Pith. sign in

REVIEW 4 major objections 6 minor 53 references

HPR3D: Hierarchical Proxy Representation for High-Fidelity 3D Reconstruction and Controllable Editing

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A sparse tree of proxy nodes can represent a 3D object compactly and support direct drag-based multi-scale editing.

desk verdict A genuinely useful hierarchical representation for 3D editing, but the reconstruction claim is inherited from an unnamed external mesh step and undercut by the MASH baseline. read the letter →

arxiv 2507.11971 v1 pith:OMG4FNOI submitted 2025-07-16 cs.GR cs.CV

Tielong Wang , Yuxuan Xiong , Jinfan Liu , Zhifan Zhang , Ye Chen , Yue Shi , Bingbing Ni This is my paper · ORCID
classification cs.GRcs.CV
keywords hierarchicalproxyrepresentation3Dreconstructioncontrollableeditingtexturetransferoctreepartitioningmulti-scaleMLPdecodingObjaverse
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

This paper proposes HPR3D, a 3D representation in which an object's shape and texture are stored on a small hierarchy of proxy points (each with a position, a normal, and a learned texture feature) organized as a tree. The authors claim this representation is compact enough to reconstruct high-fidelity geometry and texture from multi-view renders, and is directly editable by dragging proxy points, with high-level points moving large regions and low-level points refining details. On 1,000 Objaverse objects, they report better texture fidelity than FlexiCubes and faster geometry fitting than MASH while also supporting texture transfer between regions. The point of the paper is that a single representation can serve reconstruction and controllable editing, which meshes, neural fields, and Gaussian splats each handle only partially.

What carries the argument

The load-bearing object is the hierarchical proxy point set, $\{(p_i^{(l)}, n_i^{(l)}, f_i^{(l)})\}$ for levels $l=1,\dots,L$: a tree of control points where each node carries a 3D position, a surface normal, and a texture feature. It is built by octree partitioning of the bottom-level mesh vertices and a curvature-based clustering criterion (CBCC): within each cell a least-squares plane fit to the points' normals produces a candidate proxy, accepted if the fitting error is below threshold $\epsilon$, otherwise the cell's points are kept as separate proxies. Texture features are propagated from the top level down and, with a positional encoding of the relative offset between a proxy and its parent, decoded by a small MLP into RGB. Deformation is propagated by the influence weight $w_{ji} = \exp(-\|p_j^{(1)} - p_i^{(l)}\|_2/\tau)$ before a final Laplacian editing pass.

What would settle it

Take one object and feed it through two different upstream mesh extraction algorithms with known different Chamfer distances; if the final HPR3D Chamfer distance is identical whenever the initial meshes differ, then the hierarchical fitting contributes no independent geometric accuracy and the reconstruction claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that hierarchical proxy nodes provide a more universal 3D representation than dense meshes, voxels, point clouds, or neural fields: geometry and texture are both encoded by L levels of proxy points plus a small MLP decoder, and every query of a surface point interpolates features from nearby and parent proxy nodes. The bottom level starts from the vertices of an input mesh; recursive octree partitioning with error-guided plane fitting builds coarser levels, and texture features are propagated top-down with positional encoding and decoded into RGB. The paper shows this achieves Chamfer Distance 0.0207, PSNR 37.12, and SSIM 0.9858 on its Objaverse test set, better texture than FlexiCubes and comparable or faster geometry than MASH, with joint geometry-texture parameter count of 2.6e6 and geometry-only fitting in 30 seconds. Editing follows directly from the representation: moving a proxy point deforms affected mesh vertices through an exponential falloff weight followed by Laplacian editing, while editing proxy features enables multi-scale texture changes and warp-free texture transfer.

Load-bearing premise

The method depends on an unnamed upstream mesh reconstruction algorithm to produce the bottom-level vertices, and if that mesh is coarse or biased the proxy hierarchy inherits its errors.

Editorial extensions

If this is right

  • Multi-scale shape editing: dragging few high-level proxy points changes whole parts, while low-level proxy points refine local details, with no manual handle selection.
  • Texture transfer: aligning same-level proxy features between two regions and interpolating them yields texture transfer that does not warp or distort, for regions of any shape or size.
  • Compact storage: joint geometry and texture fit in about 2.6 million parameters, and geometry-only fitting takes 30 seconds rather than the 1.5 hours reported for the anchor-point baseline.
  • Input modality flexibility: the same hierarchy can be built from a mesh, point cloud, implicit field, or multi-view images, because only a mesh initialization is required upstream.
  • Combined editing: because geometry and texture live on the same proxy nodes, an artist can mix large-scale shape changes with fine texture edits in one representation.

Reading between the lines

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

  • If the representation is as modular as claimed, one natural test the paper does not run is to swap the upstream mesh reconstruction algorithm and measure how much of the final error is inherited; the paper fixes no baseline for that step.
  • The same proxy hierarchy could be treated as a generative latent space: if the feature vectors at each level are statistically regular, sampling or interpolating them might generate new shapes, an extension the paper leaves implicit.
  • The texture-transfer mechanism rests on rigid alignment of proxy points between regions; applying it to non-rigidly corresponding or partially occluded surfaces would require an alignment step the paper does not provide.
  • Because the bottom-level proxy points are exactly mesh vertices, the representation's geometric ceiling is the input mesh's resolution; a testable prediction is that finer input meshes should improve reconstruction but not editing locality, which depends on the hierarchy levels.
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 / 6 minor

Summary. The paper introduces HPR3D, a hierarchical proxy-point representation for 3D objects. Given a multimodal input, the pipeline first reconstructs a mesh using an unspecified 'existing mesh reconstruction algorithm,' initializes the bottom-level proxy points from that mesh's vertices, and builds higher levels through octree spatial partitioning and error-guided plane fitting (Eq. 2). Texture features are attached to proxy points at each level and optimized with a rendering loss, then decoded by a small MLP. Editing is performed by displacing proxy points with an exponential influence weight (Eq. 6) followed by Laplacian editing, and by transferring texture features between proxy points. Experiments on 1,000 Objaverse objects report Chamfer distance for geometry and PSNR/SSIM for texture compared with DMTet, FlexiCubes, and MASH, along with qualitative editing results and a component ablation.

Significance. If the central claims hold, HPR3D would be a useful representation: a single compact hierarchy supporting both reconstruction and multi-scale editing, with reported low parameter counts, 30s geometry optimization, and higher texture PSNR than FlexiCubes. The idea of hierarchical control points with distance-based influence is simple and plausible, and the paper includes quantitative comparisons, an ablation, and clear descriptions of the proxy construction. However, the current evaluation does not fully establish the headline reconstruction claim because the bottom-level geometry is inherited from an unnamed external mesh reconstruction step, and the paper's own Table 1 shows MASH achieving a lower Chamfer distance. The editing contribution is supported only by qualitative figures. With controlled experiments and qualified claims, the core idea is likely salvageable.

major comments (4)
  1. [Methodology, Adaptive Proxy Reconstruction] The pipeline says 'we first reconstruct its geometry as a mesh using an existing mesh reconstruction algorithm' but never names that algorithm, and the bottom-level proxy points are initialized directly from the mesh vertices with no subsequent refinement of vertex positions. As a result, the Chamfer distance reported in Table 1 measures the fidelity of the external mesh plus the proxy compression of that mesh, not HPR3D's reconstruction from the multimodal input. The reported geometry-only time of 30s and geometry parameter count also exclude or conflate this upstream step. The authors should name the external algorithm, use a fixed input representation for all methods, and either include the upstream cost or restrict the claim to mesh-compression and editing rather than reconstruction.
  2. [Table 1 and Abstract/Contributions] Table 1 reports MASH with CD 0.0145, which is lower than HPR3D's 0.0207, yet the abstract and the contribution list claim 'state-of-the-art performance on tasks such as reconstruction.' As written, this is internally contradicted by the paper's own quantitative results. If the intended comparison is joint geometry-and-texture reconstruction or a speed-accuracy trade-off, the claim should be qualified explicitly; otherwise the reconstruction claim should be revised.
  3. [Easy 3D Editing] Editing is presented as a core contribution, but the evaluation in Figs. 3 and 4 is purely qualitative. The only baseline is a manual Laplacian-editing workflow in Blender, with no quantitative measure of deformation quality, editing effort, geometric distortion, or texture-preservation accuracy. To support the claimed 'superior editability,' the authors should provide a quantitative editing evaluation (e.g., error against a target deformation, local distortion metrics, or a user study) and a controlled comparison with an automatic editing baseline.
  4. [Ablation Study, Table 2] The component ablation evaluates only PSNR and SSIM on a subset and never reports geometric metrics such as Chamfer distance. Since the curvature-based clustering criterion (CBCC) is claimed to allocate proxy points based on local geometric complexity, its effect on geometric reconstruction quality should be measured. The same applies to the multi-level geometry hierarchy; without geometry ablations, the contribution of the hierarchical proxy construction to the reported CD is unverified.
minor comments (6)
  1. [Methodology, Eq. (5)-(6)] The influence weight w^{(l)}_{ji} is used in Eq. (5) before its definition in Eq. (6); please reorder or add a forward reference, and clarify whether the influence is computed only from the bottom-level point to the edited proxy or propagated recursively through intermediate levels.
  2. [Experiments, Implementation Details] The clustering error threshold ε is set to 5.0, but the units or normalization of the least-squares error in Eq. (2) are not specified; without this, the threshold value is not interpretable or reproducible.
  3. [Table 1] The time columns 'Time (G)' and 'Time (G+T)' are not defined consistently across methods; for HPR3D, clarify whether the 30s geometry time includes the external mesh reconstruction step, and for DMTet and MASH, state what input representation they receive.
  4. [Easy 3D Editing, Texture Editing] The texture-transfer procedure is described only verbally ('rigidly align two regions... perform linear interpolation'); equations or pseudo-code for the alignment, correspondence, and interpolation steps are needed for reproducibility.
  5. [Experiments, Dataset] The paper does not state how the 1,000-object test set is selected from Objaverse or which input modality (mesh, point cloud, or multi-view images) is provided to each baseline; these details affect the comparability of the results.
  6. [Related Works and Contributions] The claim that MASH 'lacks support for intuitive and flexible editing' is not demonstrated in the experiments; since MASH also uses anchor points, the paper should explain more concretely why MASH's anchors cannot be edited in a similar way.

Circularity Check

1 steps flagged · score 6.0 of 10

Geometry reconstruction reduces by construction to the external mesh used as input; texture and editing claims remain independent.

  1. fitted input called prediction [Methodology, Adaptive Proxy Reconstruction, final paragraph; Fig. 1 caption; Implementation Details; Table 1]
    "Given the multimodal input S, we first reconstruct its geometry as a mesh, denoted as M = (V, F), where V is the set of vertices. We initialize V as C(1), and from the recursive relations above, we obtain the multi-level proxy points from C(1) to C(L)."

    The bottom-level proxy points C(1) are initialized to be exactly the vertex set V of the mesh produced by an unspecified 'existing mesh reconstruction algorithm'. The subsequent adaptive construction (Eq. 2 and the threshold rule) either clusters those vertices into higher-level proxy points or retains them individually when the fitting error exceeds the threshold; no step updates the positions p(1) to improve geometric fidelity. Therefore the final bottom-level geometry is identical to the input mesh's vertex set by construction. The Chamfer Distance reported in Table 1 (CD 0.0207) is consequently the CD of the external mesh-reconstruction algorithm's output, not a quantity derived or refined by HPR3D.

full rationale

The central geometry-reconstruction result is not self-contained: the bottom level of the hierarchical proxy representation is seeded from, and never moves beyond, the vertices of a mesh obtained from an unnamed 'existing mesh reconstruction algorithm'. Because the adaptive clustering in Eq. 2 only partitions and re-labels those vertices (retaining them individually when error is high), the final geometry used for CD evaluation is the input mesh itself. This makes the headline reconstruction number inherited rather than produced by the proposed representation. The texture pipeline is different: texture features and the MLP decoder are optimized per object against a multi-view rendering loss, so PSNR/SSIM measure HPR3D's own fitted representation rather than an imported result. The editing demonstrations are qualitative and rely on the hierarchical proxy structure plus Laplacian editing, which are independent of the external mesh's accuracy. Self-citations to Xiong et al. 2025 and Shi et al. 2021/2023/2025 are related-work references and are not load-bearing for the derivation. Because one of the two headline capabilities (geometry reconstruction) reduces by construction to its input, while texture reconstruction and editing retain independent content, the circularity score is 6.

Assumptions & free parameters 7 free parameters · 3 assumptions · 1 invented entities

The central claim rests on several hand-set hyperparameters and an under-specified upstream mesh reconstruction. The proxy node data structure is an invented computational entity, not a physically motivated one. The plane-fitting and exponential-weight models are additional assumptions that the paper does not independently validate.

free parameters (7)
  • clustering error threshold epsilon = 5.0
    Hand-set threshold in Eq. 2 that controls whether a proxy point is accepted or all points are retained; directly controls hierarchy density and is not tuned via any reported search.
  • influence temperature tau = 1.0
    Hand-set scalar in Eq. 6 controlling the spatial falloff of proxy-point displacement on lower-level points.
  • texture feature dimensions per level = F^(1)=32, F^(2)=24, F^(3)=12
    Hand-chosen per-level feature widths for the texture decoder; no justification or sensitivity analysis is reported.
  • positional encoding dimension = 60
    Hand-chosen dimension for PE in Eq. 3.
  • loss weight lambda = 0.5
    Hand-set weight for auxiliary texture loss in Eq. 4.
  • number of levels L and octree max resolution R = L=3, R=7
    Architectural choices for the hierarchy; no sensitivity analysis is reported.
  • per-object texture features and decoder weights = optimized
    Fitted per object to multi-view renderings via Lrender; these are the representation parameters and are not independently constrained before optimization.
assumptions (3)
  • domain assumption The target object can be reconstructed into a high-quality mesh by an existing algorithm, and mesh vertices are a sufficient bottom-level geometry representation.
    Stated in the overview: 'we first reconstruct its geometry as a mesh using an existing mesh reconstruction algorithm.' The accuracy of the final geometry is therefore bounded by an unstated external method.
  • standard math Local surface patches can be approximated as planes defined by point normals, so least-squares fitting in Eq. 2 yields meaningful proxy points.
    This is a standard local-planarity assumption used in point-cloud processing; it may fail at sharp features, which the paper mitigates by retaining points when the fitting error exceeds a threshold.
  • ad hoc to paper Displacement of a proxy point propagates to bottom-level points through the exponential weight exp(-d/tau), and Laplacian editing then preserves shape coherence.
    This is the paper's chosen deformation model; it is not derived from physics or from a prior deformation theory and is only shown qualitatively in Fig. 3.
invented entities (1)
  • Hierarchical proxy node
    purpose: A new compact 3D representation storing position, normal, and texture features at multiple tree-structured levels, used for reconstruction and editing.
    The proxy nodes are a computational construct introduced by this paper; they have no independent falsifiable prediction outside the method. Their utility is measured only through the paper's own reconstruction and editing experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HPR3D: Hierarchical Proxy Representation for High-Fidelity 3D Reconstruction and Controllable Editing." pith.science (2026). https://pith.science/paper/OMG4FNOI

@misc{pith2026250711971,
  author       = {Pith},
  title        = {Pith review of: HPR3D: Hierarchical Proxy Representation for High-Fidelity 3D Reconstruction and Controllable Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OMG4FNOI}},
  note         = {Machine review of arXiv:2507.11971}
}
read the original abstract

Current 3D representations like meshes, voxels, point clouds, and NeRF-based neural implicit fields exhibit significant limitations: they are often task-specific, lacking universal applicability across reconstruction, generation, editing, and driving. While meshes offer high precision, their dense vertex data complicates editing; NeRFs deliver excellent rendering but suffer from structural ambiguity, hindering animation and manipulation; all representations inherently struggle with the trade-off between data complexity and fidelity. To overcome these issues, we introduce a novel 3D Hierarchical Proxy Node representation. Its core innovation lies in representing an object's shape and texture via a sparse set of hierarchically organized (tree-structured) proxy nodes distributed on its surface and interior. Each node stores local shape and texture information (implicitly encoded by a small MLP) within its neighborhood. Querying any 3D coordinate's properties involves efficient neural interpolation and lightweight decoding from relevant nearby and parent nodes. This framework yields a highly compact representation where nodes align with local semantics, enabling direct drag-and-edit manipulation, and offers scalable quality-complexity control. Extensive experiments across 3D reconstruction and editing demonstrate our method's expressive efficiency, high-fidelity rendering quality, and superior editability.

Figures

Figures reproduced from arXiv: 2507.11971 by the authors.

Figure 1
Figure 1. Overview of our framework. We propose a parameterized representation based on a hierarchical structure of proxy points, which enables accurate and efficient reconstruction of 3D objects while supporting multi-scale geometry and texture editing. Given a 3D object in any modality, we first reconstruct its mesh using an existing mesh reconstruction algorithm. The vertices of the reconstructed mesh are then used to init… view at source ↗
Figure 2
Figure 2. Qualitative comparison on Objaverse dataset. Qualitative reconstruction results are presented. DMTet (Shen et al. 2021) and MASH (Li et al. 2025) show only geometric reconstructions, as they do not support texture modeling. In contrast, our method and FlexiCubes (Shen et al. 2023) provide full reconstructions, including both geometry and texture. The last column visualizes the hierarchical proxy point structure prod… view at source ↗
Figure 3
Figure 3. The second column displays large-scale edits pro [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 46 canonical work pages

  1. [1]

    Athar, S.; Xu, Z.; Sunkavalli, K.; Shechtman, E.; and Shu, Z. 2022. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20364--20373

    work page 2022

  2. [2]

    Baorui, M.; Zhizhong, H.; Yu-Shen, L.; and Matthias, Z. 2021. Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces. In Proceedings of the International Conference on Machine Learning

    work page 2021

  3. [3]

    Barr, A. H. 1984. Global and local deformations of solid primitives. In Proceedings of the 11th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '84, 21–30. New York, NY, USA: Association for Computing Machinery. ISBN 0897911385

    work page 1984

  4. [4]

    T.; Mildenhall, B.; Tancik, M.; Hedman, P.; Martin - Brualla, R.; and Srinivasan, P

    Barron, J. T.; Mildenhall, B.; Tancik, M.; Hedman, P.; Martin - Brualla, R.; and Srinivasan, P. P. 2021. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 5835--5844

    work page 2021

  5. [5]

    Chen, A.; Xu, Z.; Geiger, A.; Yu, J.; and Su, H. 2022. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, 333--350. Springer

    work page 2022

  6. [6]

    Deitke, M.; Schwenk, D.; Salvador, J.; Weihs, L.; Michel, O.; VanderBilt, E.; Schmidt, L.; Ehsani, K.; Kembhavi, A.; and Farhadi, A. 2023. Objaverse: A Universe of Annotated 3D Objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 13142--13153

    work page 2023

  7. [7]

    Gropp, A.; Yariv, L.; Haim, N.; Atzmon, M.; and Lipman, Y. 2020. Implicit Geometric Regularization for Learning Shapes. In Proceedings of the International Conference on Machine Learning, 3569--3579

    work page 2020

  8. [8]

    Gu \'e don, A.; and Lepetit, V. 2024. SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5354--5363

    work page 2024

Show all 53 references
  1. [9]

    Hamdi, A.; Melas-Kyriazi, L.; Mai, J.; Qian, G.; Liu, R.; Vondrick, C.; Ghanem, B.; and Vedaldi, A. 2024. Ges: Generalized exponential splatting for efficient radiance field rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19812--19822

    2024

  2. [10]

    Huang, B.; Yu, Z.; Chen, A.; Geiger, A.; and Gao, S. 2024. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In ACM SIGGRAPH 2024 Conference Papers. Association for Computing Machinery

    2024

  3. [11]

    Hui, K.-H.; Li, R.; Hu, J.; and Fu, C.-W. 2022. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 18572--18582

    2022

  4. [12]

    Jakab, T.; Tucker, R.; Makadia, A.; Wu, J.; Snavely, N.; and Kanazawa, A. 2021. Keypointdeformer: Unsupervised 3d keypoint discovery for shape control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12783--12792

    2021

  5. [13]

    Jiang, Y.; Tu, J.; Liu, Y.; Gao, X.; Long, X.; Wang, W.; and Ma, Y. 2024. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5322--5332

    2024

  6. [14]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics (ToG), 42(4): 139--1

    2023

  7. [15]

    Li, C.; Xin, Y.; Zhou, X.; Shamir, A.; Zhang, H.; Liu, L.; and Hu, R. 2025. MASH: Masked Anchored SpHerical Distances for 3D Shape Representation and Generation. arXiv preprint arXiv:2504.09149

    2025 arXiv

  8. [16]

    Li, J.; Shi, Y.; Cao, J.; Ni, B.; Zhang, W.; Zhang, K.; and Gool, L. V. 2024. Mipmap-GS: Let Gaussians Deform with Scale-specific Mipmap for Anti-aliasing Rendering. In International Conference on 3D Vision

    2024

  9. [17]

    Li, M.; and Zhang, H. 2021. D2im-net: Learning detail disentangled implicit fields from single images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10246--10255

    2021

  10. [18]

    Liang, Z.; Zhang, Q.; Feng, Y.; Shan, Y.; and Jia, K. 2024. Gs-ir: 3d gaussian splatting for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21644--21653

    2024

  11. [19]

    Lin, J.; Li, Z.; Tang, X.; Liu, J.; Liu, S.; Liu, J.; Lu, Y.; Wu, X.; Xu, S.; Yan, Y.; et al. 2024. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5166--5175

    2024

  12. [20]

    Liu, L.; Gu, J.; Zaw Lin, K.; Chua, T.-S.; and Theobalt, C. 2020. Neural sparse voxel fields. In Advances in Neural Information Processing Systems, volume 33, 15651--15663

    2020

  13. [21]

    Liu, S.; Zhang, X.; Zhang, Z.; Zhang, R.; Zhu, J.; and Russell, B. 2021. Editing Conditional Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 5753--5763

    2021

  14. [22]

    E.; and Cline, H

    Lorensen, W. E.; and Cline, H. E. 1998. Marching cubes: A high resolution 3D surface construction algorithm. In Seminal Graphics: Pioneering Efforts that Shaped the Field, 347--353

    1998

  15. [23]

    S.; Barron, J

    Martin-Brualla, R.; Radwan, N.; Sajjadi, M. S.; Barron, J. T.; Dosovitskiy, A.; and Duckworth, D. 2021. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7210--7219

    2021

  16. [24]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

    2021

  17. [25]

    M\"uller, T.; Evans, A.; Schied, C.; and Keller, A. 2022. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Transactions on Graphics (ToG), 41(4): 102:1--102:15

    2022

  18. [26]

    Oechsle, M.; Peng, S.; and Geiger, A. 2021. UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision

    2021

  19. [27]

    J.; Florence, P.; Straub, J.; Newcombe, R.; and Lovegrove, S

    Park, J. J.; Florence, P.; Straub, J.; Newcombe, R.; and Lovegrove, S. 2019. DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2019

  20. [28]

    T.; Bouaziz, S.; Goldman, D

    Park, K.; Sinha, U.; Barron, J. T.; Bouaziz, S.; Goldman, D. B.; Seitz, S. M.; and Martin-Brualla, R. 2021. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 5865--5874

    2021

  21. [29]

    Pumarola, A.; Corona, E.; Pons-Moll, G.; and Moreno-Noguer, F. 2021. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10318--10327

    2021

  22. [30]

    Reiser, C.; Peng, S.; Liao, Y.; and Geiger, A. 2021. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 14335--14345

    2021

  23. [31]

    Shen, T.; Gao, J.; Yin, K.; Liu, M.-Y.; and Fidler, S. 2021. Deep Marching Tetrahedra: a Hybrid Representation for High-resolution 3d Shape Synthesis. In Advances in Neural Information Processing Systems, volume 34, 6087--6101

    2021

  24. [32]

    Shen, T.; Munkberg, J.; Hasselgren, J.; Yin, K.; Wang, Z.; Chen, W.; Gojcic, Z.; Fidler, S.; Sharp, N.; and Gao, J. 2023. Flexible Isosurface Extraction for Gradient-Based Mesh Optimization. ACM Transactions on Graphics (ToG), 42(4): 37--1

    2023

  25. [33]

    Shi, Y.; Ni, B.; Liu, J.; Rong, D.; Qian, Y.; and Zhang, W. 2021. Geometric Granularity Aware Pixel-To-Mesh. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 13097--13106

    2021

  26. [34]

    Shi, Y.; Rong, D.; Chen, C.; Ma, C.; Ni, B.; and Zhang, W. 2025. DARF: Depth-Aware Generalizable Neural Radiance Field. Displays, 88: 102996

    2025

  27. [35]

    Shi, Y.; Xiong, Y.; Chai, J.; Ni, B.; and Zhang, W. 2023. USR: Unsupervised Separated 3D Garment and Human Reconstruction via Geometry and Semantic Consistency. arXiv preprint arXiv:2302.10518

    2023 arXiv

  28. [36]

    Sorkine, O.; and Alexa, M. 2007. As-Rigid-As-Possible Surface Modeling. In Proceedings of EUROGRAPHICS/ACM SIGGRAPH Symposium on Geometry Processing, 109--116

    2007

  29. [37]

    Sorkine, O.; Cohen-Or, D.; Lipman, Y.; Alexa, M.; R\" o ssl, C.; and Seidel, H.-P. 2004. Laplacian surface editing. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH Symposium on Geometry Processing, SGP '04, 175–184. New York, NY, USA: Association for Computing Machinery. ...

    2004

  30. [38]

    Wang, P.; Liu, L.; Liu, Y.; Theobalt, C.; Komura, T.; and Wang, W. 2021. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction. In Ranzato, M.; Beygelzimer, A.; Dauphin, Y. N.; Liang, P.; and Vaughan, J. W., eds., Advances in Neural Informat...

    2021

  31. [39]

    Wang, X.; Zhu, J.; Ye, Q.; Huo, Y.; Ran, Y.; Zhong, Z.; and Chen, J. 2023 a . Seal-3D: Interactive Pixel-Level Editing for Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 17637--17647

    2023

  32. [40]

    Wang, Y.; Han, Q.; Habermann, M.; Daniilidis, K.; Theobalt, C.; and Liu, L. 2023 b . NeuS2: Fast Learning of Neural Implicit Surfaces for Multi-view Reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision

    2023

  33. [41]

    C.; Sheikh, H

    Wang, Z.; Bovik, A. C.; Sheikh, H. R.; and Simoncelli, E. P. 2004. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4): 600--612

    2004

  34. [42]

    Wolf, Y.; Bracha, A.; and Kimmel, R. 2024. GS 2 M esh: Surface Reconstruction from G aussian Splatting via Novel Stereo Views. In European Conference on Computer Vision

    2024

  35. [43]

    Xiong, Y.; Shi, Y.; Dou, Y.; and Ni, B. 2025. DualNeRF: Text-Driven 3D Scene Editing via Dual-Field Representation. arXiv preprint arXiv:2502.16302

    2025 arXiv

  36. [44]

    F.; Chen, Y.; and Lee, G

    Yan, Z.; Low, W. F.; Chen, Y.; and Lee, G. H. 2024. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20923--20931

    2024

  37. [45]

    Yariv, L.; Gu, J.; Kasten, Y.; and Lipman, Y. 2021. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems

    2021

  38. [46]

    Yu, A.; Li, R.; Tancik, M.; Li, H.; Ng, R.; and Kanazawa, A. 2021. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 5752--5761

    2021

  39. [47]

    Yu, Z.; Chen, A.; Huang, B.; Sattler, T.; and Geiger, A. 2024. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19447--19456

    2024

  40. [48]

    Yu, Z.; Peng, S.; Niemeyer, M.; Sattler, T.; and Geiger, A. 2022. MonoSDF: Exploring Monocular Geometric Cues for Neural Implicit Surface Reconstruction. In Advances in Neural Information Processing Systems

    2022

  41. [49]

    Yu, Z.; Sattler, T.; and Geiger, A. 2024. Gaussian Opacity Fields: Efficient Adaptive Surface Reconstruction in Unbounded Scenes. ACM Transactions on Graphics (ToG)

    2024

  42. [50]

    Yuan, Y.-J.; Sun, Y.-T.; Lai, Y.-K.; Ma, Y.; Jia, R.; and Gao, L. 2022. NeRF-editing: geometry editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 18353--18364

    2022

  43. [51]

    Zhang, K.; Riegler, G.; Snavely, N.; and Koltun, V. 2020. NeRF++: Analyzing and Improving Neural Radiance Fields. arXiv preprint arXiv:2010.07492

    2020 arXiv

  44. [52]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  45. [53]

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