Cheng Wang, and Hao Su. Zerorf: Fast sparse view 360deg reconstruction with zero pretrain- ing. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 21114–21124,
[19]

Deep image prior

Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446–9454,
[20]

Sparsenerf: Distilling depth ranking for few-shot novel view synthesis

Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zi- wei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 9065–9076,
[21]

Reconfusion: 3d reconstruction with diffusion priors

Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patt...

2024
[22]

FreeNeRF: Im- proving few-shot neural rendering with free frequency reg- ularization

Jiawei Yang, Marco Pavone, and Yue Wang. FreeNeRF: Im- proving few-shot neural rendering with free frequency reg- ularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254–8263,
[23]

pixelNeRF: Neural radiance fields from one or few images

Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2, 3

2021
[24]

Fsgs: Real-time few-shot view synthesis using gaussian splatting

Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European conference on computer vision, pages 145–163. Springer, 2025. 2, 3 9 FourieRF: Few-Shot NeRFs via Progressive Fourier Frequency Control Supplem...

2025
[25]

and ZeroRF [18]. As can be seen in the provided web- site, our method significantly outperforms those baselines in terms of quality of reconstructions in the few-shot set- ting, and has fewer artefacts, especially when considering very few input views. We provide qualitative c...

2000

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Barron, Ben Mildenhall, Matthew Tancik, Pe- ter Hedman, Ricardo Martin-Brualla, and Pratul P

Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Pe- ter Hedman, Ricardo Martin-Brualla, and Pratul P. Srini- vasan. Mip-NeRF: A multiscale representation for anti- aliasing neural radiance fields. ICCV, 2021. 2

work page 2021

[2] [2]

MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo

Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international con- ference on computer vision , pages 14124–14133, 2021. 2, 3

work page 2021

[3] [3]

TensoRF: Tensorial radiance fields

Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2, 4, 5, 7, 8, 1

work page 2022

[4] [4]

Putting nerf on a diet: Semantically consistent few-shot view synthesis

Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885–5894, 2021. 2, 3, 7

work page 2021

[5] [5]

NeRFshop: Interactive editing of neural radiance fields

Cl ´ement Jambon, Bernhard Kerbl, Georgios Kopanas, Stavros Diolatzis, Thomas Leimk ¨uhler, and George Dret- takis. NeRFshop: Interactive editing of neural radiance fields. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 6(1), 2023. 2

work page 2023

[6] [6]

TensoIR: Tensorial inverse rendering

Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Song- fang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2

work page 2023

[7] [7]

3d gaussian splatting for real-time radiance field rendering

Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2

2023

[8] [8]

InfoNeRF: Ray entropy minimization for few-shot neural volume ren- dering

Mijeong Kim, Seonguk Seo, and Bohyung Han. InfoNeRF: Ray entropy minimization for few-shot neural volume ren- dering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 12912– 12921, 2022. 2, 3

work page 2022