Pith. sign in

REVIEW 5 major objections 5 minor 40 references

DoubleDiffusion: Combining Heat Diffusion with Denoising Diffusion for Texture Generation on 3D Meshes

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

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

Pith's one-line read DoubleDiffusion generates RGB textures directly on 3D mesh surfaces by combining heat diffusion with denoising diffusion, reporting higher coverage and 8.1x faster sampling than a field-based baseline.

desk verdict Sensible engineering with released code, but the 'first' claim and the quantitative edge over MDF are both shakier than the paper admits. read the letter →

arxiv 2501.03397 v5 pith:GLKP4HUP submitted 2025-01-06 cs.CV

Xuyang Wang , Ziang Cheng , Zhenyu Li , Jiayu Yang , Haorui Ji , Pan Ji , Mehrtash Harandi , Richard Hartley
show 1 more author
Hongdong Li
This is my paper · ORCID
classification cs.CV
keywords meshtexturegenerationdenoisingdiffusionprobabilisticmodelheatLaplacian-Beltramioperatorper-vertexRGB3Dmanifoldlearningspectrallow-passfilteringgeometry-aware
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 tries to establish that denoising diffusion models can learn and generate RGB textures directly on the surface of a 3D mesh, without going through 2D images, UV unwrapping, or multi-view projection. It does this by treating per-vertex colors as a signal on the mesh manifold and letting heat diffusion act as the geometry-aware denoiser inside a standard denoising diffusion pipeline. If the claim holds, texture generation becomes native to the mesh: smooth across the surface, consistent from every viewpoint, and fast enough to handle meshes with roughly 100,000 vertices. In the reported experiments the method matches the field-based baseline on distribution distance while covering far more of the test distribution, using a smaller model and an 8.1x faster per-step sampling.

What carries the argument

The load-bearing object is the heat kernel $H(s) = \Phi e^{-s\Lambda}\Phi^T$, built from the eigenvectors $\Phi$ and eigenvalues $\Lambda$ of the cotangent-weight Laplacian of the mesh. Applied to per-vertex features, it projects the signal into a spectral domain, discounts high-frequency components through the factors $e^{-s\lambda_i}$, and thereby acts as a geometric low-pass filter that propagates color across the surface while suppressing noise. This heat-diffusion module is combined with spatial gradient operators and a timestep-aware per-vertex MLP in each DoubleDiffusion Block, and the eigenvectors are precomputed once per mesh so inference stays cheap.

What would settle it

Train DoubleDiffusion on a mesh that mixes very large flat faces with finely tessellated curved regions. If the generated texture comes out over-smoothed or blotchy precisely on the large faces while fine faces retain detail, the truncated heat kernel is not a faithful denoiser and the central claim fails; the paper's own limitation on large faces points to this test. A quantitative version is to measure per-face reconstruction error of a known test pattern after one heat-diffusion step and show it grows sharply with face area.

Watch

Extended reading notes

Core claim

The central discovery is that a denoising diffusion probabilistic model can operate directly on a fixed triangle mesh by replacing the usual image-domain noise predictor with a per-vertex network built from heat diffusion. The forward process adds Gaussian noise to per-vertex RGB values; the reverse process predicts the noise using stacked DoubleDiffusion Blocks, each combining a spectral heat-diffusion low-pass filter, spatial gradient operators, and a timestep-aware per-vertex MLP. The authors claim this is the first framework that directly learns and generates signals on the 3D mesh surface, and that because the denoiser respects mesh geometry rather than a 2D grid, the outputs are smooth along the surface and consistent across views. Supporting evidence includes a single-manifold experiment on a bunny mesh with human-face textures, where the method keeps minimum matching distance on par with the field-based baseline while raising coverage from 7.75 to 32.01, and a per-category texture experiment on a large 3D object dataset showing geometry-adapted colors on chairs and other categories.

Load-bearing premise

The central bet is that a mesh-smoothing filter built from a small set of geometry eigenvectors acts as a reliable denoiser for per-vertex colors on arbitrary triangle meshes, so that smoothing the surface does not erase the fine texture the model must generate.

Editorial extensions

If this is right

  • Texture generation no longer needs multi-view rendering or UV unwrapping, which removes the usual sources of seam and baking artifacts.
  • View consistency follows from generation on the mesh itself: a single per-vertex texture is defined for the whole surface, so there are no independent views to reconcile.
  • The same framework can model both a single fixed mesh and a whole category of different meshes, generating textures that adapt to each shape's geometry.
  • Because the model operates per vertex without global attention, it scales to roughly 100,000 vertices in a single pass, where the field-based baseline is limited to about 5,000.
  • With 2.4M parameters and 0.14s per denoising step, the method is substantially lighter and faster, making direct mesh texturing more practical for real assets.

Reading between the lines

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

  • Editorial inference: because the heat filter keeps only 128 eigenvectors, generated texture detail is capped by that spectral resolution; raising $k$ should sharpen fine features at higher computational cost, a direct testable knob.
  • Editorial inference: the same architecture should transfer to other per-vertex signals beyond RGB, such as normals, semantic labels, or physical fields, since the denoiser is agnostic to the meaning of the vertex feature.
  • Editorial inference: the per-category results suggest a category embedding or text condition could extend the model to multi-class generation once more paired shape-texture data is available, which the paper itself flags as future work.
  • Editorial inference: the headline coverage improvement is measured against an unofficial reimplementation of the baseline, so the exact margin may change if the original baseline code is released; the speed and scalability advantages are less sensitive to that comparison.
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. This paper proposes DoubleDiffusion, a generative model that produces per-vertex RGB textures directly on 3D triangle meshes. The method treats vertex colors as signals on a manifold and replaces standard convolutional layers with a heat-diffusion module inherited from DiffusionNet: a cotangent-Laplacian eigenbasis is used to implement a geometric low-pass filter, augmented by spatial-gradient operators and a timestep-aware per-vertex MLP. Denoising follows a DDPM-style objective, with the paper also referring to EDM for fast sampling. Experiments compare against a reproduction of Manifold Diffusion Fields (MDF) on the Stanford bunny with CelebA-HQ colors, reporting comparable MMD, a 312.82% relative COV improvement, 4.7x fewer parameters, and 8.1x faster sampling. A second experiment trains per-category texture generators on ShapeNetCore chairs and shows qualitative samples on chairs, pianos, and faucets. The abstract and contributions claim that DoubleDiffusion is the first DDPM framework that directly learns and generates signals on the 3D mesh surface.

Significance. If the quantitative claims hold, the contribution is practically useful: native mesh diffusion avoids multi-view inconsistency and UV baking artifacts, and the DiffusionNet-style operator is discretization-agnostic and efficient enough to process large meshes. The paper ships code, uses an established spectral operator, and frames the task cleanly. However, the load-bearing quantitative evidence is currently anchored to an unreleased, unverified MDF reproduction with a reversed metric convention, no error bars, and one dataset; the ShapeNet experiment is qualitative only. These gaps need to be addressed before the efficiency and quality claims can be accepted.

major comments (5)
  1. [§5.1.2, Table 1] The headline 312.82% coverage improvement and the 8.1x speedup are measured only against the authors' reproduction of MDF, since MDF has no official code. The paper states that it meticulously reproduces MDF and that it followed the Achlioptas metric convention "a practical contrary to that used by MDF." Because the COV/MMD definitions swap the roles of reference and generated sets, the reproduced MDF numbers may not be comparable even to MDF's published values, and the comparison is not independently anchored. Please validate the reproduction against MDF's reported numbers where possible, report both metric conventions, and release the baseline implementation and checkpoints used for Table 1.
  2. [§5.1, Table 1] No training seeds, random seeds for sampling, or error bars are reported. With 2,824 test samples and a single run, the MMD values 0.259 versus 0.284 are close enough that the "comparable MMD" statement could change under resampling, and the COV gap of 7.75 versus 32.01 needs a confidence interval. Please report means and standard deviations over at least three training runs and bootstrap confidence intervals for COV and MMD.
  3. [§4.3 and §5.1.1] Equations (7)–(9) define a standard DDPM with a linear noise scheduler, but Section 5.1.1 says the network employs the EDM framework with 18 timesteps. EDM uses a different noise-level parameterization, preconditioning, and sampling schedule, so as written the method is under-specified and the 8.1x speedup cannot be reproduced from the equations. Please present the exact training and sampling procedures used, or align the equations with the actual EDM formulation.
  4. [§5.2, §5.1.3] The ShapeNetCore experiment is evaluated only qualitatively (Figs. 6–8), despite being listed as evidence for the framework's versatility, and the contribution text claims processing of roughly 100k vertices although the largest reported mesh has about 52k vertices. Please add quantitative metrics for per-category texture generation (for example, fidelity/coverage metrics or a user study) and either demonstrate the 100k-vertex case or qualify the claim.
  5. [§2.3 and Abstract] The abstract and contributions call DoubleDiffusion "the first framework for denoising diffusion probabilistic model that directly learns and generates signals on the 3D mesh surface," but the paper cites Foti et al. [8], "UV-free texture generation with denoising and geodesic heat diffusion," and does not explain how the two approaches differ. The related-work text groups [8] with methods that "apply positional encoding to sampled points," which is not supported by the cited title or by any discussion in the paper. Please provide a direct comparison with [8] and, if appropriate, revise the novelty claim.
minor comments (5)
  1. [Figs. 3–4, §4.2.2, §5.1.2, §5.1.3] There are several typos: "spacial gradient" should be "spatial gradient", "a practical contrary" should be "a practice contrary", "modal" should be "model", and "CelebaA-HQ" should be "CelebA-HQ".
  2. [§4.2.1, Eqs. (4)–(6)] Please state whether the eigendecomposition is the generalized Laplacian eigenproblem with the mass matrix and how the cotangent Laplacian is normalized before truncation; the equations as written use an ordinary transpose, which is ambiguous for a mass-weighted discrete Laplace-Beltrami operator.
  3. [Fig. 5, §5.1.3] The 25k and 52k vertex results are shown without quantitative metrics or a comparison baseline; please add measurements or clarify that these are demonstrations of scalability only.
  4. [§5.1.3] The claim that MDF cannot handle more than roughly 5k vertices is not backed by memory or runtime measurements; adding such measurements would make the scalability argument more convincing.
  5. [References] References [20] and [21] are the same Rombach et al. paper, and [27] and [28] duplicate the LGM citation; please consolidate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DoubleDiffusion's denoising loss is the standard DDPM objective, its heat-diffusion operator is inherited from external work (DiffusionNet), and its headline comparisons are against an externally published baseline, so no prediction reduces to a fitted value or to the authors' own prior results.

full rationale

The paper's derivation chain is self-contained against external benchmarks. The training objective (Eq. 9) is exactly the standard DDPM MSE loss between true and predicted noise; the target is the additive noise, not a parameter fitted to the evaluation data, so no fitted quantity is later renamed as a prediction. The heat-diffusion module (Eqs. 3-6) is adopted from DiffusionNet (Ref. 23), an external prior work, and is used as a geometric feature propagation operator; the paper does not need its own conclusions to justify this operator. The claim of directly learning and generating signals on the mesh surface follows from the architecture and training procedure rather than from a definition that presupposes the result. The quantitative claims (312.82% COV improvement, 4.7x smaller model, 8.1x faster sampling) are measured against a reproduction of MDF (Sec. 5.1.2), an external baseline, and the paper explicitly notes that it follows the Achlioptas metric convention 'a practical contrary to that used by MDF'; while this raises a legitimate reproducibility and comparability concern, it is not circularity because the baseline is external and no equation in the paper forces the outcome to equal the input. The limitation acknowledging difficulties with large faces is an honest scope statement, not a hidden premise. No load-bearing self-citation was found, and no step in the claimed derivation can be exhibited as equivalent to its inputs by construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its free parameters are standard hyperparameters and learned scales. The main axiomatic load is the heuristic that heat diffusion works as a denoiser for mesh signals, which is inherited from DiffusionNet.

free parameters (4)
  • heat diffusion timescale s = learned per feature channel
    Controls spectral attenuation in Eq. 6; learned during training, so the denoising behavior is fit to data rather than derived.
  • number of Laplacian eigenvectors k = 128
    Truncation of the spectral basis used for heat diffusion (Sec. 5.1.1); chosen by hand and affects quality, speed, and memory.
  • denoising timesteps (EDM) = 18
    Inference schedule from EDM [14]; a hyperparameter that trades quality for speed and is not derived.
  • number of DoubleDiffusion blocks = 8
    Network depth selected for the experiments (Sec. 5.1.1); architecture capacity is a free design choice.
assumptions (4)
  • standard math The cotangent-weight Laplacian (Eq. 2) is a valid discrete Laplace-Beltrami operator for triangle meshes.
    Used throughout Sec. 4.2; standard in mesh processing [17,25].
  • standard math The heat kernel can be approximated by truncating the eigen-decomposition of the Laplacian to the k smallest eigenvectors (Eqs. 4-6).
    Standard spectral method from DiffusionNet [23]; requires the mesh Laplacian to be symmetric and the eigenbasis to capture geometry.
  • domain assumption Texture generation can be treated as denoising per-vertex RGB signals on a fixed mesh.
    Frames the task; assumes vertex colors are a sufficient representation of texture and that the mesh geometry is fixed and known.
  • ad hoc to paper Heat diffusion acts as an effective denoiser inside a DDPM because it suppresses high-frequency noise.
    Central heuristic of Sec. 4.2.1; not proven, and the paper's Limitation section notes difficulty with large faces.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DoubleDiffusion: Combining Heat Diffusion with Denoising Diffusion for Texture Generation on 3D Meshes." pith.science (2026). https://pith.science/paper/GLKP4HUP

@misc{pith2026250103397,
  author       = {Pith},
  title        = {Pith review of: DoubleDiffusion: Combining Heat Diffusion with Denoising Diffusion for Texture Generation on 3D Meshes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GLKP4HUP}},
  note         = {Machine review of arXiv:2501.03397}
}
read the original abstract

This paper addresses the problem of generating textures for 3D mesh assets. Existing approaches often rely on image diffusion models to generate multi-view image observations, which are then transformed onto the mesh surface to produce a single texture. However, due to the gap between multi-view images and 3D space, such process is susceptible to arange of issues such as geometric inconsistencies, visibility occlusion, and baking artifacts. To overcome this problem, we propose a novel approach that directly generates texture on 3D meshes. Our approach leverages heat dissipation diffusion, which serves as an efficient operator that propagates features on the geometric surface of a mesh, while remaining insensitive to the specific layout of the wireframe. By integrating this technique into a generative diffusion pipeline, we significantly improve the efficiency of texture generation compared to existing texture generation methods. We term our approach DoubleDiffusion, as it combines heat dissipation diffusion with denoising diffusion to enable native generative learning on 3D mesh surfaces.

Figures

Figures reproduced from arXiv: 2501.03397 by the authors.

Figure 1
Figure 1. Overview of the DoubleDiffusion. texture directly on a given mesh has become an essential approach. Due to the lack of a unified 3D foundation model for meshes, some prior works [5, 35] unwrap the surface into 2D texture maps and harness powerful image priors such as Stable Diffusion [20]. Although UV maps lie on a 2D plane, they are having geometric distortion and the texture cannot be generated directly via a 2D d… view at source ↗
Figure 2
Figure 2. Illustration of the cotangent weighted differential coor [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Method Overview. Our (a) DoubleDiffusion Network contains consecutive (b) DoubleDiffusion Blocks (DDB). Each (c) DDB consists of the Heat Diffusion module aggregating vertice features over the spatial domain on mesh surfaces and the Timestep-Aware MLP that injects timestep embedding in a residual manner. the h in Eq. 1, in our task, f are initially sampled from the Gaussian Noise fT ∼ N (0, 1). These noisy vertex co… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Overview of the DoubleDiffusion Block, containing (a) Geometric-aware Low Pass Filter; (b) Spacial gradient Feature and (c) [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison between the (a) MDF (baseline) [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Texture learning across multiple chair shapes in the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Generative result of the textured chairs from five distinct [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 24 canonical work pages

  1. [8]

    Uv-free texture generation with denoising and geodesic heat diffu- sion

    Simone Foti, Stefanos Zafeiriou, and Tolga Birdal. Uv-free texture generation with denoising and geodesic heat diffu- sion. Advances in Neural Information Processing Systems , 37:128053–128081, 2025. 2

    work page 2025

  2. [18]

    Single mesh diffusion models with field latents for texture generation

    Thomas W Mitchel, Carlos Esteves, and Ameesh Makadia. Single mesh diffusion models with field latents for texture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7953– 7963, 2024. 7

    work page 2024

  3. [1]

    Learning representations and generative models for 3d point clouds

    Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40–49. PMLR, 2018. 6

    work page 2018

  4. [2]

    Layer normalization

    Jimmy Lei Ba. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5

    arXiv 2016

  5. [3]

    Meta 3d texturegen: Fast and consistent texture generation for 3d objects

    Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 1, 2

    arXiv 2024

  6. [4]

    Shapenet: An information-rich 3d model repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 7, 8

    arXiv 2015

  7. [5]

    Text2tex: Text-driven texture synthesis via diffusion models

    Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 1

    arXiv 2023

  8. [6]

    Flashtex: Fast relightable mesh texturing with lightcontrolnet

    Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. Flashtex: Fast relightable mesh texturing with lightcontrolnet. arXiv preprint arXiv:2402.13251, 2024. 2

    arXiv 2024

Show all 40 references
  1. [7]

    Manifold diffusion fields

    Ahmed AA Elhag, Yuyang Wang, Joshua M Susskind, and Miguel ´Angel Bautista. Manifold diffusion fields. In The Twelfth International Conference on Learning Representa- tions. 2, 3, 6, 7

  2. [9]

    Im- proving neural network surface processing with principal curvatures

    Josquin Harrison, James Benn, and Maxime Sermesant. Im- proving neural network surface processing with principal curvatures. Advances in Neural Information Processing Sys- tems, 37:122384–122405, 2024. 4

    2024

  3. [10]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 5

    2020

  4. [11]

    Man- ifoldplus: A robust and scalable watertight manifold sur- face generation method for triangle soups

    Jingwei Huang, Yichao Zhou, and Leonidas Guibas. Man- ifoldplus: A robust and scalable watertight manifold sur- face generation method for triangle soups. arXiv preprint arXiv:2005.11621, 2020. 6

    2005 arXiv

  5. [12]

    Perceiver io: A general architecture for structured inputs & outputs

    Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Kop- pula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 7

    2021 arXiv

  6. [13]

    Progressive growing of gans for improved qual- ity, stability, and variation.arXiv preprint arXiv:1710.10196,

    Tero Karras. Progressive growing of gans for improved qual- ity, stability, and variation.arXiv preprint arXiv:1710.10196,

    arXiv

  7. [14]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565–26577, 2022. 6

    2022

  8. [15]

    Laplace-beltrami eigenfunctions towards an algorithm that” understands” geometry

    Bruno L ´evy. Laplace-beltrami eigenfunctions towards an algorithm that” understands” geometry. In IEEE Interna- tional Conference on Shape Modeling and Applications 2006 (SMI’06), pages 13–13. IEEE, 2006. 4

    2006

  9. [16]

    Latent-nerf for shape-guided generation of 3d shapes and textures

    Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 12663–12673, 2023. 1

    2023

  10. [17]

    Discrete differential-geometry operators for triangu- lated 2-manifolds

    Mark Meyer, Mathieu Desbrun, Peter Schr ¨oder, and Alan H Barr. Discrete differential-geometry operators for triangu- lated 2-manifolds. In Visualization and mathematics III , pages 35–57. Springer, 2003. 3

    2003

  11. [19]

    Dreamfusion: Text-to-3d using 2d diffusion

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representa- tions. 1

  12. [20]

    High-resolution image syn- thesis with latent diffusion models, 2021

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models, 2021. 1

    2021

  13. [21]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2

    2022

  14. [22]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 5

    2022

  15. [23]

    Diffusionnet: Discretization agnostic learning on surfaces

    Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM Transactions on Graphics (TOG), 41(3): 1–16, 2022. 2, 3, 4, 5, 6

    2022

  16. [24]

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

    Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d gen- eration. In The Twelfth International Conference on Learn- ing Representations. 1

  17. [25]

    Laplacian mesh processing

    Olga Sorkine. Laplacian mesh processing. Eurographics (State of the Art Reports), 4(4):1, 2005. 2 9

    2005

  18. [26]

    PyVista: 3D plot- ting and mesh analysis through a streamlined interface for the Visualization Toolkit (VTK)

    Bane Sullivan and Alexander Kaszynski. PyVista: 3D plot- ting and mesh analysis through a streamlined interface for the Visualization Toolkit (VTK). Journal of Open Source Software, 4(37):1450, 2019. 6

    2019

  19. [27]

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

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 1

    2024

  20. [28]

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

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1–18. Springer, 2025. 2

    2025

  21. [29]

    Mvdiffusion++: A dense high- resolution multi-view diffusion model for single or sparse- view 3d object reconstruction

    Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Fu- rukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high- resolution multi-view diffusion model for single or sparse- view 3d object reconstruction. In European Conferen...

    2024

  22. [30]

    A signal processing approach to fair surface design

    Gabriel Taubin. A signal processing approach to fair surface design. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques , pages 351– 358, 1995. 2, 4

    1995

  23. [31]

    The stanford 3d scanning repos- itory, 1994

    Greg Turk and Marc Levoy. The stanford 3d scanning repos- itory, 1994. 5

    1994

  24. [32]

    Crm: Single image to 3d textured mesh with convolutional reconstruction model

    Zhengyi Wang, Yikai Wang, Yifei Chen, Chendong Xi- ang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 2

    2024 arXiv

  25. [33]

    Group normalization

    Yuxin Wu and Kaiming He. Group normalization. In Pro- ceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5

    2018

  26. [34]

    Graph convolutional networks using heat kernel for semi-supervised learning

    Bingbing Xu, Huawei Shen, Qi Cao, Keting Cen, and Xueqi Cheng. Graph convolutional networks using heat kernel for semi-supervised learning. arXiv preprint arXiv:2007.16002,

    2007 arXiv

  27. [35]

    Texture generation on 3d meshes with point- uv diffusion

    Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with point- uv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206–4216, 2023. 1

    2023

  28. [36]

    Spectral mesh processing

    Hao Zhang, Oliver Van Kaick, and Ramsay Dyer. Spectral mesh processing. In Computer graphics forum, pages 1865–

  29. [37]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836–3847, 2023. 2

    2023

  30. [38]

    Clay: A controllable large-scale generative model for creat- ing high-quality 3d assets

    Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creat- ing high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1–20, 2024. 2

    2024

  31. [39]

    Diffusion probabilistic fields

    Peiye Zhuang, Samira Abnar, Jiatao Gu, Alex Schwing, Joshua M Susskind, and Miguel Angel Bautista. Diffusion probabilistic fields. In The Eleventh International Confer- ence on Learning Representations, 2023. 3, 7 10

    2023

  32. [1894]

    Wiley Online Library, 2010. 2

    2010

Pith tools

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