Pith. sign in

REVIEW 5 major objections 5 minor 90 references

H3R: Hybrid Multi-view Correspondence for Generalizable 3D Reconstruction

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A hybrid of epipolar latent volumes and camera-aware attention sets new state-of-the-art results for generalizable 3D reconstruction while training twice as fast.

desk verdict Solid hybrid 3DGS paper with a real backbone finding, but the multi-view scaling claim rests on an uncontrolled MVSplat comparison. read the letter →

arxiv 2508.03118 v1 pith:VWRPIJMD submitted 2025-08-05 cs.CV

classification cs.CV
keywords generalizable3DreconstructionGaussiansplattingmulti-viewcorrespondenceplane-sweeplatentvolumePlückercoordinatesspatial-alignedvisualencodersnovelviewsynthesissparse-view
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 claims that the long-standing trade-off in generalizable 3D reconstruction — explicit geometric methods are precise but fragile, while implicit attention methods are robust but slow — can be broken by fusing both into one network. H3R combines a volumetric latent stage, which warps per-view features onto depth planes using epipolar geometry, with a camera-aware Transformer that refines correspondences through Plücker ray coordinates. The paper reports state-of-the-art PSNR on RealEstate10K, ACID, and DTU, with 2× faster convergence than prior methods, and finds that backbone choice decides quality: spatial-aligned encoders such as SD-VAE clearly beat semantic-aligned ones such as DINOv2. If the claims hold, the practical recipe for generalizable 3D reconstruction is to enforce geometry explicitly in a latent volume, keep the encoder's pixel alignment, and let attention clean up ambiguous regions.

What carries the argument

The load-bearing object is the latent volume built by plane-sweep epipolar warping of a frozen spatial-aligned encoder's features. Equation (3) warps each neighbor view's latent $x_j$ onto $d$ inverse-depth planes through a differentiable homography; Eq. (4) concatenates the warped latents along the depth axis; and the cost-free variant (Eq. 7) keeps the raw warped features rather than computing a correlation or difference cost, letting later layers learn task-specific matching. The complementary mechanism is the camera-aware Transformer, where Plücker ray coordinates (Eq. 9) are embedded into the visual tokens before multi-view self-attention (Eq. 10), injecting per-pixel camera geometry into every attention layer. These two modules carry the argument: removing the latent volume costs about 4 dB on RealEstate10K, and removing the Transformer costs 1.07 dB (Tab. 5).

What would settle it

Take the trained H3R model and degrade test-time camera poses on RealEstate10K by adding controlled rotation and translation noise (for example 0.1°–2° and 0.5–5 cm); if PSNR drops faster for H3R than for an attention-only baseline at the same noise levels, the epipolar-warped latent volume is the fragile link. A second check: replace SD-VAE with a random spatial autoencoder of similar size and see whether the volume term still contributes its roughly 4 dB gain — if not, the spatial-aligned representation claim is conditional on this specific pretraining.

Watch

Extended reading notes

Core claim

The central claim is that explicit and implicit correspondence modeling are not competing alternatives but complementary stages of one pipeline. H3R first builds a discretized latent volume by warping frozen SD-VAE features from context views onto uniformly sampled inverse-depth planes via differentiable homography, concatenating the warped features across depth without any handcrafted similarity cost (the 'cost-free' strategy, which the paper finds generalizes better across datasets than correlation or difference costs). A camera-aware Transformer then aggregates these volume-augmented tokens, with Plücker coordinates of each pixel ray added as positional embeddings, so attention is conditioned on geometry rather than learned purely from appearance. The refined latents are decoded by a hierarchical CNN into pixel-aligned 3D Gaussians, with ray distances computed as a softmax-weighted sum over 128 depth hypotheses. The paper argues that the explicit volume enforces geometric consistency where photometric cues are reliable, while attention repairs ambiguous regions, and that this division of labor yields both the reported +0.59/+1.06/+0.22 dB PSNR gains over prior state of the art and the 2× training speedup.

Load-bearing premise

The pipeline assumes that frozen SD-VAE latent features remain pixel-aligned under plane-sweep homography warping and that the input SfM camera poses are accurate enough for those warps to land correctly — if either fails, the latent volume injects false correspondences that attention cannot fully correct.

Editorial extensions

If this is right

  • Generalizable 3D reconstruction encoders should be chosen for spatial alignment, not semantics: the paper shows SD-VAE beats DINOv2 by roughly 0.8 dB on average across its three test settings while being far smaller (34M vs 86M parameters).
  • Handcrafted matching costs are unnecessary inside the volume: the cost-free concatenation (Eq. 7) matches or beats correlation and difference costs on the training domains and improves cross-dataset DTU results, suggesting learned matching generalizes better than fixed metrics.
  • Training time for this model class can be halved: H3R reaches TranSplat-level quality at 150k training steps where TranSplat needs 300k (Tab. 5), changing the compute budget required to build such systems.
  • Additional input views scale positively instead of negatively: H3R-α gains +2.78 PSNR going from 2 to 8 views while MVSplat loses 6.01 dB, so real deployments can feed arbitrary view counts without retraining.
  • Target-view camera poses can be turned into geometry without images: H3R-α supplies zero visual features but valid Plücker coordinates for target views, gaining +0.19 PSNR on RealEstate10K and improving coverage of unobserved regions.

Reading between the lines

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

  • The SD-VAE finding suggests a broader principle the paper states only for reconstruction: for dense geometric tasks such as depth, optical flow, or matching, a backbone trained with pixel-reconstruction supervision may beat a much larger semantic backbone; rerunning the Tab. A2 sweep on those tasks would test it directly.
  • Because the frozen latent space is assumed homography-warpable, a cheaper encoder route may exist: any autoencoder with strictly pixel-aligned latents (for example, VQ-GAN variants) might substitute for SD-VAE, and the cost-free volume design makes the encoding's warp-compatibility the real constraint rather than the matching metric.
  • The pose-normalization result (Tab. A5, -0.39 dB on DTU) exposes sensitivity to coordinate transforms that the paper does not pursue; an untested deployment risk is robustness to SfM pose noise in real captures, which a perturbation study could quantify.
  • The largest gains appear at low overlap ([0.60, 0.65), +1.73 dB over MVSplat), hinting the hybrid also suits wide-baseline matching; evaluating on large-baseline stereo benchmarks where epipolar priors are weakest is a natural extension.
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

5 major / 5 minor

Summary. The paper proposes H3R, a feed-forward generalizable 3D Gaussian Splatting method that combines a plane-sweep latent volume built from frozen SD-VAE features with a camera-aware Transformer using Plücker coordinates. Two extensions are introduced: H3R-α for variable 2-8 input views and target-pose integration, and H3R-β for 512×512 resolution inputs. The authors report state-of-the-art performance claims on RealEstate10K, ACID, and DTU, along with cross-dataset generalization, an ablation of visual encoders/cost strategies, and a 2× faster convergence claim based on a 150k-step comparison. The manuscript includes extensive ablations and a public code release.

Significance. If the claims hold, this is a solid systems contribution: the hybrid explicit/implicit correspondence design is simple and effective, the encoder comparison is a useful empirical guide for the community, the cross-dataset and overlap-robustness evaluations are thorough, and the code release supports reproducibility. However, several headline claims are not supported as stated. The abstract's PSNR gains use different variants and baselines across datasets; the base two-view model is not state-of-the-art on ACID; the multi-view scalability comparison is not matched in training protocol; and the 2× convergence claim needs clarification about whether the 150k row is an independent run or an intermediate checkpoint. These issues are fixable but currently prevent acceptance.

major comments (5)
  1. [Abstract; Tables 1–2] The three headline PSNR gains are not computed against a consistent baseline or variant. The abstract reports +0.59 dB on RealEstate10K, +1.06 dB on ACID, and +0.22 dB on DTU. In the tables, 0.59 dB is the difference between H3R-β (512×512) and DepthSplat† (which uses additional pre-training data) on RealEstate10K (Tab. 1); 1.06 dB matches H3R-β vs MVSplat on zero-shot ACID (Tab. 2, 29.23 vs 28.15); 0.22 dB matches base H3R vs TranSplat on DTU (Tab. 2, 15.15 vs 14.93). Because the variant, input resolution, and baseline differ by dataset, the summary statistics do not support the stated improvements as a single comparison. Please re-report gains for one protocol, e.g., base H3R vs the strongest same-resolution baseline on each dataset, and keep the abstract and tables consistent.
  2. [§4.1, Tab. 1] The base H3R is not state-of-the-art on ACID. H3R achieves 28.29 PSNR, while eFreeSplat reports 28.30 and TranSplat 28.35; H3R is also worse in LPIPS than eFreeSplat (0.144 vs 0.140). The 'state-of-the-art across multiple benchmarks' claim is therefore only true for H3R-α (28.44) and H3R-β (28.71), which use additional views/target poses or 512×512 inputs. Specify which variant and input configuration the SOTA claim refers to, and do not describe the base two-view model as SOTA on ACID.
  3. [§4.1, Fig. 4 and Tab. A7] The multi-view scalability comparison is not matched. H3R-α was fine-tuned with 2–8 context views and target-pose integration, while MVSplat appears to be evaluated out-of-the-box at 4/6/8 views without fine-tuning; MVSplat's degradation (26.36→20.35 PSNR) is expected for a two-view-trained model receiving unseen view counts. To support the 'robust multi-view adaptability' contribution, compare against a baseline fine-tuned on multi-view inputs under matched compute/data, or report MVSplat's own multi-view training results. As written, the +9.89 dB gap at 8 views reflects training-protocol differences rather than architectural ones.
  4. [§4.2, Tab. 5 and Appendix C.3] The '2× faster convergence' claim is not sufficiently supported. The table compares H3R at 150k steps with baselines at 300k steps, but the main H3R model is pre-trained for 1M steps (Appendix C.3), and the 150k 'Full' row reaches only 26.68 PSNR in this table while the Tab. 1 model reaches 27.60. State whether the 150k row is a from-scratch run or an intermediate checkpoint, and report wall-clock time per step. If the 150k row is a checkpoint, the '2×' claim is about reaching TranSplat-level quality, not about the final model, and this distinction must be made explicit.
  5. [§4.1, Tab. 1] The table is headed 'Two-view Novel View Synthesis' but includes H3R-α with 5×256×256 inputs and target-view cameras. This row is not comparable to the two-view baselines. Report an H3R-α variant with exactly two context views and no target poses, and isolate the marginal effect of additional views and of target-pose integration, so the reader can separate the architectural contribution from the input-extension contribution.
minor comments (5)
  1. [Eqs. (4)–(8)] The correlation-based volume in Eq. (5) has shape h×w×d, while the difference and cost-free volumes in Eqs. (6)–(7) have shape h×w×d×c; Eq. (8) then writes Linear2: R^{d×c}→R^{c′} for all cases. Please clarify the projection dimensions for the correlation strategy.
  2. [Tab. 1 footnote] The symbols 'H3R-�' and 'H3R-�' render as unicode boxes; replace them with H3R-α and H3R-β throughout the table and text.
  3. [Appendix A.3, Tab. A5] The sentence 'We attribute this discrepancy stems from different camera motion characteristics' is ungrammatical; revise to 'We attribute this discrepancy to different camera motion characteristics.'
  4. [Appendix C.2] The model details report '12 layers with hidden dimensions of 512' but the decoder and volume dimensions are only specified in Appendix B; a short forward-path summary with tensor shapes at each stage would aid reproducibility.
  5. [Sec. 3.2, Eq. (13)] The averaging operation over j≠i should specify the tensor dimensions over which the mean is taken; for the cost-free volume this is not obvious from the notation x′_i = (1/(N-1)) Σ x′_{j→i}.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: H3R's claims are empirical, externally benchmarked, and do not reduce by construction to their inputs.

full rationale

This is an empirical systems paper. The central claims—state-of-the-art PSNR/SSIM/LPIPS on RealEstate10K, ACID, and DTU; 2x faster convergence; robustness to variable views; and the superiority of spatial-aligned encoders—are established by held-out evaluations against external benchmarks and by ablation studies, not by an analytical derivation from fitted constants. The method's components (latent volume, Plücker-coordinate Transformer, cost-free volume, SD-VAE encoder) are justified by controlled comparisons in Tables 3, 4, 5 and A2. No parameter is fitted to the test set and then renamed a prediction. The design choices are not derived from the results they are used to explain. The few self-citations (GaussianBlock, GS2-GNESF, View-consistent 3D editing) appear in related-work context and are not load-bearing for the paper's contributions. The concern that the H3R-alpha versus MVSplat multi-view comparison may be an unmatched training-protocol comparison is a fairness and correctness issue, not a circularity issue: the paper does not define H3R's performance in terms of MVSplat's outputs. No uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via citation. The derivation chain, such as it is, is architectural and empirical rather than deductive, so there is no step in which an output is equivalent to an input by construction.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central empirical claims rest on network hyperparameters and benchmark assumptions rather than on mathematical derivation. The most fragile premise is that frozen SD-VAE features can be warped along epipolar lines to form a meaningful latent volume; this is validated only by the paper's own ablations. There are no invented physical entities.

free parameters (7)
  • Depth hypothesis count d=128 = 128
    Uniformly sampled depth planes in inverse depth space (Appendix B). The volumetric fusion result depends on this discretization.
  • Near/far depth planes = 1 to 100 (RealEstate10K/ACID); 2.215 to 4.525 (DTU)
    Dataset-specific bounds that define the epipolar volume and directly affect reconstructed geometry.
  • Pixel-space scale range [s_min, s_max] = 0.5 to 15.0
    Manually set sigmoid bounds for Gaussian scales (Appendix B), affecting the rendering footprint.
  • LPIPS loss weight lambda = 0.05
    Balances the reconstruction loss in Eq. (12); chosen empirically.
  • Pixel gradient loss weight = 1.0
    Set empirically (Appendix C.3); the ablation shows it contributes a small quality gain.
  • EMA decay = 0.999
    Adopted as default because it improves final metrics by 0.18 PSNR (Tab. A4).
  • Max frame distance schedule = 25 to 45 over 150k steps
    Training schedule for view baselines (Appendix C.3); affects the difficulty of correspondence learning.
assumptions (5)
  • standard math Plane-sweep homography warping with calibrated cameras produces valid epipolar correspondences
    Eq. (3) assumes pinhole camera geometry and accurate relative projections P_i_to_j.
  • domain assumption 3D Gaussian Splatting is a differentiable and sufficient scene representation for generalizable view synthesis
    The pipeline renders Gaussians parameterized as in Appendix B and supervises with image losses; the whole approach depends on this representation.
  • ad hoc to paper Frozen SD-VAE features preserve spatial alignment suitable for pixel-wise feature warping
    The main encoder choice (Sec. 4.2, Tab. A2) is justified empirically within this paper rather than by an independent theory; this is a load-bearing premise.
  • domain assumption Benchmark camera poses from SfM are accurate enough for epipolar warping
    The volumetric module and target-view Gaussian generation assume reliable poses; the paper does not quantify pose noise.
  • domain assumption Zero-shot evaluation from indoor RealEstate10K to outdoor ACID and object-centric DTU is a valid generalization test
    The setup follows MVSplat and pixelSplat (Sec. C.1); the generalization conclusions depend on this protocol.

how reviews work

0 comments
Cite this review

Pith. "Pith review of H3R: Hybrid Multi-view Correspondence for Generalizable 3D Reconstruction." pith.science (2026). https://pith.science/paper/VWRPIJMD

@misc{pith2026250803118,
  author       = {Pith},
  title        = {Pith review of: H3R: Hybrid Multi-view Correspondence for Generalizable 3D Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VWRPIJMD}},
  note         = {Machine review of arXiv:2508.03118}
}
abstract

Despite recent advances in feed-forward 3D Gaussian Splatting, generalizable 3D reconstruction remains challenging, particularly in multi-view correspondence modeling. Existing approaches face a fundamental trade-off: explicit methods achieve geometric precision but struggle with ambiguous regions, while implicit methods provide robustness but suffer from slow convergence. We present H3R, a hybrid framework that addresses this limitation by integrating volumetric latent fusion with attention-based feature aggregation. Our framework consists of two complementary components: an efficient latent volume that enforces geometric consistency through epipolar constraints, and a camera-aware Transformer that leverages Pl\"ucker coordinates for adaptive correspondence refinement. By integrating both paradigms, our approach enhances generalization while converging 2$\times$ faster than existing methods. Furthermore, we show that spatial-aligned foundation models (e.g., SD-VAE) substantially outperform semantic-aligned models (e.g., DINOv2), resolving the mismatch between semantic representations and spatial reconstruction requirements. Our method supports variable-number and high-resolution input views while demonstrating robust cross-dataset generalization. Extensive experiments show that our method achieves state-of-the-art performance across multiple benchmarks, with significant PSNR improvements of 0.59 dB, 1.06 dB, and 0.22 dB on the RealEstate10K, ACID, and DTU datasets, respectively. Code is available at https://github.com/JiaHeng-DLUT/H3R.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

90 extracted references · 73 canonical work pages

  1. [1]

    Mvsformer++: Revealing the devil in transformer’s details for multi-view stereo

    Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer++: Revealing the devil in transformer’s details for multi-view stereo. In ICLR, 2024. 2

  2. [2]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In ICCV, 2021. 2

  3. [3]

    Generative novel view synthesis with 3d-aware diffusion models

    Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In ICCV, 2023. 1

  4. [4]

    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. In CVPR,

  5. [5]

    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. In ICCV, 2021. 1

  6. [6]

    Lara: Efficient large-baseline radiance fields

    Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In ECCV, 2024. 2

  7. [7]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, 2024. 1, 2, 4, 5, 6, 7, 8, 14, 15, 16

  8. [8]

    Mvsplat360: Feed-forward 360 scene synthesis from sparse views

    Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In NeurIPS, 2024. 1, 5

Show all 90 references
  1. [9]

    A space-sweep approach to true multi- image matching

    Robert T Collins. A space-sweep approach to true multi- image matching. In CVPR, 1996. 2, 3

  2. [10]

    Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts

    Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, and Zhangyang Wang. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In ICCV,

  3. [11]

    Objaverse-xl: A universe of 10m+ 3d objects

    Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram V oleti, Samir Yitzhak Gadre, et al. Objaverse-xl: A universe of 10m+ 3d objects. In NeurIPS, 2023. 2

  4. [12]

    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. In CVPR, 2023. 2

  5. [13]

    Transmvsnet: Global context-aware multi-view stereo network with trans- formers

    Yikang Ding, Wentao Yuan, Qingtian Zhu, Haotian Zhang, Xiangyue Liu, Yuanjiang Wang, and Xiao Liu. Transmvsnet: Global context-aware multi-view stereo network with trans- formers. In CVPR, 2022. 2

  6. [14]

    Learning to render novel views from wide-baseline stereo pairs

    Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitz- mann. Learning to render novel views from wide-baseline stereo pairs. In CVPR, 2023. 1, 3, 5

  7. [15]

    Pixelgaussian: Gen- eralizable 3d gaussian reconstruction from arbitrary views

    Xin Fei, Wenzhao Zheng, Yueqi Duan, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Jiwen Lu. Pixelgaussian: Gen- eralizable 3d gaussian reconstruction from arbitrary views. arXiv:2410.18979, 2024. 1

  8. [16]

    Cascade cost volume for high-resolution multi-view stereo and stereo matching

    Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In CVPR, 2020. 2

  9. [17]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR,

  10. [18]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 3, 6, 12

  11. [19]

    Query-key normalization for transformers

    Alex Henry, Prudhvi Raj Dachapally, Shubham Pawar, and Yuxuan Chen. Query-key normalization for transformers. In EMNLP, 2020. 16

  12. [20]

    Training compute-optimal large language mod- els

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language mod- els. In NeurIPS, 2022. 13

  13. [21]

    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, 2023. 1, 2, 3, 4

  14. [22]

    Dpsnet: End-to-end deep plane sweep stereo

    Sunghoon Im, Hae-Gon Jeon, Stephen Lin, and In So Kweon. Dpsnet: End-to-end deep plane sweep stereo. In ICLR, 2019. 3

  15. [23]

    Sparsplat: Fast multi-view reconstruction with generalizable 2d gaussian splatting

    Shubhendu Jena, Shishir Reddy Vutukur, and Adnane Boukhayma. Sparsplat: Fast multi-view reconstruction with generalizable 2d gaussian splatting. In CVPR Workshops,

  16. [24]

    Large scale multi-view stereopsis evalu- ation

    Rasmus Jensen, Anders Dahl, George V ogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evalu- ation. In CVPR, 2014. 16

  17. [25]

    Leap: Liberate sparse-view 3d modeling from camera poses

    Hanwen Jiang, Zhenyu Jiang, Yue Zhao, and Qixing Huang. Leap: Liberate sparse-view 3d modeling from camera poses. In ICLR, 2023. 1

  18. [26]

    Gaussianblock: Building part-aware compositional and editable 3d scene by primitives and gaus- sians

    Shuyi Jiang, Qihao Zhao, Hossein Rahmani, De Wen Soh, Jun Liu, and Na Zhao. Gaussianblock: Building part-aware compositional and editable 3d scene by primitives and gaus- sians. In ICLR, 2025. 2

  19. [27]

    Lvsm: A large view synthesis model with minimal 3d inductive bias

    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. In ICLR, 2024. 1, 3 9

  20. [28]

    Geonerf: Generalizing nerf with geometry priors

    Mohammad Mahdi Johari, Yann Lepoittevin, and Franc ¸ois Fleuret. Geonerf: Generalizing nerf with geometry priors. In CVPR, 2022. 1, 3

  21. [29]

    Scaling laws for neural language models

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv:2001.08361, 2020. 13

  22. [30]

    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 TOG, 2023. 1

  23. [31]

    Berg, Wan-Yen Lo, et al

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chlo´e Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, et al. Segment any- thing. In ICCV, 2023. 3, 12

  24. [32]

    Wildgaussians: 3d gaussian splatting in the wild

    Jonas Kulhanek, Songyou Peng, Zuzana Kukelova, Marc Pollefeys, and Torsten Sattler. Wildgaussians: 3d gaussian splatting in the wild. In NeurIPS, 2024. 2

  25. [33]

    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, 2025. 3, 6, 12, 14

  26. [34]

    Gp- nerf: Generalized perception nerf for context-aware 3d scene understanding

    Hao Li, Dingwen Zhang, Yalun Dai, Nian Liu, Lechao Cheng, Jingfeng Li, Jingdong Wang, and Junwei Han. Gp- nerf: Generalized perception nerf for context-aware 3d scene understanding. In CVPR, 2024. 1

  27. [35]

    Infinite nature: Perpetual view generation of natural scenes from a single image

    Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In ICCV, 2021. 5, 16

  28. [36]

    Se- mantic ray: Learning a generalizable semantic field with cross-reprojection attention

    Fangfu Liu, Chubin Zhang, Yu Zheng, and Yueqi Duan. Se- mantic ray: Learning a generalizable semantic field with cross-reprojection attention. In CVPR, 2023. 1

  29. [37]

    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 ICCV, 2023. 4

  30. [38]

    When epipolar constraint meets non-local operators in multi-view stereo

    Tianqi Liu, Xinyi Ye, Weiyue Zhao, Zhiyu Pan, Min Shi, and Zhiguo Cao. When epipolar constraint meets non-local operators in multi-view stereo. In ICCV, 2023. 1, 3

  31. [39]

    Mvsgaussian: Fast generalizable gaussian splatting re- construction 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 re- construction from multi-view stereo. In ECCV, 2024. 1

  32. [40]

    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. In CVPR, 2022. 1

  33. [41]

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

    Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view syn- thesis with prescriptive sampling guidelines. ACM TOG ,

  34. [42]

    Epipolar-free 3d gaussian splatting for generalizable novel view synthesis

    Zhiyuan Min, Yawei Luo, Jianwen Sun, and Yi Yang. Epipolar-free 3d gaussian splatting for generalizable novel view synthesis. In NeurIPS, 2024. 1, 2, 5, 8

  35. [43]

    V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2023. 2, 6, 12, 14

  36. [44]

    Rethinking depth estimation for multi- view stereo: A unified representation

    Rui Peng, Rongjie Wang, Zhenyu Wang, Yawen Lai, and Ronggang Wang. Rethinking depth estimation for multi- view stereo: A unified representation. In CVPR, 2022. 2

  37. [45]

    Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to- 3d

    Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mu- tian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to- 3d. In CVPR, 2024. 2

  38. [46]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In ICML, 2021. 3, 6, 12, 14

  39. [47]

    Sam 2: Seg- ment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chlo´e Rolland, Laura Gustafson, et al. Sam 2: Seg- ment anything in images and videos. In ICLR, 2024. 3, 12

  40. [48]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, 2022. 2, 3, 6, 7, 12, 13

  41. [49]

    Hiera: A hi- erarchical vision transformer without the bells-and-whistles

    Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, Jitendra Ma- lik, Yanghao Li, and Christoph Feichtenhofer. Hiera: A hi- erarchical vision transformer without the bells-and-whi...

  42. [50]

    Mehdi S. M. Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani V ora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view syn- thesis through set-latent scene representations. In CVPR,

  43. [51]

    Structure-from-motion revisited

    Johannes Lutz Sch ¨onberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 16

  44. [52]

    Glu variants improve transformer

    Noam Shazeer. Glu variants improve transformer. arXiv:2002.05202, 2020. 16

  45. [53]

    Gamba: Marry gaussian splatting with mamba for single-view 3d recon- struction

    Qiuhong Shen, Zike Wu, Xuanyu Yi, Pan Zhou, Hanwang Zhang, Shuicheng Yan, and Xinchao Wang. Gamba: Marry gaussian splatting with mamba for single-view 3d recon- struction. IEEE TPAMI, 2025. 2

  46. [54]

    Mvdream: Multi-view diffusion for 3d gen- eration

    Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d gen- eration. In ICLR, 2023. 4

  47. [55]

    Philosophical Transactions of the Royal Society of London

    Royal society of London. Philosophical Transactions of the Royal Society of London. 1864. 2, 4

  48. [56]

    Generalizable patch-based neural render- ing

    Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural render- ing. In ECCV, 2022. 1, 3, 5

  49. [57]

    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. In CVPR, 2024. 2

  50. [58]

    Lgm: Large multi-view gaus- sian model for high-resolution 3d content creation

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaus- sian model for high-resolution 3d content creation. InECCV,

  51. [59]

    Hisplat: Hierarchical 3d gaus- sian splatting for generalizable sparse-view reconstruction

    Shengji Tang, Weicai Ye, Peng Ye, Weihao Lin, Yang Zhou, Tao Chen, and Wanli Ouyang. Hisplat: Hierarchical 3d gaus- sian splatting for generalizable sparse-view reconstruction. ICLR, 2024. 1, 2, 4

  52. [60]

    Deit iii: Revenge of the vit

    Hugo Touvron, Matthieu Cord, and Herv ´e J ´egou. Deit iii: Revenge of the vit. In ECCV, 2022. 3, 6, 12

  53. [61]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv:2302.13971, 2023. 13

  54. [62]

    Grf: Learning a general ra- diance field for 3d representation and rendering

    Alex Trevithick and Bo Yang. Grf: Learning a general ra- diance field for 3d representation and rendering. In ICCV,

  55. [63]

    Gs2-gnesf: Geometry- semantics synergy for generalizable neural semantic fields

    Chengshun Wang and Na Zhao. Gs2-gnesf: Geometry- semantics synergy for generalizable neural semantic fields. In ACM MM, 2024

  56. [64]

    Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021

  57. [65]

    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. In CVPR,

  58. [66]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. In CVPR, 2024. 3, 12, 14

  59. [67]

    Real-esrgan: Training real-world blind super-resolution with pure synthetic data

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV, 2021. 13

  60. [68]

    Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes

    Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. NeurIPS, 2024. 1

  61. [69]

    View-consistent 3d editing with gaus- sian splatting

    Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaus- sian splatting. In ECCV, 2024. 1

  62. [70]

    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 TIP, 2004. 16

  63. [71]

    latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction

    Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. In ECCV,

  64. [72]

    Lrm-zero: Training large reconstruction models with syn- thesized data

    Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, S ¨oren Pirk, Arie Kaufman, Xin Sun, and Hao Tan. Lrm-zero: Training large reconstruction models with syn- thesized data. NeurIPS, 2024. 1

  65. [73]

    Cl-mvsnet: Unsupervised multi-view stereo with dual-level contrastive learning

    Kaiqiang Xiong, Rui Peng, Zhe Zhang, Tianxing Feng, Jianbo Jiao, Feng Gao, and Ronggang Wang. Cl-mvsnet: Unsupervised multi-view stereo with dual-level contrastive learning. In ICCV, 2023. 2

  66. [74]

    Unifying flow, stereo and depth estimation

    Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. ICCV Workshops, 2023. 2, 3

  67. [75]

    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. In CVPR, 2024. 5

  68. [76]

    Depthsplat: Connecting gaussian splatting and depth.CVPR,

    Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth.CVPR,

  69. [77]

    Dmv3d: Denoising multi- view diffusion using 3d large reconstruction model

    Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Ji- ahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, and Kai Zhang. Dmv3d: Denoising multi- view diffusion using 3d large reconstruction model. InICLR,

  70. [78]

    Grm: Large gaussian reconstruction model for ef- ficient 3d reconstruction and generation

    Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wet- zstein. Grm: Large gaussian reconstruction model for ef- ficient 3d reconstruction and generation. In ECCV, 2024. 1, 2, 3, 4

  71. [79]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 3, 12, 14

  72. [80]

    Depth any- thing v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2. NeurIPS, 2024. 3, 12, 14

  73. [81]

    Mvsnet: Depth inference for unstructured multi-view stereo

    Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In ECCV, 2018. 2, 3

  74. [82]

    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. 5

  75. [83]

    Geolrm: Geometry-aware large re- construction model for high-quality 3d gaussian generation

    Chubin Zhang, Hongliang Song, Yi Wei, Chen Yu, Jiwen Lu, and Yansong Tang. Geolrm: Geometry-aware large re- construction model for high-quality 3d gaussian generation. In NeurIPS, 2024. 2

  76. [84]

    Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers

    Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. In AAAI, 2024. 1, 2, 4, 5, 6, 8

  77. [85]

    Gs-lrm: Large recon- struction model for 3d gaussian splatting

    Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large recon- struction model for 3d gaussian splatting. In ECCV, 2024. 1, 2, 3, 4, 13, 14, 15

  78. [86]

    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. 4, 16

  79. [87]

    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. In CVPR, 2024. 1, 2

  80. [88]

    Stereo magnification: Learning view syn- thesis using multiplane images

    Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view syn- thesis using multiplane images. In ACM TOG, 2018. 5, 16

  81. [89]

    Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers

    Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, 2024. 2 11 A. Additional Analysis A.1. Visual Encoder Analysis ...

  82. [2024]

    1, 2, 4, 5, 6, 14, 15, 16

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.