Pith. sign in

REVIEW 2 major objections 6 minor 2 cited by

Intern-GS: Vision Model Guided Sparse-View 3D Gaussian Splatting

T0 review · 2 major / 6 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 Intern-GS shows that replacing SfM with a dense stereo prior and diffusion-refined pseudo views pushes sparse-view 3D Gaussian rendering to state-of-the-art quality.

desk verdict A coherent engineering contribution in sparse-view 3DGS, but the SOTA claim is undercut by a concrete numerical inconsistency between Table 1 and Table 3. read the letter →

arxiv 2505.20729 v1 pith:3FU35QR3 submitted 2025-05-27 cs.CV

Xiangyu Sun , Runnan Chen , Mingming Gong , Dong Xu , Tongliang Liu This is my paper
classification cs.CV
keywords sparse-viewreconstruction3DGaussiansplattingnovelviewsynthesisvisionfoundationmodelsdiffusionpriordepthregularizationmulti-viewstereoinitializationfew-shotrendering
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

The paper sets out to show that the two weak points of sparse-view 3D Gaussian Splatting, thin and unreliable point-cloud initialization and missing supervision for unobserved viewpoints, can both be fixed with pretrained vision models. It builds a dense, scale-aligned, redundancy-reduced initial point cloud from a multi-view stereo network instead of an SfM point cloud. During optimization it adds depth-correlation losses on training views and on nearby pseudo views, plus a photometric loss that compares rendered pseudo views against diffusion-refined images. On three-view LLFF, DTU, and Tanks and Temples benchmarks, the method reports the best or near-best PSNR, LPIPS, and SSIM among the listed baselines, with the largest single ablation gain coming from the dense initialization and the next from the diffusion appearance refiner.

What carries the argument

The machinery is a guided initialization plus a three-term hybrid regularizer. In initialization, a multi-view stereo network predicts pixel-wise point maps and confidence maps, and a Redundancy-Free masking rule adds new Gaussians only where track density is low or where rendered depth is badly in front of the predicted depth. In optimization, depth regularization uses Pearson correlation between rendered and predicted depth, which removes scale ambiguity, on both training views and pseudo views offset by five degrees of rotation; appearance refinement renders pseudo views, passes them through a diffusion refiner conditioned on semantic and pixel-level image latents from pretrained vision encoders, and feeds the refined images into a photometric loss (Eq. 17). The Pearson correlation is the scale-ambiguity workhorse, and the pseudo-view photometric loss is the mechanism that extends supervision into unseen regions.

What would settle it

Choose a held-out test viewpoint far from all three training viewpoints, render it with Intern-GS, and measure PSNR and LPIPS separately on image regions whose 3D surfaces are occluded from all training views. If those occlusion-only regions show no advantage over a no-diffusion variant, or if they show hallucinated texture, then the diffusion pseudo-view loss is filling gaps with invented content rather than recovering the scene's actual appearance.

Watch

Extended reading notes

Core claim

The paper's central claim is that sparse-view 3D Gaussian Splatting fails not because the splatting optimization is weak, but because it starts from too little geometry and is supervised only at observed viewpoints. Intern-GS replaces SfM initialization with a dense point cloud produced by a pretrained multi-view stereo network, removes redundant Gaussians with a masking rule based on track density and depth error, and then regularizes optimization with scale-invariant depth-correlation losses on both training views and nearby pseudo views. A pretrained diffusion model redraws rendered pseudo-view images, and those redrawn images become photometric ground truth for regions the true views never see. The overall claim is that consistent geometry and appearance in unobserved regions can be recovered from pretrained priors, and doing so yields state-of-the-art rendering quality with only three training views.

Load-bearing premise

The diffusion model's synthetic pseudo-view images are treated as ground truth for unseen regions, so if those images contain plausible but invented details, the Gaussians are optimized toward the generator's preconceptions rather than the true scene, and the reported gains partly measure agreement with the prior instead of reconstruction accuracy.

Editorial extensions

If this is right

  • If the claim holds, SfM initialization is no longer a prerequisite for high-quality sparse-view 3DGS; a learned stereo prior can seed the Gaussians directly.
  • Training-view and pseudo-view depth regularization are compatible and each contributes independent gains, with the training-view depth constraint being the larger depth contributor.
  • The pipeline transfers beyond forward-facing captures to large-scale scenes, where it still reports top PSNR and LPIPS with three views.
  • Because the diffusion refiner is frozen, the method gains rendering quality without per-scene generative training, and each scene trains within minutes.

Reading between the lines

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

  • The largest ablation jump (PSNR 19.64 to 20.49) comes from the diffusion appearance refiner, which suggests the reported superiority over NeRF-style baselines may partly reflect the generative prior's agreement with the test distribution rather than pure reconstruction of unseen geometry; a fair test is to evaluate on scenes whose appearance is far from the diffusion model's training distribution.
  • The same recipe, dense multi-view stereo initialization plus pseudo-view diffusion supervision, should transfer to other explicit representations such as surface-based splatting, because nothing in the argument depends on Gaussians specifically.
  • A stronger variant of the claim would predict that pseudo-view supervision can be replaced by a pure consistency regularizer, for example rendering the same pseudo view from two nearby camera positions and enforcing agreement, which would remove the hallucination risk while keeping the geometric benefit; the paper does not test this.
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

2 major / 6 minor

Summary. The paper proposes Intern-GS, a sparse-view 3D Gaussian Splatting method that replaces SfM-based initialization with a DUSt3R-derived dense, redundancy-free point cloud, and adds two regularization terms during optimization: depth constraints on training and pseudo views, and a diffusion-based Multi-view Appearance Refinement that generates pseudo-view images used as color supervision. Experiments on LLFF, DTU, and Tanks and Temples under three training views report state-of-the-art PSNR/LPIPS/SSIM, and ablations decompose the contributions of the three components.

Significance. If the reported numbers are reliable, the paper offers a useful recipe for sparse-view 3DGS: dense multi-view stereo initialization plus external depth and diffusion priors can mitigate overfitting, with consistent gains across forward-facing and large-scale scenes. The ablations are informative and the method is clearly described. The main evidence, however, is weakened by an inconsistency between Tables 1 and 3 for the same configuration, and the diffusion-based pseudo-view supervision carries a hallucination-bias risk that is not yet directly validated. No code release is indicated, which also limits reproducibility.

major comments (2)
  1. [Tables 1 and 3 (Sections 4.2 and 4.3)] Table 1 reports Intern-GS on LLFF with three training views as PSNR 20.49, LPIPS 0.212, SSIM 0.693, while Table 3, fourth row, reports the same method, dataset, and training views as PSNR 20.49, LPIPS 0.304, SSIM 0.656. Because PSNR is identical but LPIPS and SSIM differ, at least one table is computed under a different protocol or contains a transcription error. This discrepancy is load-bearing: the headline state-of-the-art claim on LLFF rests on the Table 1 values. If Table 3 is correct, Intern-GS's SSIM (0.656) is below SparseGS's 0.668 and its LPIPS advantage over FreeNeRF is marginal, so the claim of state-of-the-art on all metrics is not supported. Please reconcile the tables and state which protocol corresponds to the final model.
  2. [Section 3.4, Eq. (17), and Table 3] The pseudo-view color loss L_cp in Eq. (17) treats diffusion-refined images as photometric ground truth for unobserved viewpoints. Because those images are produced by a pretrained diffusion model, the Gaussians can be optimized toward the prior's hallucinated appearance rather than the true scene; the ablation attributes the largest single PSNR jump (19.64 to 20.49 in Table 3) to MAR, so this concern is load-bearing. I do not view the use of the diffusion prior as circular, since test-view metrics are external to the method, but the hallucination-bias risk needs a concrete test. Please provide evidence that the pseudo-view images are consistent with the observed views, for example by measuring the discrepancy between diffusion-refined pseudo views and the nearest observed view, or by validating on a subset where held-out ground truth is available.
minor comments (6)
  1. [Tables 1 and 2] SparseGS is cited as "Zhu et al. (2024)" in the tables but as Fu et al. (2024b) in the text and reference list; since FSGS is also Zhu et al. (2024), this citation ambiguity makes it difficult to identify which baseline results are being compared.
  2. [Section 2 headings] The heading "Novel View Synthsis" in Sections 2.1 and 2.2 contains a typo; it should read "Novel View Synthesis."
  3. [Eq. (9)] The mask condition "L1(Dp) > 50MDE" is unclear: MDE is mentioned in the text as median depth error, but the equation does not define the L1 operation or the units, so the threshold behavior is ambiguous.
  4. [Section 4.1 and Appendix B.1] The text says "downsampling rates of 8 and 4 for LLFF and DTU," but Appendix B.1 describes LLFF downsampled by a factor of 8 and DTU resized from 1600x1200 to 400x300; the latter is not a uniform downsampling rate of 4 in both dimensions, so the description should be made consistent.
  5. [Section 3.4, Eq. (14)] The UNet U_theta is described as "learnable," but the diffusion model appears to be a pretrained prior; if it is frozen during Gaussian optimization, the wording should be corrected to avoid implying fine-tuning.
  6. [Appendix D] The limitation section appropriately acknowledges difficulties with outward scene extrapolation, but it does not address the risk that diffusion-refined pseudo views may introduce hallucinated appearance in unseen regions; this should be discussed explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: external pretrained priors and held-out test-view evaluation keep the derivation self-contained; the flagged metric inconsistency is a correctness issue, not a circularity reduction.

full rationale

Intern-GS's derivation chain is not circular under the defined rubric. The initialization uses DUSt3R, an externally published pretrained model, to produce dense point clouds; the depth losses (Eqs. 11-12) compare rendered depth against depth from DUSt3R and MiDaS, both external priors; and the appearance refinement (Eqs. 13-17) uses a pretrained diffusion model conditioned on rendered images and CLIP/DINO latents. No parameter is fitted to the held-out test views, and the final SOTA claim is evaluated on external train/test splits (LLFF, DTU, Tanks and Temples). The pseudo-view targets in L_cp are generated from the model's own renderings through a fixed pretrained diffusion refiner, so they are self-conditioned training regularizers rather than independent measurements; this is a potential hallucination or bias risk, but it does not make the central claim true by construction, and the held-out evaluation is not defined in terms of those pseudo-labels. The grid-searched weights (lambda_3, lambda_4) are hyperparameters, not predictions. There are no load-bearing self-citations, no imported uniqueness theorem, and no ansatz smuggled in via citation. Separately, the paper's evidence has a serious internal inconsistency: Table 1 reports the full LLFF model at LPIPS 0.212 / SSIM 0.693, while Table 3 reports the same configuration at LPIPS 0.304 / SSIM 0.656 with identical PSNR 20.49; this is an evidence/correctness problem to be resolved by the correctness pass, but it is not a circularity reduction because no equation or fitted parameter forces the SOTA claim by definition.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The method depends on several pretrained models and hand-set thresholds rather than deriving new mathematical results. The free parameters are mostly loss weights and mask thresholds chosen by hand or grid search, and the 'axioms' are the pretrained vision models and the 3DGS framework that are taken as given. No new physical or conceptual entity is introduced.

free parameters (5)
  • lambda_3 (pseudo-depth loss weight) = 0.05
    Selected via grid search (Section 3.5) to balance depth regularization from pseudo views.
  • lambda_4 (pseudo-color loss weight) = 0.001
    Selected via grid search (Section 3.5) for the diffusion-refined pseudo-view color loss.
  • pseudo-view rotation deviation = 5 degrees
    Pseudo views are generated with a fixed 5-degree rotation perturbation from training views (Section 3.3); no sensitivity analysis.
  • track density threshold in mask = 0.5
    Mask adds Gaussians where track density S_p < 0.5 (Eq. 9); threshold is hand-set.
  • depth error threshold in mask = 50 x median depth error
    Mask adds Gaussians when depth error exceeds 50 times the median depth error (Eq. 9); hand-set.
assumptions (6)
  • domain assumption 3D Gaussian Splatting rendering and optimization equations (Eqs. 1-5) are a valid scene representation and remain the base model.
    The method builds directly on the 3DGS framework (Kerbl et al.) without modifying the core representation.
  • domain assumption DUSt3R provides scale-aligned, accurate point maps and camera poses for sparse input images.
    Section 3.2 relies on DUSt3R's pretrained weights for dense initialization and poses.
  • domain assumption MiDaS provides reliable monocular depth for pseudo views.
    Pseudo-view depth regularization uses MiDaS depth as the target (Section B.3).
  • ad hoc to paper The diffusion refiner (Wang et al. 2024) produces photometrically consistent images that can serve as color supervision.
    The pseudo-view color loss (Eq. 17) assumes diffusion-refined images are valid ground truth for unseen viewpoints.
  • domain assumption Pearson correlation is an appropriate scale-invariant depth loss.
    Used to handle scale ambiguity between rendered and predicted depth (Eqs. 11-12).
  • ad hoc to paper Pseudo views 5 degrees from training views are close enough to be useful yet different enough to reduce overfitting.
    The 5-degree deviation is fixed without ablation (Section 3.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Intern-GS: Vision Model Guided Sparse-View 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/3FU35QR3

@misc{pith2026250520729,
  author       = {Pith},
  title        = {Pith review of: Intern-GS: Vision Model Guided Sparse-View 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3FU35QR3}},
  note         = {Machine review of arXiv:2505.20729}
}
read the original abstract

Sparse-view scene reconstruction often faces significant challenges due to the constraints imposed by limited observational data. These limitations result in incomplete information, leading to suboptimal reconstructions using existing methodologies. To address this, we present Intern-GS, a novel approach that effectively leverages rich prior knowledge from vision foundation models to enhance the process of sparse-view Gaussian Splatting, thereby enabling high-quality scene reconstruction. Specifically, Intern-GS utilizes vision foundation models to guide both the initialization and the optimization process of 3D Gaussian splatting, effectively addressing the limitations of sparse inputs. In the initialization process, our method employs DUSt3R to generate a dense and non-redundant gaussian point cloud. This approach significantly alleviates the limitations encountered by traditional structure-from-motion (SfM) methods, which often struggle under sparse-view constraints. During the optimization process, vision foundation models predict depth and appearance for unobserved views, refining the 3D Gaussians to compensate for missing information in unseen regions. Extensive experiments demonstrate that Intern-GS achieves state-of-the-art rendering quality across diverse datasets, including both forward-facing and large-scale scenes, such as LLFF, DTU, and Tanks and Temples.

Figures

Figures reproduced from arXiv: 2505.20729 by the authors.

Figure 1
Figure 1. Comparison of the SOTA SparseNeRF Wang et al. (2023a), SparseGS Fu et al. (2024b) in 3 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of point cloud initialization of original 3D Gaussian Kerbl et al. (2023) and our method [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. In our framework, we first utilize a multi-view stereo to predict point maps. This technique [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Results on LLFF dataset Mildenhall et al. (2019) and DTU dataset Aanæs et al. (2016) in 3 training [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Results on Tanks dataset Knapitsch et al. (2017) in 3 training views. In comparison, SparseNeRF [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The qualitative results of Intern-GS on Tanks and Temples dataset under 3 training views. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: The qualitative results of Intern-GS on DTU dataset under 3 training views. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: The qualitative results of Intern-GS on DTU dataset under 3 training views. [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: The qualitative results of Intern-GS on LLFF dataset under 3 training views. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MAC-Splat: Multi-Attribute Consistency for High-Fidelity Sparse-View Reconstruction

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Semantically enriched MASt3R correspondences plus a multi-attribute 3D consistency loss raise sparse-view ScanNet++ PSNR by >4.5 dB over Splatt3R and preserve quality under wide baselines.

  2. Sparse-View 3D Reconstruction: Recent Advances and Open Challenges

    cs.CV 2025-07 conditional novelty 3.0 of 10

    A comprehensive survey that organizes sparse-view 3D reconstruction methods into geometry-based, NeRF, 3DGS, and diffusion-based categories, with benchmarks and open challenges.

Reference graph

Works this paper leans on

58 extracted references · 38 canonical work pages · cited by 2 Pith papers

  1. [1]

    Large-scale data for multiple-view stereopsis

    Henrik Aan s, Rasmus Ramsb l Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, pp.\ 1--16, 2016

    work page 2016

  2. [2]

    Novel view synthesis in tensor space

    Shai Avidan and Amnon Shashua. Novel view synthesis in tensor space. In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp.\ 1034--1040. IEEE, 1997

    work page 1997

  3. [3]

    Mip-nerf: A multiscale representation for anti-aliasing neural 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 neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 5855--5864, 2021

    work page 2021

  4. [4]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 16123--16133, 2022

    work page 2022

  5. [5]

    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 conference on computer vision, pp.\ 14124--14133, 2021

    work page 2021

  6. [6]

    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, pp.\ 333--350. Springer, 2022

    work page 2022

  7. [7]

    Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes

    Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 7911--7920, 2021

    work page 2021

  8. [8]

    Depth-supervised nerf: Fewer views and faster training for free

    Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12882--12891, 2022

    work page 2022

Show all 58 references
  1. [9]

    Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds

    Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2, 2024

    2024 arXiv

  2. [10]

    Colmap-free 3d gaussian splatting

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 20796--20805, 2024 a

    2024

  3. [11]

    Colmap-free 3d gaussian splatting

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 20796--20805, 2024 b

    2024

  4. [12]

    Nerf: Neural radiance field in 3d vision, a comprehensive review

    Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022

    2022 arXiv

  5. [13]

    Fastnerf: High-fidelity neural rendering at 200fps

    Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 14346--14355, 2021

    2021

  6. [14]

    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, pp.\ 5885--5894, 2021

    2021

  7. [15]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35: 0 26565--26577, 2022

    2022

  8. [16]

    Splatam: Splat, track & map 3d gaussians for dense rgb-d slam

    Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126, 2023

    2023 arXiv

  9. [17]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk \"u hler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 0 (4): 0 1--14, 2023

    2023

  10. [18]

    Tanks and temples: Benchmarking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 0 (4): 0 1--13, 2017

    2017

  11. [19]

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

    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 normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 20775--...

    2024

  12. [20]

    Neural sparse voxel fields

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33: 0 15651--15663, 2020

    2020

  13. [21]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 10012--10022, 2021

    2021

  14. [22]

    A review of 3d reconstruction techniques in civil engineering and their applications

    Zhiliang Ma and Shilong Liu. A review of 3d reconstruction techniques in civil engineering and their applications. Advanced Engineering Informatics, 37: 0 163--174, 2018

    2018

  15. [23]

    Gaussian splatting slam

    Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. arXiv preprint arXiv:2312.06741, 2023

    2023 arXiv

  16. [24]

    Gnerf: Gan-based neural radiance field without posed camera

    Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 6351--6361, 2021

    2021

  17. [25]

    Local light field fusion: Practical view synthesis 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 synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38 0 (4): 0 1--14, 2019

    2019

  18. [26]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65 0 (1): 0 99--106, 2021

    2021

  19. [27]

    Giraffe: Representing scenes as compositional generative neural feature fields

    Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11453--11464, 2021

    2021

  20. [28]

    Regnerf: 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. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp....

    2022

  21. [29]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp...

    2021

  22. [30]

    Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer

    Ren \'e Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44 0 (3): 0 1623--1637, 2020

    2020

  23. [31]

    Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps

    Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 14335--14345, 2021

    2021

  24. [32]

    Scaling vision with sparse mixture of experts

    Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, Andr \'e Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. Advances in Neural Information Processing Systems, 34: 0 8583--8595, 2021

    2021

  25. [33]

    Nerf-slam: Real-time dense monocular slam with neural radiance fields

    Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp.\ 3437--3444. IEEE, 2023

    2023

  26. [34]

    Learning representations by back-propagating errors

    David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. nature, 323 0 (6088): 0 533--536, 1986

    1986

  27. [35]

    Structure-from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 4104--4113, 2016

    2016

  28. [36]

    Pixelwise view selection for unstructured multi-view stereo

    Johannes Lutz Sch\" o nberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016

    2016

  29. [37]

    Graf: Generative radiance fields for 3d-aware image synthesis

    Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33: 0 20154--20166, 2020

    2020

  30. [38]

    imap: Implicit mapping and positioning in real-time

    Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 6229--6238, 2021

    2021

  31. [39]

    Dreamgaussian: Generative gaussian splatting for efficient 3d content creation

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023

    2023 arXiv

  32. [40]

    Sparf: Neural radiance fields from sparse and noisy poses

    Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 4190--4200, 2023

    2023

  33. [41]

    The interpretation of structure from motion

    Shimon Ullman. The interpretation of structure from motion. Proceedings of the Royal Society of London. Series B. Biological Sciences, 203 0 (1153): 0 405--426, 1979

    1979

  34. [42]

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

    Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 9065--9076, 2023 a

    2023

  35. [43]

    Vistadream: Sampling multiview consistent images for single-view scene reconstruction

    Haiping Wang, Yuan Liu, Ziwei Liu, Wenping Wang, Zhen Dong, and Bisheng Yang. Vistadream: Sampling multiview consistent images for single-view scene reconstruction. arXiv preprint arXiv:2410.16892, 2024

    2024 arXiv

  36. [44]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. arXiv preprint arXiv:2312.14132, 2023 b

    2023 arXiv

  37. [45]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13 0 (4): 0 600--612, 2004

    2004

  38. [46]

    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

  39. [47]

    Nerf-ds: Neural radiance fields for dynamic specular objects

    Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8285--8295, 2023

    2023

  40. [48]

    Freenerf: Improving few-shot neural rendering with free frequency regularization

    Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8254--8263, 2023

    2023

  41. [49]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 5752--5761, 2021 a

    2021

  42. [50]

    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 Vision and Pattern Recognition, pp.\ 4578--4587, 2021 b

    2021

  43. [51]

    Cor-gs: sparse-view 3d gaussian splatting via co-regularization

    Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: sparse-view 3d gaussian splatting via co-regularization. In European Conference on Computer Vision, pp.\ 335--352. Springer, 2024

    2024

  44. [52]

    Nerf++: Analyzing and improving neural radiance fields

    Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020

    2010 arXiv

  45. [53]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 586--595, 2018

    2018

  46. [54]

    View synthesis by appearance flow

    Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision--ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11--14, 2016, Proceedings, Part IV 14, pp.\ 286--301. Springer, 2016

    2016

  47. [55]

    Deep review and analysis of recent nerfs

    Fang Zhu, Shuai Guo, Li Song, Ke Xu, Jiayu Hu, et al. Deep review and analysis of recent nerfs. APSIPA Transactions on Signal and Information Processing, 12 0 (1), 2023

    2023

  48. [56]

    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, pp.\ 145--163. Springer, 2024

    2024

  49. [57]

    Nice-slam: Neural implicit scalable encoding for slam

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12786--12796, 2022

    2022

  50. [58]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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