StructuredField: Unifying Structured Geometry and Radiance Field

Jinkai Cui; Juyong Zhang; Kaiwen Song; Zherui Qiu

REVIEW 2 major objections 5 minor 5 references

StructuredField: Unifying Structured Geometry and Radiance Field

T0 review · 2 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read StructuredField renders like 3DGS while keeping a simulation-ready tetrahedral mesh

desk verdict Useful structured-mesh radiance field with a real overclaim: the homeomorphism alone does not guarantee inversion-free meshes, and Eq. 9's determinant is wrong; fix the theory and it's a solid paper. read the letter →

arxiv 2501.18152 v2 pith:FHMDP36R submitted 2025-01-30 cs.GR

Kaiwen Song , Jinkai Cui , Zherui Qiu , Juyong Zhang This is my paper · ORCID

classification cs.GR

keywords tetrahedralmeshradiancefield3DGaussianSplattinginvertibleneuralnetworkdifferentiablerenderingphysicssimulationreconstructionconformal

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

StructuredField aims to close the gap between two families of 3D representations: unstructured point-based radiance fields, which reconstruct scenes with high fidelity but cannot be used directly in existing graphics pipelines, and structured tetrahedral meshes, which are ideal for simulation and editing but hard to optimize for rendering. The paper claims that a tetrahedral mesh can be made fully differentiable by reparameterizing each tetrahedron's attributes as the parameters of a 3D Gaussian primitive, so image-based losses flow directly into vertex positions. To keep the mesh valid during optimization, the authors constrain vertex motion with an orientation-preserving homeomorphism implemented as an invertible neural network, and they add an implicit hierarchical subdivision so that fine detail is captured without breaking the conformality of the base mesh. The payoff is a single representation from which one obtains both high-fidelity novel views (comparable to 3D Gaussian Splatting on standard benchmarks) and a conformal, inversion-free mesh that can be handed to physics simulation, deformation, and level-of-detail pipelines.

What carries the argument

The machinery has three parts. (1) The reparameterization map $F_r$ (Eq. 1) sends each tetrahedron's vertex positions, spherical-harmonic coefficients, and weights to the mean, covariance, rotation, color, and opacity of a 3D Gaussian, making the mesh directly optimizable by image loss. (2) Hierarchical implicit subdivision splits base tetrahedra 1-to-4 with a new optimizable control point parameterized by barycentric coordinates inside the parent, so the root mesh stays conformal for simulation while leaf tetrahedra provide adaptive multi-scale detail for rendering. (3) The orientation-preserving homeomorphism is a permutation-based Real-NVP invertible network whose coupling layers keep the Jacobian determinant positive, applied to vertex positions to define a feasible deformation space that is supposed to exclude inverted elements and self-intersections. Together these parts let appearance optimization and geometry quality coexist in one object.

What would settle it

Train the model on a fixed scene with the orientation-preserving homeomorphism active but the quality loss disabled, at maximum subdivision depth, and count inverted tetrahedra in the final straight-edged mesh; any nonzero count for a scene that was supposed to be guaranteed inversion-free would refute the claim that the homeomorphism alone secures mesh validity.

Watch

Extended reading notes

Core claim

The central claim is that structured geometry and radiance-field rendering need not be separate representations: one can optimize a tetrahedral mesh directly from multi-view images by reparameterizing it into 3D Gaussian primitives. Each tetrahedron $T_k$ is mapped to a Gaussian $\Theta_k = F_r(T_k)$ whose mean and initial covariance are computed from the weighted vertex positions via a PCA-style estimate $\Sigma'_k$, together with an optimizable rotation, per-vertex spherical-harmonic colors, and per-tetrahedron opacity. Rendering then follows the standard 3D Gaussian splatting procedure, but gradients reach the mesh vertices, so appearance drives geometry. To keep the geometry usable, the base mesh is optimized through an orientation-preserving homeomorphism implemented as a permuted Real-NVP invertible network, whose positive Jacobian determinant is claimed to prevent element inversion and self-intersection. Detail is added by implicit 1-to-4 subdivision: root tetrahedra keep the conformal topology used for simulation, while leaf tetrahedra form the rendered mesh. The paper reports rendering quality comparable to 3DGS on the NeRF Synthetic, Shelly, and Mip-NeRF360 datasets, with final meshes that are conformal and, when the quality term is included, have zero inverted tetrahedra.

Load-bearing premise

The paper's guarantee of an inversion-free mesh rests on the continuous orientation-preserving homeomorphism, but the mesh that is actually used is a discrete straight-edged tetrahedralization; Table 1(c) shows that with the homeomorphism alone roughly 1,300 inverted tetrahedra remain, so the discrete mesh is not in fact guaranteed to stay inversion-free without the additional quality loss.

Editorial extensions

If this is right

A multi-view reconstruction yields one object that can be both rendered photorealistically and used directly by physics simulation, deformation, and level-of-detail pipelines, with no conversion step or representation switch.
Rendering quality is close to 3DGS on standard benchmarks: on NeRF Synthetic the method reports PSNR 33.53 versus 33.78 for 3DGS, and it slightly exceeds 3DGS on the Shelly dataset (PSNR 39.76 versus 39.61).
The final base mesh is conformal and, when the quality loss $L_{\text{quality}}$ is included, has zero inverted tetrahedra in the reported scenes, so it satisfies the topological requirements of standard simulation meshes.
Because every Gaussian is confined to its parent tetrahedron, deforming the mesh automatically deforms the radiance field; the paper demonstrates this with XPBD mass-spring simulation and lattice deformation.
Level-of-detail rendering is obtained by collapsing leaf tetrahedra back into their parents (taking the parent opacity as the max of the children), giving a continuous LOD chain from a single representation.

Reading between the lines

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

The continuous homeomorphism guarantees a positive Jacobian determinant for a continuous deformation, but the straight-edged discrete mesh is a different object: the paper's Table 1(c) reports about 1,300 inverted tetrahedra remaining with the homeomorphism alone, so the 'guarantee' acts as a strong regularizer and the quality term is what actually drives the count to zero.
The reparameterization scheme is not tied to 3DGS: the same $F_r$ idea could bind 2D Gaussians or convex splats to tetrahedra, yielding other hybrid representations with the same structured-geometry advantage.
A natural stress test is to push the homeomorphism outside its training regime, for instance by applying a larger lattice deformation or subdividing more deeply, and counting inverted elements; if the count rises, the continuous guarantee does not transfer to the discrete mesh under refinement.
The inverse direction, feeding a simulated or deformed mesh back through the reparameterization for re-rendering, could support appearance-aware physics editing; the paper demonstrates deformation but not editing of material or lighting, which would be a direct 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.

Desk Editor's Note

read the letter

Quick take: StructuredField is a useful, practical representation that mostly delivers what it promises, but the paper's headline guarantee—that the homeomorphism constraint keeps the mesh inversion-free—is contradicted by the paper's own ablation. Fix that and it's a solid TOG submission.

What's actually new: they reparameterize tetrahedron attributes into 3D Gaussian parameters (Eqs. 1–4), so you get differentiable rendering directly from a tetrahedral mesh. The hierarchical implicit subdivision keeps a conformal base mesh while allowing adaptive detail, which is a neat solution to the non-conformality problem. The applications (simulation, deformation, LOD) are real and demonstrated. Code is public.

The soft spots are real, and one is load-bearing. Table 1(c) shows that using the homeomorphism alone, about 1,300 inverted tetrahedra remain. That directly contradicts the stated guarantee that the mesh stays inversion-free and self-intersection-free during optimization and in the final result. The reason is not mysterious: H is a continuous invertible map, but the mesh uses straight edges between mapped vertices; the linear interpolation is not the image of the tetrahedron under H, so orientation preservation of H doesn't transfer to the discrete mesh. The final method gets to 0 inversions only after adding the L_quality loss, so the guarantee should be reframed as: homeomorphism reduces inversions, quality loss eliminates them. Also, Eq. (9) miscomputes the determinant: the displayed matrix has det = exp(s) > 0, so the claim that the non-permuted split fails to keep the determinant positive is wrong. The permutation strategy still makes the Jacobian triangular, which is a legitimate way to guarantee positivity, but the motivating example needs correction.

One more thing worth noting: the final pipeline (H + L_quality) trades rendering quality for mesh quality. On the ablation, PSNR drops from 34.64 to 33.64 and LPIPS from 0.022 to 0.0455 compared to H alone. That's a real cost, and the paper should be more transparent that the quality loss, not the homeomorphism, is what gives the clean mesh.

On the evidence and citation side: the comparisons to 3DGS, 2DGS, Adaptive Shells, etc., are reasonable. The same-group citations (NDR) are contextual, not load-bearing. No circularity.

Who this is for: anyone working on differentiable rendering with structured geometry, mesh-based radiance fields, or physics-aware reconstruction. It deserves peer review—the representation is interesting and the experiments are honest enough—but the authors should be asked to correct the math and reword the guarantee before acceptance.

Referee Report

2 major / 5 minor

Summary. The paper proposes StructuredField, a tetrahedral-mesh representation whose per-tetrahedron attributes are reparameterized as 3D Gaussian parameters (Eq. 1), enabling differentiable rendering through a Gaussian-splatting-like pipeline. A hierarchical implicit subdivision scheme (Sec. 3.2) adds multi-scale detail while keeping a conformal base mesh for downstream use. To preserve mesh quality during optimization, the vertex positions are constrained through an orientation-preserving homeomorphism implemented as a Real-NVP-style invertible network with a permutation strategy (Sec. 3.3). The authors claim this constraint guarantees that the tetrahedral mesh remains inversion-free and self-intersection-free during optimization and in the final result. Experiments on NeRF Synthetic, Shelly, and Mip-NeRF 360 show rendering quality comparable to 3DGS, while the resulting mesh supports physical simulation, deformation, and level-of-detail rendering.

Significance. If the claims hold, StructuredField is a valuable bridge between unstructured point-based radiance fields and structured tetrahedral geometry, enabling high-fidelity rendering and downstream simulation/editing from a single optimized representation. The reparameterization of tetrahedra as Gaussian primitives is simple and effective, and the ablation (Table 1) clearly demonstrates the benefit of the homeomorphism constraint over unconstrained optimization and a signed-volume loss. The paper ships code and includes a range of applications. However, the central guarantee of an inversion-free mesh is overstated and the theoretical derivation contains a concrete error, so the contribution needs revision before it can be fully accepted.

major comments (2)

[3.3, Table 1] The paper claims that the orientation-preserving homeomorphism 'guarantees' the tetrahedral mesh remains inversion-free and self-intersection-free during optimization and in the final result, but the ablation in Table 1 (c) shows that approximately 1.3k inverted tetrahedra remain when the homeomorphism is used alone. The homeomorphism H is applied only to vertex positions, while the mesh being evaluated is the straight-edged tetrahedron defined by those mapped vertices; a positive Jacobian for H does not imply that the convex hull of the mapped vertices has positive signed volume. The final inversion-free result in Table 1 (d) is achieved only with the additional L_quality loss (lambda_4 = 10). The guarantee claim should be weakened to describe a strong regularizer, or the authors must supply a proof that the discrete straight-edged mesh is also inversion-free under the proposed constraint.
[3.3, Eq. (9)] The determinant of the matrix displayed in Eq. (9) is exp(s) > 0, because the matrix is block lower-triangular with diagonal entries 1, exp(s), and 1. The text's assertion that 'its determinant does not remain positive' is therefore incorrect. This invalidates the stated motivation for the permutation strategy. The permutation approach may still be valid (det(P) = 1 for the given P in Eq. (10)), but the determinant computation should be corrected and the argument for why the permutation is needed should be reworked or removed.

minor comments (5)

[3.3, Eq. (10)] The text states that the permutation is used 'when we need to keep the x and z components unchanged and transform the y component', but the matrix P in Eq. (10) maps (x, y, z) to (z, x, y), which does not keep x and z unchanged. This description should be reconciled with the actual permutation.
[4.1] The sentence 'We present more visual results of the optimized mesh in Fig. 12 and numerical results in Tab. 5 and Tab. 5' references Table 5 twice; the second reference appears to be a typo for another table.
[Fig. 11] The caption 'LoD 0 LoD 1 LoD 5 LoD 3' does not appear to be in a consistent order; the LOD levels should be listed monotonically.
[3.3, Eq. (5)] The sentence introducing L_sv mentions T_k^0 as the initial tetrahedron, but the equation itself does not use T_k^0; the definition should be completed or the reference removed.
[3.4, Sec. 4.3] The statement in Sec. 4.3 that 'the primitives in our representation are always confined within the corresponding tetrahedra' is imprecise: while the Gaussian mean is a weighted combination of vertex positions (hence inside the convex hull), the 3D Gaussian's support is unbounded and the covariance after the learned rotation can extend outside the tetrahedron. Rephrase to say the mean is confined, or clarify the intended meaning.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: StructuredField's mesh and Gaussian parameters are jointly optimized against external multi-view images, and the only same-group citation (NDR) is contextual rather than load-bearing.

full rationale

The paper's core mapping Θ_k = F_r(T_k) (Eq. 1) is a representation choice, not a prediction: the tetrahedron parameters are the unknowns fitted to multi-view RGB supervision via L_render, and the rendered output of those same parameters is compared against held-out views of external benchmarks (NeRF Synthetic, Shelly, Mip-NeRF 360). No fitted quantity is later reported as an independent prediction, and no benchmark result is imported from a self-citation. The orientation-preserving homeomorphism (Sec. 3.3) is imposed as a constraint on the feasible vertex positions, and its Jacobian-positivity argument is checked locally per coupling layer; this is a constraint, not a hidden input equal to the output. The related-work citation to NDR (Cai et al. 2022) is from the same group (Juyong Zhang is a co-author), but it is only used to contextualize invertible-network deformations; the actual coupling-layer construction follows the external Real-NVP citation (Dinh et al. 2017), so the self-citation is not load-bearing. The reviewer-flagged mismatch between the continuous homeomorphism guarantee and the remaining ~1,300 inverted tetrahedra in Table 1(c) is a correctness/over-claim concern about discrete straight-edged meshes, not a circular definition or a fitted parameter renamed as a prediction; it does not raise the circularity score. The ablation and comparison tables provide external, falsifiable evidence that the method works as a representation, so the honest finding is no significant circularity.

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

The representation relies on standard learned parameters and a few domain assumptions. The load-bearing assumption is the homeomorphism guarantee, which is empirically contradicted in the ablation.

free parameters (7)

per-tetrahedron rotation quaternion q_k = learned
Eq. (3) applies a learned rotation to the PCA-derived covariance; without it rendering quality drops (Table 2, W/o Rotation).
per-vertex weights w_{k i} = learned
Used in Eqs. (2) and (4) to form the Gaussian mean, covariance, and color; ReLU-clamped.
per-tetrahedron opacity o_k = learned
Controls masking and blending.
invertible network MLP parameters = learned
Real-NVP scale and translation functions (Eq. 7) are MLPs with 2 hidden layers of 128 units.
loss weights lambda1..lambda4 = 0.8, 0.2, 0.5, 10.0
Chosen by hand in Implementation Details; trade off rendering, mask, and quality losses.
subdivision gradient threshold delta = 0.0002
Same as 3DGS; decides when to split tetrahedra.
pruning threshold epsilon = 0.05
Masks low-opacity tetrahedra.

assumptions (5)

standard math Real-NVP coupling layers are orientation-preserving diffeomorphisms
Dinh et al. 2017; used to build the homeomorphism H.
domain assumption Initial tetrahedral mesh is valid (no inversions or self-intersections)
Sec. 3.4 initialization from NeuS2/fTetWild or uniform grid; the guarantee is only as good as the input.
ad hoc to paper Positive-Jacobian global homeomorphism implies the straight-edged image mesh is also valid
Sec. 3.3; this is the core guarantee claim, but Table 1 (c) contradicts it, so the axiom is false in the discrete setting.
domain assumption The PCA-derived covariance in Eq. (2) is positive definite and captures the desired anisotropy
Assumes non-degenerate weighted vertex positions; degenerate tetrahedra would give singular covariances.
domain assumption Implicit child tetrahedra stay inside their parent and do not affect the base mesh conformality
Sec. 3.2; barycentric parameterization ensures containment by construction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of StructuredField: Unifying Structured Geometry and Radiance Field." pith.science (2026). https://pith.science/paper/FHMDP36R

@misc{pith2026250118152,
  author       = {Pith},
  title        = {Pith review of: StructuredField: Unifying Structured Geometry and Radiance Field},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FHMDP36R}},
  note         = {Machine review of arXiv:2501.18152}
}

read the original abstract

Recent point-based differentiable rendering techniques have achieved significant success in high-fidelity reconstruction and fast rendering. However, due to the unstructured nature of point-based representations, they are difficult to apply to modern graphics pipelines designed for structured meshes, as well as to a variety of simulation and editing algorithms that work well with structured mesh representations. To this end, we propose StructuredField, a novel representation that achieves both a structured geometric representation of the reconstructed object and high-fidelity rendering reconstruction. We employ structured tetrahedral meshes to represent the reconstructed object. We reparameterize the geometric attributes of these tetrahedra into the parameters of 3D Gaussian primitives, thereby enabling differentiable, high-fidelity rendering directly from the mesh. Furthermore, a hierarchical implicit subdivision strategy is utilized to ensure a conformal mesh structure while empowering the representation to capture multi-scale details. To maintain geometric integrity during optimization, we propose a novel inversion-free homeomorphism that constrains the tetrahedral mesh, guaranteeing it remains both inversion-free and self-intersection-free during the optimization process and in the final result. Based on our proposed StructuredField, we achieve high-quality structured meshes that are completely inversion-free and conformal, while also attaining reconstruction results comparable to those of 3DGS. We also demonstrate the applicability of our representation to various applications such as physical simulation, deformation, and level-of-detail.

Figures

Figures reproduced from arXiv: 2501.18152 by the authors.

**Figure 1.** StructuredField represents and reconstructs scene using a structured tetrahedral mesh. This novel structured 3D representation enables a variety of applications, including physical simulations, deformations, and more. Recent point-based differentiable rendering techniques have achieved significant success in high-fidelity reconstruction and fast rendering. However, due to the unstructured nature of point-based repr… view at source ↗

**Figure 2.** Overview of StructuredField. Given multi-view images as input, we reconstruct the 3D scene using a structured tetrahedral mesh. Our implicit multi-level subdivision module iteratively refines the initial input tetrahedra, creating a hierarchical structure. The root nodes of this hierarchy (the base tetrahedra) form a high-quality geometric foundation; their vertex positions are optimized using an orientation-preserv… view at source ↗

**Figure 3.** Data Structure of Hierarchical Implicit Subdivision. Base tetrahedra can be iteratively subdivided in a 1-to-4 manner, forming a quadtreelike hierarchy. The root node tetrahedra constitute a high-quality and conformal geometry, designated for downstream applications such as physical simulation. Leaf node tetrahedra are utilized for high-quality rendering. The positions of the implicitly defined subdivision contro… view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: Invertible neural network architecture. By applying a permutation strategy, the input vector is divided into two contiguous parts, ensuring that the resulting Jacobian matrix is upper triangular with positive diagonal elements. This strategy guarantees that the determi…

**Figure 5.** Figure 5: Two causes leading to mesh anomalies: self intersection and element inversion. A naive solution is to introduce a loss function to constrain the mesh optimization process. Locally injective mapping [Schüller et al. 2013] uses a barrier function to strictly prevent elem…

**Figure 6.** Figure 6: Visualization of tetrahedral mesh cross-sections optimized with different constraints. bonsai garden kitchen stump [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Gallery of Rendering Results on MipNeRF 360 dataset. ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: August 2025 [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

**Figure 8.** Figure 8: Our results on the test-views of Shelly dataset and NeRF Synthetic dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

**Figure 9.** Figure 9: Gallery of deformation results. Videos are provided in the supplementary materials for further visualization. Level-of-Detail. The hierarchical nature of our implicit subdivision naturally facilitates an efficient Level of Detail (LOD) rendering strategy. This allows …

**Figure 10.** Figure 10: Gallery of physical simulation results. The original shape is shown in the leftmost column, with the corresponding physical simulations displayed in the right column. Videos are provided in the supplementary materials for further visualization [PITH_FULL_IMAGE:figure…

**Figure 11.** Figure 11: Demonstration of our method’s LoD capabilities on Shelly Dataset and NeRF Synthetic Dataset. Promotion Association CAS (No. 2018495) and the Fundamental Research Funds for the Central Universities (No. WK3470000021). References Hervé Abdi and Lynne J Williams. 2010. P…

**Figure 12.** Figure 12: Gallery of mesh results on Shelly dataset and NeRF Synthetic dataset. Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. 2023. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Con…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 3 canonical work pages

[5]

ACM Transactions on Graphics (TOG) 38, 6 (2019), 1–14

Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG) 38, 6 (2019), 1–14. ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: August 2025. StructuredField • 15 Yizhou Yu, Kun Zhou, Dong Xu, Xiaohan Shi, Hujun Bao, Baining Guo, and Heung- Yeung Shum. 2004. Mesh editing with poisson-based gradient...

work page 2019
[2019]

ACM Trans

Taichi: a language for high-performance computation on spatially sparse data structures. ACM Trans. Graph. 38, 6, Article 201 (nov 2019), 16 pages. https: //doi.org/10.1145/3355089.3356506 ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: August 2025. 14 • Song et al. Yixin Hu, Teseo Schneider, Bolun Wang, Denis Zorin, and Daniele Panozzo. 202...

arXiv 2020
[2021]

Computational Visual Media 7 (2021), 289–318

Inversion-free geometric mapping construction: A survey. Computational Visual Media 7 (2021), 289–318. Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. 2010. Towards Internet-scale multi-view stereo. In 2010 IEEE Computer Society Conference on Com- puter Vision and Pattern Recognition . 1434–1441. https://doi.org/10.1109/CVPR.2010....

arXiv 2021
[2022]

In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5470–5479. ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: August 2025. StructuredField • 13 Fig. 12. Gallery of mesh results on Shelly dataset and NeRF Synthetic dataset. Jonathan T Barron, Ben ...

work page 2025
[2023]

In Proceedings of the IEEE/CVF International Conference on Computer Vision

Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 19697–19705. Jens Behrmann, Will Grathwohl, Ricky TQ Chen, David Duvenaud, and Jörn-Henrik Jacobsen. 2019. Invertible residual networks. In International conference on machine learning. PMLR, 573–582. Mario Botsch, Alexand...

work page 2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [5]

ACM Transactions on Graphics (TOG) 38, 6 (2019), 1–14

Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG) 38, 6 (2019), 1–14. ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: August 2025. StructuredField • 15 Yizhou Yu, Kun Zhou, Dong Xu, Xiaohan Shi, Hujun Bao, Baining Guo, and Heung- Yeung Shum. 2004. Mesh editing with poisson-based gradient...

work page 2019

[2] [2019]

ACM Trans

Taichi: a language for high-performance computation on spatially sparse data structures. ACM Trans. Graph. 38, 6, Article 201 (nov 2019), 16 pages. https: //doi.org/10.1145/3355089.3356506 ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: August 2025. 14 • Song et al. Yixin Hu, Teseo Schneider, Bolun Wang, Denis Zorin, and Daniele Panozzo. 202...

arXiv 2020

[3] [2021]

Computational Visual Media 7 (2021), 289–318

Inversion-free geometric mapping construction: A survey. Computational Visual Media 7 (2021), 289–318. Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. 2010. Towards Internet-scale multi-view stereo. In 2010 IEEE Computer Society Conference on Com- puter Vision and Pattern Recognition . 1434–1441. https://doi.org/10.1109/CVPR.2010....

arXiv 2021

[4] [2022]

In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5470–5479. ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: August 2025. StructuredField • 13 Fig. 12. Gallery of mesh results on Shelly dataset and NeRF Synthetic dataset. Jonathan T Barron, Ben ...

work page 2025

[5] [2023]

In Proceedings of the IEEE/CVF International Conference on Computer Vision

Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 19697–19705. Jens Behrmann, Will Grathwohl, Ricky TQ Chen, David Duvenaud, and Jörn-Henrik Jacobsen. 2019. Invertible residual networks. In International conference on machine learning. PMLR, 573–582. Mario Botsch, Alexand...

work page 2022