Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Pl\"uckeRF: A Line-based 3D Representation for Few-view Reconstruction

T0 review · 4 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 PlückeRF claims that a 3D representation built from Plücker-coordinate lines, with attention biased by physical line-to-line distance, improves few-view reconstruction over triplane and feed-forward baselines.

desk verdict Clean inductive bias, over-claimed results: the distance bias demonstrably helps, but the SOTA comparison is contradicted by the paper's own Cars numbers. read the letter →

arxiv 2506.03713 v1 pith:UKGJA7AH submitted 2025-06-04 cs.CV

Sam Bahrami , Dylan Campbell This is my paper
classification cs.CV
keywords Plückercoordinatesfew-viewreconstructiontriplanerepresentationattentionbiasnovelviewsynthesisneuralradiancefieldsfeed-forwardgeometricinductive
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 is trying to establish that a feed-forward few-view 3D reconstruction model works better when its 3D representation is made of structured lines, not just feature planes. The proposed PlückeRF representation attaches a Plücker-coordinate ray to every pixel of three orthogonal feature grids, and biases transformer attention by the shortest distance between those 3D lines and the rays of the input images. This gives the model a geometric prior: nearby and intersecting lines share information, distant lines are penalized. On two-view reconstruction of synthetic chairs and cars, the authors report higher PSNR and lower LPIPS than the same model without the bias and than the compared triplane and feed-forward baselines. If the claim is right, it offers a simple, inexpensive way to inject 3D locality into transformer reconstruction without volumetric convolutions or per-scene optimization.

What carries the argument

The load-bearing object is the PlückeRF line representation: three orthogonal axis-aligned feature grids $T_{xy},T_{yz},T_{zx}$, with each grid pixel defining a ray perpendicular to its plane, converted to Plücker coordinates $l=(d,m)=(d,o\times d)$. The identity that carries the argument is the line-to-line distance of Eq. (2): for non-parallel lines $d(l_1,l_2)=|d_1^\top m_2+d_2^\top m_1|/\|d_1\times d_2\|_2$, and for parallel lines $\|d_1\times(m_1-(d_1^\top d_2)m_2)\|_2$. This distance is used to bias attention logits as $\mathrm{softmax}(QK^\top/\sqrt{d_I}-\gamma d(l_q,l_k))V$, in both cross-attention (image rays vs. 3D lines) and self-attention (3D lines vs. each other). The bias turns camera pose information into a concrete geometric prior: nearby and intersecting lines share information, distant lines do not, without requiring a volumetric data structure.

What would settle it

Train the same two-view model with and without the distance bias on an object category or real-capture dataset not in the training distribution. If the bias stops improving novel-view PSNR, or hurts it, the claim that geometric line-to-line distance is a generally useful inductive bias would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that representing the output 3D field as three orthogonal feature grids, each grid pixel carrying a Plücker-coordinate line perpendicular to its plane, turns camera geometry into a usable inductive bias. The model computes line-to-line distances between image-patch rays and those grid lines, and between every pair of grid lines, then subtracts a learnable multiple of the distance from attention logits before softmax. In the paper's experiments, this bias improves two-view novel-view synthesis on ShapeNet-SRN Chairs (PSNR 28.22 vs. 27.67 without bias) and Cars (25.54 vs. 25.26), with larger relative gains on extrapolated views at least 90 degrees from both inputs. The authors conclude that geometric proximity between rays and 3D locations is a useful prior for feed-forward few-view reconstruction.

Load-bearing premise

The paper's experiments cover only two synthetic object categories, so the load-bearing assumption is that the learned line-distance bias will also help on other object classes and on real images; the paper gives no evidence for that transfer.

Editorial extensions

If this is right

  • On the two-view ShapeNet-SRN Chairs and Cars benchmarks, adding the line–line distance bias improves PSNR over the identical model without the bias, so geometric attention is a direct driver of reconstruction quality.
  • The bias's benefit is larger on extrapolated views (out-of-plane rotation at least 90 degrees from both inputs), meaning the representation helps the model infer genuinely unseen sides rather than only interpolate.
  • Ablations show both finetuning the DINOv2 encoder and keeping the learnable bias scale $\gamma$ are important: removing either drops PSNR by roughly 1 to 1.4 dB in the shortened schedule, so the mechanism is not a free lunch.
  • Using Plücker line parameters as transformer input tokens mainly speeds convergence rather than changing final accuracy, suggesting the geometric value comes from the distance bias and positional encoding, not from the token content.

Reading between the lines

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

  • Inference: the same line–line distance bias could be grafted onto other explicit representations such as Gaussian splat primitives or hash grids, giving a testable prediction that geometric attention improves those too.
  • Inference: because the bias penalizes long-range attention, it may suppress useful symmetric correspondences such as the two sides of a chair; a model with orientation- or layer-dependent $\gamma$ might recover that information without losing locality.
  • Inference: the strong dependence on DINOv2 finetuning suggests that part of the gain is the encoder learning dataset-specific appearance, so the purely geometric contribution should be measured against a frozen-encoder version on real data before generalizing the claim.
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

4 major / 5 minor

Summary. The paper proposes PlückeRF, a feed-forward few-view 3D reconstruction model that represents the target object as three orthogonal feature grids and augments each grid cell with Plücker coordinates of the ray through that cell. The model uses a transformer with cross- and self-attention, and biases the attention logits by the line-line distance between image rays and grid rays and between grid rays. The model is trained end-to-end with NeRF-style volumetric rendering losses and is evaluated on ShapeNet-SRN Cars and Chairs against OpenLRM, pixelNeRF, and Splatter Image. The main evidence is that the distance-bias variant consistently outperforms the no-bias variant in the reported settings.

Significance. If the claimed improvements hold, the paper offers a simple, geometrically motivated inductive bias for feed-forward few-view reconstruction, with the benefit of a structured and interpretable 3D representation. The ablation evidence is a real strength: the distance bias improves PSNR from 24.21 to 23.44 in Table 3 and from 28.22 to 27.67 on Chairs in Table 1, and the code is released. However, the broader claim of state-of-the-art performance is not established by the paper's own quantitative results on Cars, and the evaluation is limited to two synthetic object categories. The central representation idea is defensible, but the evidence and claims need to be brought into alignment.

major comments (4)
  1. [Sec. 4.2, Table 1] The abstract and conclusion claim improvements over "state-of-the-art feedforward reconstruction methods," but the paper's own quantitative results on SRN Cars do not support that claim: 2-view Splatter Image achieves PSNR 26.01 while the proposed method achieves 25.54, and the paper does not report Splatter Image's LPIPS on Cars, so no perceptual comparison is available there. The qualitative Cars comparison in Figure 5 (and supplementary Figure 7) uses the single-view Splatter Image (n=1) model rather than the stronger 2-view model used in Table 1, which overstates the qualitative advantage. The claim should be restricted to the included baselines and to the Chairs dataset, or the comparisons should be completed and updated.
  2. [Sec. 4.2, Table 1 and Sec. 4.4, Table 3] No error bars, standard deviations, or multiple-seed results are reported anywhere. The core improvement of the distance bias is 0.55 dB on Chairs (28.22 vs 27.67) and 0.28 dB on Cars (25.54 vs 25.26) in Table 1, and 0.77 dB in the Table 3 ablation (24.21 vs 23.44). Without variance estimates, these differences cannot be distinguished from run-to-run noise, so the central claim that the bias improves reconstruction is not fully supported. Please report at least three seeds with mean and standard deviation, or another statistical comparison.
  3. [Sec. 4.1, Sec. 4.2, Sec. 4.4] The evaluation is restricted to two synthetic object categories (ShapeNet-SRN Chairs and Cars). The paper's title and abstract present the line-based representation and distance bias as a general mechanism for few-view reconstruction, but generalizability to other object classes or real-world scenes is untested. The large drop when DINOv2 finetuning is removed (Table 3: PSNR 24.21 vs 23.10) suggests the model is heavily dependent on the pretrained backbone and the training distribution, so the geometric bias may not transfer. Add at least one cross-category or real-data experiment, or explicitly scope the claims to category-specific synthetic objects.
  4. [Sec. 4.1, Table 1] The "state-of-the-art" comparison set is incomplete and partly dated. OpenLRM is a single-view method included at n=1, pixelNeRF is from 2021, and no recent generalizable feed-forward baselines (e.g., pixelSplat, MVSplat, or other LRM variants) are evaluated. Therefore, even where the proposed method wins on Chairs, it has not been shown to be state of the art. Please either add recent relevant baselines or replace the "state-of-the-art" claim with a claim about the specific compared methods.
minor comments (5)
  1. [Sec. 4.2, Figure 5] The caption notes that Splatter Image results are from the single-view model, while Table 1 compares with the two-view model; the main text should explicitly explain this discrepancy and justify why the qualitative comparison is against a weaker baseline.
  2. [Table 1 and Table 2] Splatter Image's LPIPS on Cars is missing in Table 1, and Table 2 has no Splatter Image results on Cars at all; please report these or state why they are unavailable.
  3. [Sec. 3.1, Eq. (2)] The parallel-line distance formula assumes unit directions so that d1^T d2 equals +1 or -1; this holds for the constructed rays but should be stated explicitly for clarity.
  4. [Throughout] The typesetting of the name "PlückeRF" and the word "Plücker" is inconsistent; a final proofread would improve presentation.
  5. [Sec. 5] The sentence "outperforming existing approaches" is too broad given the Cars numbers and should be aligned with the revised claims.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity: the distance-bias improvement is a controlled held-out ablation, not a fitted prediction.

full rationale

PlückeRF's claimed contribution is the line–line distance-biased attention of Eq. (4) applied to Plücker-coordinate line tokens. The central empirical claim—that this bias improves over the equivalent triplane representation—is supported by a controlled ablation (Ours w/o bias vs Ours, Table 1: Chairs PSNR 27.67→28.22; Cars 25.26→25.54) evaluated on held-out test views. The bias is not fitted to the test set; the learnable scale γ is trained on the training split and the improvement transfers to the disjoint test split. The triplane baseline is the same model without the bias, so the comparison isolates the contribution. The Plücker-coordinate line construction (Eq. 1) and distance formula (Eq. 2) are standard geometry, not defined in terms of the rendered output. The use of pretrained DINOv2 features is a dependency, and the SOTA claim is weakened by Table 1's Cars numbers (Ours 25.54 PSNR vs Splatter Image 2-view 26.01), but that is an evidence/correctness issue rather than circular reasoning. The only self-citation is Flash3D [30] in Related Work, which is not load-bearing for the core derivation. No step in the paper reduces by construction to its inputs.

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

The PlückeRF line representation is a data structure, not a new physical entity in the sense of the schema. The free parameters are the loss weight, grid resolution, and the learnable bias scale gamma, all standard model choices or hyperparameters. The axioms reflect the method's dependence on a pretrained backbone, the specific geometric distance formula, and the triplane rendering pipeline.

free parameters (3)
  • LPIPS loss weight alpha = 0.01 (after warm-up)
    Set to 0 initially and increased to 0.01 at a later training iteration, chosen to balance MSE and perceptual loss.
  • Feature grid resolution M = 64 (main), 48 (ablation)
    Higher resolution improves detail but increases memory; 64 chosen for main experiments, 48 for ablations to save compute.
  • Distance bias scale gamma = learned, initialized at 0
    A learnable scalar per attention layer controlling the strength of the line-line distance bias; fixing it to 1 in the ablation hurts performance.
assumptions (3)
  • domain assumption DINOv2 pretrained features provide a strong image prior that can be finetuned for 3D reconstruction
    The method relies on a pretrained vision transformer for patch features; the ablation shows a large drop without finetuning.
  • ad hoc to paper The line-line distance in Eq. (2) is the correct geometric measure for attention bias
    The paper introduces this specific distance formulation and does not compare with alternatives such as point-to-line distance or Euclidean distance between line endpoints.
  • standard math The triplane rendering formulation of Chan et al. applies to this setting
    The method reuses the triplane NeRF rendering pipeline without modification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pl\"uckeRF: A Line-based 3D Representation for Few-view Reconstruction." pith.science (2026). https://pith.science/paper/UKGJA7AH

@misc{pith2026250603713,
  author       = {Pith},
  title        = {Pith review of: Pl\"uckeRF: A Line-based 3D Representation for Few-view Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UKGJA7AH}},
  note         = {Machine review of arXiv:2506.03713}
}
read the original abstract

Feed-forward 3D reconstruction methods aim to predict the 3D structure of a scene directly from input images, providing a faster alternative to per-scene optimization approaches. Significant progress has been made in single-view and few-view reconstruction using learned priors that infer object shape and appearance, even for unobserved regions. However, there is substantial potential to enhance these methods by better leveraging information from multiple views when available. To address this, we propose a few-view reconstruction model that more effectively harnesses multi-view information. Our approach introduces a simple mechanism that connects the 3D representation with pixel rays from the input views, allowing for preferential sharing of information between nearby 3D locations and between 3D locations and nearby pixel rays. We achieve this by defining the 3D representation as a set of structured, feature-augmented lines; the Pl\"uckeRF representation. Using this representation, we demonstrate improvements in reconstruction quality over the equivalent triplane representation and state-of-the-art feedforward reconstruction methods.

Figures

Figures reproduced from arXiv: 2506.03713 by the authors.

Figure 1
Figure 1. PluckeRF: A feed-forward method that directly predicts a 3D representation from posed input images. The model uses a line–line ¨ distance biased attention mechanism that helps the model interpret which parts of the images are important for constructing specific areas of the 3D output. The resulting 3D representation is structured as three orthogonal feature grids, enabling efficient novel view synthesis. Abstract Fe… view at source ↗
Figure 2
Figure 2. Flowchart for the PluckeRF model. We use a pre-trained vision transformer (DINOv2 [ ¨ 23]) to encode the input images (Sec. 3.6). We take the camera information from the images and create rays corresponding with the image patches from our image encoder, which we convert to Plucker coordinates (Sec. ¨ 3.3). Our final input is the set of lines corresponding to the pixels in the orthogonal feature grid representation, … view at source ↗
Figure 3
Figure 3. A visual representation of the Plucker coordinate dis- ¨ tance between a line from our transformer keys (represented by the red arrow), and the lines corresponding to the queries—the PluckeRF line representation. The shading on the squares on each ¨ axis-aligned plane represents the distance between {lPi } 3N2 i=1 and the red arrow. In practice, those with a greater distance (lighter squares) will be penalized in at… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Shapenet-SRN Chairs qualitative comparison. The [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Shapenet-SRN Cars qualitative comparison. The model [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Additional Shapenet-SRN Chairs qualitative results on chairs in the testing split. Given two [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Additional Shapenet-SRN Cars qualitative results on cars in the testing split. Given two input [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages

  1. [1]

    Chan, Connor Z

    Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 3, 6

    work page 2022

  2. [2]

    pixelSplat: 3D Gaussian Splats from Im- age Pairs for Scalable Generalizable 3D Reconstruction

    David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelSplat: 3D Gaussian Splats from Im- age Pairs for Scalable Generalizable 3D Reconstruction. In CVPR, pages 19457–19467, 2024. 2

    work page 2024

  3. [3]

    TensoRF: Tensorial Radiance Fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In ECCV,

    work page

  4. [4]

    Ray Conditioning: Trading Photo- consistency for Photo-realism in Multi-view Image Genera- tion

    Eric Ming Chen, Sidhanth Holalkere, Ruyu Yan, Kai Zhang, and Abe Davis. Ray Conditioning: Trading Photo- consistency for Photo-realism in Multi-view Image Genera- tion. In ICCV, pages 23242–23251, 2023. 3, 4

    work page 2023

  5. [5]

    Objaverse: A Universe of Annotated 3D Objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A Universe of Annotated 3D Objects. arXiv, 2022. 2

    work page 2022

  6. [6]

    NerfDiff: Single-image View Synthesis with NeRF-guided Distillation from 3D-aware Diffusion

    Jiatao Gu, Alex Trevithick, Kai-En Lin, Josh Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. NerfDiff: Single-image View Synthesis with NeRF-guided Distillation from 3D-aware Diffusion. In ICML, 2023. 2, 3

    work page 2023

  7. [7]

    Openlrm: Open-source large reconstruction models

    Zexin He and Tengfei Wang. Openlrm: Open-source large reconstruction models. https://github.com/ 3DTopia/OpenLRM, 2023. 6, 7, 8, 1

    work page 2023

  8. [8]

    LRM: Large Reconstruction Model for Single Image to 3D

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large Reconstruction Model for Single Image to 3D. In ICLR, 2024. 2, 3, 6, 1

    work page 2024

Show all 43 references
  1. [9]

    Codenerf: Disentan- gled neural radiance fields for object categories

    Wonbong Jang and Lourdes Agapito. Codenerf: Disentan- gled neural radiance fields for object categories. In CVPR, pages 12949–12958, 2021. 2

    2021

  2. [10]

    Few-View Object Reconstruction with Unknown Cat- egories and Camera Poses

    Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-View Object Reconstruction with Unknown Cat- egories and Camera Poses. International Conference on 3D Vision (3DV), 2024. 3

    2024

  3. [11]

    LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias, 2024

    Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias, 2024. 2, 3, 4

    2024

  4. [12]

    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. 1, 2, 8

    2023

  5. [13]

    An im- age is worth 16x16 words: Transformers for image recogni- tion at scale

    Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weis- senborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Syl- vain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An im- age is worth 16x16 words: Transformers for image recogni-...

    2021

  6. [14]

    Ground- ing Image Matching in 3D with MASt3R

    Vincent Leroy, Yohann Cabon, and J´erˆome Revaud. Ground- ing Image Matching in 3D with MASt3R. In ECCV, pages 71–91. Springer, 2024. 3

    2024

  7. [15]

    Instant3d: Fast Text-to-3D with Sparse-view Generation and Large Reconstruction Model

    Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast Text-to-3D with Sparse-view Generation and Large Reconstruction Model. In ICLR, 2024. 2, 3

    2024

  8. [16]

    Neural Sparse V oxel Fields

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural Sparse V oxel Fields. NeurIPS,

  9. [17]

    Zero-1-to-3: Zero-shot One Image to 3D Object

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tok- makov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot One Image to 3D Object. In CVPR, pages 9298– 9309, 2023. 2

    2023

  10. [18]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR, 2019. 1

    2019

  11. [19]

    RealFusion: 360 Reconstruction of Any Object from a Single Image

    Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. RealFusion: 360 Reconstruction of Any Object from a Single Image. In CVPR, 2023. 3

    2023

  12. [20]

    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 Synthesis. In ECCV, 2020. 1, 2

    2020

  13. [21]

    Instant Neural Graphics Primitives with a Mul- tiresolution Hash Encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant Neural Graphics Primitives with a Mul- tiresolution Hash Encoding. ACM Trans. Graph. , 41(4): 102:1–102:15, 2022. 1, 2

    2022

  14. [22]

    AutoRF: Learning 3D Object Radiance Fields from Single View Observations

    Norman M ¨uller, Andrea Simonelli, Lorenzo Porzi, Samuel Rota Bul `o, Matthias Nießner, and Peter Kontschieder. AutoRF: Learning 3D Object Radiance Fields from Single View Observations. In CVPR, 2022. 2

    2022

  15. [23]

    Maxime Oquab, Timoth ´ee Darcet, Theo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Rus- sell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang- Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Ni...

    2023

  16. [24]

    DeepSDF: Learning Continuous Signed Distance Functions for Shape Represen- tation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning Continuous Signed Distance Functions for Shape Represen- tation. In CVPR, 2019. 2

    2019

  17. [25]

    Julius Pl ¨ucker. XVII. On a New Geometry of Space. Philo- sophical Transactions of the Royal Society of London, pages 725–791, 1865. 3

  18. [26]

    Scene Representation Networks: Continuous 3D- Structure-Aware Neural Scene Representations

    Vincent Sitzmann, Michael Zollh ¨ofer, and Gordon Wet- zstein. Scene Representation Networks: Continuous 3D- Structure-Aware Neural Scene Representations. In NeurIPS,

  19. [27]

    Free- man, Joshua B

    Vincent Sitzmann, Semon Rezchikov, William T. Free- man, Joshua B. Tenenbaum, and Fredo Durand. Light Field Networks: Neural Scene Representations with Single- Evaluation Rendering. In NeurIPS, 2021. 3, 4

    2021

  20. [28]

    Splatt3r: Zero-shot Gaussian Splatting from Uncalibrated Image Pairs

    Brandon Smart, Chuanxia Zheng, Iro Laina, and Vic- tor Adrian Prisacariu. Splatt3r: Zero-shot Gaussian Splatting from Uncalibrated Image Pairs. 2024. 2, 3, 8 9

    2024

  21. [29]

    Viewset Diffusion: (0-)Image-Conditioned 3D Generative Models from 2D Data

    Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset Diffusion: (0-)Image-Conditioned 3D Generative Models from 2D Data. ICCV, 2023. 2

    2023

  22. [30]

    Flash3D: Feed-Forward Gen- eralisable 3D Scene Reconstruction from a Single Image

    Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao Henriques, Christian Rup- precht, and Andrea Vedaldi. Flash3D: Feed-Forward Gen- eralisable 3D Scene Reconstruction from a Single Image. arXiv, 2024. 2, 3

    2024

  23. [31]

    Splatter Image: Ultra-Fast Single-View 3D Recon- struction

    Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter Image: Ultra-Fast Single-View 3D Recon- struction. In CVPR, 2024. 2, 3, 6, 7, 8, 1

    2024

  24. [32]

    Geometry-biased Transformers for Novel View Synthesis

    Naveen Venkat, Mayank Agarwal, Maneesh Singh, and Shubham Tulsiani. Geometry-biased Transformers for Novel View Synthesis. arXiv, 2023. 2, 3

    2023

  25. [33]

    VGGSfM: Visual geometry grounded deep structure from motion

    Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. VGGSfM: Visual geometry grounded deep structure from motion. In CVPR, pages 21686–21697, 2024. 3

    2024

  26. [34]

    Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images

    Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images. In ECCV, 2018. 2

    2018

  27. [35]

    Efros, and Angjoo Kanazawa

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3D Per- ception Model with Persistent State, 2025. 3

    2025

  28. [36]

    DUSt3R: Geometric 3D Vision Made Easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D Vision Made Easy. In CVPR, 2024. 3

    2024

  29. [37]

    SynSin: End-to-end view synthesis from a single image

    Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 2

    2020

  30. [38]

    PQ-NET: A Generative Part Seq2Seq Network for 3D Shapes

    Rundi Wu, Yixin Zhuang, Kai Xu, Hao Zhang, and Baoquan Chen. PQ-NET: A Generative Part Seq2Seq Network for 3D Shapes. In CVPR, 2020. 2

    2020

  31. [39]

    Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv, 2024. 2

    2024

  32. [40]

    FvOR: Robust Joint Shape and Pose Optimization for Few-view Object Recon- struction

    Zhenpei Yang, Zhile Ren, Miguel Angel Bautista, Zaiwei Zhang, Qi Shan, and Qixing Huang. FvOR: Robust Joint Shape and Pose Optimization for Few-view Object Recon- struction. In CVPR, pages 2497–2507, 2022. 3

    2022

  33. [41]

    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, 6, 7, 8, 1

    2021

  34. [42]

    GS-LRM: Large Re- construction Model for 3D Gaussian Splatting

    Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large Re- construction Model for 3D Gaussian Splatting. ECCV, 2024. 2, 3, 8

    2024

  35. [43]

    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 CVPR, 2018. 6 10 Pl ¨uckeRF: A Line-based 3D Representation for Few-view Reconstruction Supplementary Material In this appen...

    2018

Pith tools

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