Pith. sign in

REVIEW 1 major objections 5 minor 48 references

Depth-Guided Bundle Sampling for Efficient Generalizable Neural Radiance Field Reconstruction

T0 review · 1 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that grouping adjacent rays into bundles and sampling each bundle as a cone of depth-guided spheres makes generalizable neural radiance fields up to twice as fast to render, sometimes with better image quality.

desk verdict A promising acceleration trick for generalizable NeRFs that is undermined as written by a max/min typo in the sample-count formula; the core idea is worth refereeing after that fix. read the letter →

arxiv 2505.19793 v1 pith:ESVZQRUS submitted 2025-05-26 cs.CV

Li Fang , Hao Zhu , Longlong Chen , Fei Hu , Long Ye , Zhan Ma This is my paper
classification cs.CV
keywords generalizableneuralradiancefieldsbundlesamplingdepth-guidedadaptivenovelviewsynthesisvolumerenderingaccelerationplenoptictheorymipmapfeaturesmulti-viewstereo
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

Generalizable neural radiance fields can already synthesize novel views from a few images, but they pay a high price: every pixel along every ray is sampled densely, so high-resolution rendering is slow. This paper claims most of that work is redundant, because natural scenes are piecewise smooth. It groups adjacent rays into $K\times K$ bundles, casts a cone through each bundle, and samples a small set of inscribed spheres shared by the whole bundle, using depth-guided adaptive counts so that smooth regions get few samples and complex regions get more. Plugged into ENeRF, the method reports up to a 1.27 dB PSNR gain and a 47% to 2.24x speedup, with average samples per ray dropping to 0.42 ($2\times 2$) and 0.104 ($4\times 4$) on DTU. If this holds, it makes real-time, high-resolution view synthesis from sparse inputs considerably closer.

What carries the argument

Bundle sampling partitions the target image into non-overlapping $K\times K$ pixel bundles; each bundle is a cone from the camera through the corresponding image disk, and the cone is sampled by a sequence of inscribed spheres with centers and radii given by Eq. (3). Feature extraction uses mipmap levels from the source-view feature volumes selected by the sphere's projected footprint radius (Eq. (5)), producing a joint bundle representation, while pixel-aligned colors for each individual ray inside the bundle are concatenated as a ray-specific representation that restores high-frequency detail. Depth-guided adaptive sampling sets the number of spheres per cone to $N_C = \max(\lceil 2R/\delta_s \rceil, N_{\max})$ (Eq. (6)), with $R$ the predicted depth range and $\delta_s$ the minimum inter-sample spacing, so smooth regions can receive as few as one sample. This machinery converts the plenoptic-sampling observation into a concrete rendering-speedup mechanism.

What would settle it

Render a scene dominated by thin structures and high-frequency texture (fences, hair, grass, specular highlights) using the method at $2\times 2$ versus $4\times 4$ bundle sizes and compare against per-ray sampling: if the larger bundle collapses PSNR and LPIPS on these regions far more than on smooth scenes, or if the depth-guided sampler visibly under-samples at depth discontinuities, the piecewise-smooth assumption is the weak point.

Watch

Extended reading notes

Core claim

The central claim is that you do not need to sample every ray in a generalizable radiance field: rays that land in the same $K\times K$ patch of the target image can be treated as one bundle, represented by a cone whose inscribed spheres carry pre-filtered mipmap features from the source views, with one shared volume-rendering pass. The number of spheres per cone is set adaptively from the predicted depth range, so smooth, confident regions receive very few samples while boundaries and depth discontinuities receive more. The paper shows that this bundle representation, combined with a per-ray color residual, actually improves rendering quality over the ENeRF base while slashing the sample count, and that the same recipe transfers to a 3D-GS base model with smaller but positive gains.

Load-bearing premise

The load-bearing premise is that all $K\times K$ rays within a bundle see essentially the same scene content, so one shared cone of spheres with depth-guided sampling can represent them without losing details; if a bundle straddles a thin structure, a sharp edge, or strong parallax, the shared representation blurs detail and the depth-confidence-based sampler may put too few samples in the wrong place.

Editorial extensions

If this is right

  • At bundle size $2\times 2$, the method raises PSNR on DTU by 1.27 dB over the ENeRF base while increasing FPS by 47%, so shared bundle sampling can improve quality while cutting cost.
  • At bundle size $4\times 4$, the same model runs at over 2x the FPS of the base (43.6 vs 19.5) with comparable PSNR, offering a tunable speed/quality trade-off.
  • The depth-guided adaptive sampler concentrates samples at object boundaries and depth discontinuities and reduces them in smooth areas, which is how the average sample count per ray drops below 0.1 without visible degradation.
  • The method transfers to a generalizable 3D-GS pipeline, giving a small quality improvement and faster inference, indicating the bundle idea is not specific to volume-density decoders.
  • Cross-dataset results on Real Forward-facing and NeRF Synthetic keep quality on par with or above the base method, so the speed gains do not obviously come from overfitting to the DTU training set.

Reading between the lines

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

  • The piecewise-smooth assumption predicts the largest speed-ups will appear on scenes with large low-texture regions; scenes packed with fine detail will need smaller bundles or a higher $N_{\max}$, and a per-bundle measure of texture energy could make that trade-off explicit.
  • The same cone-of-spheres sampling could be applied at training time or to scene-specific radiance fields, where the depth range $R$ is known from geometry, potentially removing the need for a learned depth-confidence branch.
  • A direct comparison against a per-ray baseline with the same total sample budget would isolate whether the quality gain comes from bundle regularization or from depth-guided allocation.
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.

Referee Report

1 major / 5 minor

Summary. The paper proposes a depth-guided bundle sampling strategy for generalizable neural radiance field rendering. Adjacent K×K rays in the target view are grouped into cones; each cone is sampled with inscribed spheres, encoded with mipmap features from source views and per-ray colors, and decoded by a small neural renderer. The number of samples per bundle is intended to be adapted to a predicted depth range R, with an upper cap N_max=6. The method is applied to ENeRF and to MVSGaussian, and is evaluated on DTU, Real Forward-facing, and NeRF Synthetic. The reported claims are up to 1.27 dB PSNR improvement over ENeRF and up to a 2.24× FPS increase on DTU, with average samples per ray of 0.422 for 2×2 bundles and 0.104 for 4×4 bundles.

Significance. The efficiency-oriented idea of sampling bundles of rays jointly with depth-guided adaptive counts is timely and, if implemented as described, would be a useful contribution to generalizable view synthesis. The paper has concrete strengths: it provides ablations isolating the sphere-based bundle sampling, the adaptive depth-guided count, the joint bundle representation, and the ray-specific representation; it evaluates across multiple datasets; the gains are compared against standard benchmarks; and the code is promised publicly. However, the central efficiency claim is currently not reproducible from the equations, because Eq. (6) uses max where the text requires a minimum cap. This is a load-bearing internal inconsistency, not a matter of disagreement with external consensus, and it must be resolved before the results can be accepted.

major comments (1)
  1. [Section 4.2, Eq. (6)] The printed formula N_C = max(ceil(2·R/δ_s), N_max), with N_max = 6 in Section 5.1, forces every bundle to contain at least six samples, because the max operator returns N_max whenever the depth term is smaller. This directly contradicts the sentence that calls N_max a 'maximum limit' and makes the reported average samples per ray in Table 1 impossible: the lower bounds are 6/4 = 1.5 for 2×2 bundles and 6/16 = 0.375 for 4×4 bundles, while Table 1 reports 0.422 and 0.104. If the implementation actually uses min(ceil(2·R/δ_s), N_max), as the surrounding text and the sample-count measurements imply, then Eq. (6), the implementation description, and the reported efficiency numbers must be reconciled before the central speed claim can be reproduced from the manuscript.
minor comments (5)
  1. [Abstract and Table 1] The claim of 'state-of-the-art rendering quality' is not supported by Table 1 on DTU, where ConvGLR reports a PSNR of 31.65 compared with 28.86 for ENeRF+Ours (2×2); the wording should be qualified, for example as state-of-the-art among methods at interactive rates.
  2. [Abstract and Table 1] The abstract reports a PSNR improvement of up to 1.27 dB over ENeRF, but the values in Table 1 give 28.86−27.61 = 1.25 dB; the discrepancy should be corrected or explained.
  3. [Table 3] In the fourth ablation column, the LPIPS value is printed as 0.933, which is implausible relative to all other LPIPS values in the table (roughly 0.07–0.11); this appears to be a typo, likely 0.093, and should be corrected.
  4. [Section 4.1.1] The statement that grouping rays into K×K bundles 'reduces the total number of samples by approximately 1/K^2' is only true at a fixed number of samples per bundle; since the later adaptive scheme changes the per-bundle count, the sentence should be clarified to avoid an implied guarantee that is not what the method delivers.
  5. [Section 4.1.1, Eq. (4)] The typesetting of the expression for r_src appears malformed in the manuscript, with unclear placement of the square root and parentheses; please check and re-set the equation so the projection radius calculation is unambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the efficiency and quality claims are validated against external benchmarks and do not reduce to fitted parameters or self-citations.

full rationale

The central claims (bundle sampling reduces samples; depth-guided adaptive sampling improves speed; ENeRF+Ours outperforms ENeRF on DTU) are evaluated on external datasets (DTU, Real Forward-facing, NeRF Synthetic) against published baselines using standard PSNR/SSIM/LPIPS and FPS metrics. The adaptive sample count N_C in Eq. (6) is a function of the predicted depth range R and hand-set constants (delta_s, N_max); it is not fitted to the target PSNR or FPS, and the reported sample averages are produced by the depth estimator's confidence rather than by construction from the metric. The cone/mipmap sampling is borrowed from external prior work [15] and integrated into a new bundle formulation, so no load-bearing claim is justified only by a self-citation (the authors do not cite themselves). The only notable manuscript issue is that Eq. (6) reads max(...) while the surrounding text describes N_max as a maximum cap; taken literally this would force at least N_max samples per bundle and is inconsistent with the reported 0.422/0.104 average samples per ray. That is a reproducibility/consistency concern, not a circularity: it does not make any predicted quantity equivalent to an input by construction. Score 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method introduces no new physical entities. It adds two hand-set hyperparameters (delta_s and N_max) and relies on three domain assumptions about scene smoothness, plenoptic bounds, and depth reliability. These are the inputs the reader must grant for the central claim to hold.

free parameters (2)
  • delta_s (minimum inter-sample spacing) = 1/64 of scene depth range
    Used in Eq. (6) to determine the sample count per bundle; hand-chosen, not derived from first principles.
  • N_max (maximum samples per bundle) = 6
    Caps the adaptive sample count in Eq. (6); set experimentally and not fit to the evaluation metrics.
assumptions (3)
  • domain assumption Plenoptic sampling bound (Eq. 2) transfers to NeRF bundle sampling.
    The paper uses the bound to justify adaptive sample counts but does not prove its validity in this setting.
  • domain assumption Natural scenes are piecewise smooth, with high-frequency content localized at boundaries.
    This justifies grouping KxK adjacent rays into a single cone; not true for all scenes (e.g., dense foliage or thin structures).
  • domain assumption The predicted depth range R from the feature volume is reliable enough for adaptive sampling.
    Sample allocation depends critically on the depth estimate; errors could misallocate samples and degrade quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Depth-Guided Bundle Sampling for Efficient Generalizable Neural Radiance Field Reconstruction." pith.science (2026). https://pith.science/paper/ESVZQRUS

@misc{pith2026250519793,
  author       = {Pith},
  title        = {Pith review of: Depth-Guided Bundle Sampling for Efficient Generalizable Neural Radiance Field Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ESVZQRUS}},
  note         = {Machine review of arXiv:2505.19793}
}
read the original abstract

Recent advancements in generalizable novel view synthesis have achieved impressive quality through interpolation between nearby views. However, rendering high-resolution images remains computationally intensive due to the need for dense sampling of all rays. Recognizing that natural scenes are typically piecewise smooth and sampling all rays is often redundant, we propose a novel depth-guided bundle sampling strategy to accelerate rendering. By grouping adjacent rays into a bundle and sampling them collectively, a shared representation is generated for decoding all rays within the bundle. To further optimize efficiency, our adaptive sampling strategy dynamically allocates samples based on depth confidence, concentrating more samples in complex regions while reducing them in smoother areas. When applied to ENeRF, our method achieves up to a 1.27 dB PSNR improvement and a 47% increase in FPS on the DTU dataset. Extensive experiments on synthetic and real-world datasets demonstrate state-of-the-art rendering quality and up to 2x faster rendering compared to existing generalizable methods. Code is available at https://github.com/KLMAV-CUC/GDB-NeRF.

Figures

Figures reproduced from arXiv: 2505.19793 by the authors.

Figure 1
Figure 1. Rendering Quality (PSNR) vs. Speed (FPS) on the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Network architecture of the proposed depth-guided bundle sampling strategy on ENeRF [ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison of ray sampling strategies: (a) Traditional [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of ENeRF+Ours (2 × 2) with state-of-the-art methods [10, 22, 24] under 3-view setting. Each image triplet includes: the reconstructed image on the left, a zoomed-in view on the upper right, and the error map on the lower right. Sphere ✗ ✓ ✓ ✓ ✓ S…
Figure 5
Figure 5. Figure 5: Visualization of sample allocation in depth-guided adap [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 32 canonical work pages

  1. [1]

    Large-scale data for multiple-view stereopsis.International Journal of Computer Vision, 120:153–168, 2016

    Henrik Aanæs, Rasmus Ramsbøl Jensen, George V ogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis.International Journal of Computer Vision, 120:153–168, 2016. 1, 6, 7

    work page 2016

  2. [2]

    Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 5855–5864,

  3. [3]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 5

    2022

  4. [4]

    Real-time neural light field on mobile devices

    Junli Cao, Huan Wang, Pavlo Chemerys, Vladislav Shakhrai, Ju Hu, Yun Fu, Denys Makoviichuk, Sergey Tulyakov, and Jian Ren. Real-time neural light field on mobile devices. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8328–8337, 2023. 2

    work page 2023

  5. [5]

    Plenoptic sampling

    Jin-Xiang Chai, Xin Tong, Shing-Chow Chan, and Heung- Yeung Shum. Plenoptic sampling. InProceedings of the 27th annual conference on Computer graphics and interac- tive techniques, pages 307–318, 2000. 2, 3

    work page 2000

  6. [6]

    Efficient geometry-aware 3d generative adversarial networks

    Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123–16133, 2022. 3

    2022

  7. [7]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction

    David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457–19467, 2024. 2

    2024

  8. [8]

    Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo

    Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo. InProceedings of the IEEE/CVF international conference on computer vision, pages 14124–14133, 2021. 1, 2, 6, 7

    work page 2021

Show all 48 references
  1. [9]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InEuropean con- ference on computer vision, pages 333–350. Springer, 2022. 2

    2022

  2. [10]

    Explicit correspondence matching for generalizable neural radiance fields.arXiv preprint arXiv:2304.12294, 2023

    Yuedong Chen, Haofei Xu, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Explicit correspondence matching for generalizable neural radiance fields.arXiv preprint arXiv:2304.12294, 2023. 1, 2, 6, 7, 8

    2023 arXiv

  3. [11]

    Mobilenerf: Exploiting the polygon ras- terization pipeline for efficient neural field rendering on mo- bile architectures

    Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and An- drea Tagliasacchi. Mobilenerf: Exploiting the polygon ras- terization pipeline for efficient neural field rendering on mo- bile architectures. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recogn...

    2023

  4. [12]

    Stereo radiance fields (srf): Learning view syn- thesis for sparse views of novel scenes

    Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view syn- thesis for sparse views of novel scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7911–7920, 2021. 1, 2

    2021

  5. [13]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 1, 2

    2022

  6. [14]

    Efficientnerf efficient neural radiance fields

    Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12902–12911, 2022. 1, 2

    2022

  7. [15]

    Tri-miprf: Tri-mip represen- tation for efficient anti-aliasing neural radiance fields

    Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip represen- tation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2023. 4, 5

    2023

  8. [16]

    Geonerf: Generalizing nerf with geometry priors

    Mohammad Mahdi Johari, Yann Lepoittevin, and Franc ¸ois Fleuret. Geonerf: Generalizing nerf with geometry priors. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365–18375, 2022. 1, 2

    2022

  9. [17]

    Relu fields: The little non-linearity that could

    Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. InACM SIGGRAPH 2022 conference proceedings, pages 1–9, 2022. 2

    2022

  10. [18]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

  11. [19]

    Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014. 6

    2014 arXiv

  12. [20]

    Adanerf: Adaptive sam- pling for real-time rendering of neural radiance fields

    Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollh¨ofer, and Markus Steinberger. Adanerf: Adaptive sam- pling for real-time rendering of neural radiance fields. In European Conference on Computer Vision, pages 254–270. Springer, 2022. 2

    2022

  13. [21]

    Mine: Towards continuous depth mpi with nerf for novel view synthesis

    Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. Mine: Towards continuous depth mpi with nerf for novel view synthesis. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 12578–12588, 2021. 1, 2

    2021

  14. [22]

    Efficient neural radiance fields for interactive free-viewpoint video

    Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. InSIGGRAPH Asia 2022 Conference Papers, pages 1–9, 2022. 1, 2, 3, 4, 5, 6, 7, 8

    2022

  15. [23]

    Neural sparse voxel fields.Advances in Neural Information Processing Systems, 33:15651–15663,

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields.Advances in Neural Information Processing Systems, 33:15651–15663,

  16. [24]

    Mvsgaussian: Fast generalizable gaussian splatting recon- struction from multi-view stereo

    Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splatting recon- struction from multi-view stereo. InEuropean Conference on Computer Vision, pages 37–53. Springer, 2024. 1, 2...

    2024

  17. [25]

    Geometry-aware reconstruc- tion and fusion-refined rendering for generalizable neural ra- diance fields

    Tianqi Liu, Xinyi Ye, Min Shi, Zihao Huang, Zhiyu Pan, Zhan Peng, and Zhiguo Cao. Geometry-aware reconstruc- tion and fusion-refined rendering for generalizable neural ra- diance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages...

    2024

  18. [26]

    Neural rays for occlusion-aware image-based render- ing

    Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based render- ing. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 7824–7833,

  19. [27]

    Local light field fusion: Practical view syn- thesis with prescriptive sampling guidelines.ACM Transac- tions on Graphics (ToG), 38(4):1–14, 2019

    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 Transac- tions on Graphics (ToG), 38(4):1–14, 2019. 6, 7

    2019

  20. [28]

    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. InECCV, 2020. 1, 6, 7

    2020

  21. [29]

    Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM transactions on graphics (TOG), 41(4):1–15, 2022

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM transactions on graphics (TOG), 41(4):1–15, 2022. 1, 2

    2022

  22. [30]

    Donerf: Towards real- time rendering of compact neural radiance fields using depth oracle networks

    Thomas Neff, Pascal Stadlbauer, Mathias Parger, Andreas Kurz, Joerg H Mueller, Chakravarty R Alla Chaitanya, Anton Kaplanyan, and Markus Steinberger. Donerf: Towards real- time rendering of compact neural radiance fields using depth oracle networks. InComputer Graphics Forum, ...

  23. [31]

    Cascaded and generalizable neural ra- diance fields for fast view synthesis.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    Phong Nguyen-Ha, Lam Huynh, Esa Rahtu, Jiri Matas, Janne Heikkil, et al. Cascaded and generalizable neural ra- diance fields for fast view synthesis.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2, 6, 7

    2023

  24. [32]

    Sampling for view synthesis: From local light field fusion to neural radiance fields and beyond.arXiv preprint arXiv:2408.04586, 2024

    Ravi Ramamoorthi. Sampling for view synthesis: From local light field fusion to neural radiance fields and beyond.arXiv preprint arXiv:2408.04586, 2024. 3

    2024 arXiv

  25. [33]

    Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction

    Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459– 5469, 2022. 2

    2022

  26. [34]

    Recent advances in im- plicit representation-based 3d shape generation.Visual Intel- ligence, 2(1):9, 2024

    Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in im- plicit representation-based 3d shape generation.Visual Intel- ligence, 2(1):9, 2024. 1

    2024

  27. [35]

    Splatter image: Ultra-fast single-view 3d recon- struction

    Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d recon- struction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208– 10217, 2024. 2

    2024

  28. [36]

    Global latent neu- ral rendering

    Thomas Tanay and Matteo Maggioni. Global latent neu- ral rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19723– 19733, 2024. 1, 2, 6, 7

    2024

  29. [37]

    Grf: Learning a general radi- ance field for 3d representation and rendering

    Alex Trevithick and Bo Yang. Grf: Learning a general radi- ance field for 3d representation and rendering. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 15182–15192, 2021. 1, 2

    2021

  30. [38]

    Learning neural duplex ra- diance fields for real-time view synthesis

    Ziyu Wan, Christian Richardt, Alja ˇz Bo ˇziˇc, Chao Li, Vijay Rengarajan, Seonghyeon Nam, Xiaoyu Xiang, Tuotuo Li, Bo Zhu, Rakesh Ranjan, et al. Learning neural duplex ra- diance fields for real-time view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision ...

    2023

  31. [39]

    Is attention all that nerf needs?Proceedings of the International Conference on Learning Representations (ICLR), 2023

    Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venu- gopalan, Zhangyang Wang, et al. Is attention all that nerf needs?Proceedings of the International Conference on Learning Representations (ICLR), 2023. 1, 2, 6, 7

    2023

  32. [40]

    Ibr- net: Learning multi-view image-based rendering

    Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibr- net: Learning multi-view image-based rendering. InPro- ceedings of the IEEE/CVF conference on computer vision and p...

    2021

  33. [41]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 6

    2004

  34. [42]

    Adaptive shells for efficient neu- ral radiance field rendering.ACM Transactions on Graphics (TOG), 42(6):1–15, 2023

    Zian Wang, Tianchang Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas M¨uller, and Zan Gojcic. Adaptive shells for efficient neu- ral radiance field rendering.ACM Transactions on Graphics (TOG), 42(6):1–15, 2023. 1

    2023

  35. [43]

    Murf: Multi-baseline radiance fields

    Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041–20050, 2024. 1, 6, 7

    2024

  36. [44]

    Plenoctrees for real-time rendering of neural radiance fields

    Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752– 5761, 2021. 2

    2021

  37. [45]

    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 Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4578–4587, 2021. 1, 2, 6, 7

    2021

  38. [46]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 6

    2018

  39. [47]

    Gps- gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis

    Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps- gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

    2024

  40. [59]

    Wiley Online Library, 2021. 1, 2

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.