Pith. sign in

REVIEW 3 major objections 4 minor 53 references

DiffGI: Differentiable Geometry Images for High-Fidelity Thin-Shell 3D Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read DiffGI claims that replacing binary occupancy with a continuous 2D truncated signed distance function and making Marching Squares differentiable allows thin-shell, open-boundary 3D surfaces to be reconstructed and generated end to end from

desk verdict Solid, practical paper: continuous 2D TSDF in geometry images with differentiable Marching Squares works and is well evidenced, but the position-map background convention is under-specified. read the letter →

arxiv 2607.13365 v1 pith:QHLQIUUP submitted 2026-07-15 cs.CV

Eungjune Shim , Hansol Lee , Eunjung Ju This is my paper
classification cs.CV
keywords 3Dgenerationgeometryimagestruncatedsigneddistancefunctiondifferentiablemarchingsquaresthin-shelllatentdiffusionnon-manifoldsurfacesboundaryprecision
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a geometry image—a 2D grid storing a 3D surface—can represent thin-shell and open-boundary objects such as garments far more faithfully if its binary occupancy channel is replaced with a continuous 2D signed distance field, and that the mesh-extraction step can be made differentiable so that 3D shape losses train the whole pipeline end to end. Binary occupancy maps encode boundaries at the pixel grid's resolution, so downsampling destroys thin details and creates staircase artifacts; a truncated signed distance function (TSDF) instead encodes where the boundary lies between pixels. The paper introduces Differentiable Marching Squares, which computes zero-crossing vertices by linear interpolation of the TSDF and samples 3D coordinates from the position map, making reconstruction backpropagatable. On garment and furniture datasets, this yields sharper boundaries and better preservation of thin shells than prior geometry-image and voxel approaches, with a VAE compressing 256×256 tensors into a 32×32×4 latent space that supports real-time conditional generation. A sympathetic reader would care because it suggests that compact, UV-friendly, open-boundary 3D generation is not limited to expensive volumetric representations.

What carries the argument

The central object is a four-channel geometry-image tensor: a three-channel position map plus a one-channel 2D TSDF. The TSDF stores the signed pixel distance to the nearest chart contour, truncated at 15 pixels, so boundary locations survive aggressive downsampling. Differentiable Marching Squares (DMS) treats each boundary vertex as the linearly interpolated zero crossing of the TSDF between opposite-signed pixels, so vertex coordinates are continuous functions of the field; bilinear sampling of the position map turns each 2D vertex into 3D coordinates. A geometry-aware normal rendering loss on the DMS-reconstructed mesh provides 3D supervision through the whole pipeline.

What would settle it

Inspect a single garment chart's TSDF and position map before and after downsampling: read the position-map values in the dilated background ring around the chart, and compare DMS-reconstructed boundary coordinates against ground-truth boundaries for cases where the zero crossing falls in the dilated region. If the background position is zero, reconstructed boundary vertices will be biased toward the global origin, measurable as a jump in Boundary Chamfer Distance.

Watch

Extended reading notes

Core claim

The central claim is that replacing the binary occupancy channel of a geometry image with a continuous 2D truncated signed distance function, combined with a differentiable Marching Squares extractor, lets boundary position be encoded at subpixel precision in a fixed-resolution grid and lets 3D surface losses backpropagate to the 2D latent. On garment and furniture data, the authors show that this representation compresses into a 32×32×4 latent and still reconstructs and generates thin, open-boundary, non-manifold surfaces with sharper boundaries than occupancy-based geometry images and voxel-based methods, while running in about a second on a consumer GPU.

Load-bearing premise

The 3D coordinates of boundary vertices are obtained by bilinear sampling of the 3-channel position map at the TSDF zero-crossing, but the paper never states what value the position map holds in the background outside charts; if that value is zero, boundary vertices will be dragged toward the origin and the reported boundary precision cannot hold (Sections 3.1 Steps 3–5 and 3.2).

Editorial extensions

If this is right

  • Boundary quality stops depending on grid resolution: the continuous TSDF preserves subpixel boundary position, so low-resolution tensors outperform binary occupancy at the same resolution and match high-resolution occupancy near 256.
  • End-to-end training becomes possible: because DMS is differentiable, normal-rendering and other 3D surface losses can be backpropagated through reconstruction to the VAE encoder and decoder, letting compression focus on high-frequency geometry.
  • Aggressive compression works: a 256×256×4 geometry image can be reduced to a 32×32×4 latent with reconstruction metrics better than those of uncompressed occupancy-based geometry images.
  • Generation becomes cheap: a transformer-based latent diffusion model with flow matching on this space generates 3D shapes in about 1.2 seconds on a consumer GPU (3.22 GB VRAM) and about 8.5 seconds on CPU, versus tens of seconds and much higher memory for volumetric baselines.
  • Thin and open-boundary shapes become obtainable: the representation handles non-manifold garment surfaces and open boundaries that watertight implicit methods fail on.

Reading between the lines

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

  • Inference: The paper leaves unspecified what value the position map takes in the background outside the charts; if that value is not the dilated edge value, zero-crossing vertices near chart borders will be pulled toward the origin, making the reported boundary fidelity depend on an undocumented implementation detail (Section 3.1 Steps 3–5 and Section 3.2).
  • Inference: The TSDF advantage is boundary-complexity dependent—on flat, straight-edged furniture the gap over occupancy shrinks, as the paper's own resolution analysis notes, so the subpixel-precision benefit matters most for wrinkle-dense garments.
  • Inference: Because each UV chart is reconstructed independently and ambiguous saddle cells are always split, visible seams appear across chart boundaries; enforcing cross-chart consistency would likely be needed before the output feeds physics simulation.
  • Inference: A natural testable extension is to apply the same continuous-field-plus-differentiable-extraction recipe to signed distance fields on other 2D parameterizations and to dual-contouring variants for sharp mechanical edges, as the paper lists as its own future work.
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

3 major / 4 minor

Summary. The paper proposes DiffGI, a geometry-image representation for 3D surfaces that replaces binary occupancy maps with a continuous 2D truncated signed distance function (TSDF), and a Differentiable Marching Squares (DMS) module that reconstructs a 3D mesh from the TSDF and 3-channel position map in a fully differentiable manner. The authors train a VAE with pixel-space L1 losses and a geometry-aware normal rendering loss to compress 256×256×4 geometry images into a 32×32×4 latent space, then train a transformer-based latent diffusion model for conditional generation. The main claims are superior reconstruction fidelity and boundary precision compared to occupancy-based geometry-image baselines, with significantly lower compute.

Significance. If the claims hold, the paper makes a useful contribution: it brings differentiable iso-surface extraction to the 2D multi-chart geometry-image setting, enabling surface-level supervision for thin-shell and non-manifold geometry. The resolution ablation in the supplementary material (Fig. S1) is a genuine strength and directly supports the TSDF-vs-occupancy advantage at low resolution. The method's efficiency (Table 3) and the reported reconstruction metrics on GarmageSet are also notable. However, the central boundary-precision claim depends on an under-specified detail in the position-map background convention, and the paper's own limitation section admits chart seams that are claimed not to affect metrics. The comparison with GarmageNet is partly run under a non-native tessellation, which complicates interpretation. These issues are addressable but require revision.

major comments (3)
  1. [Sec. 3.1 Step 3 and Sec. 3.2, Eq. (1)] The DMS module computes boundary vertex 3D coordinates by bilinear sampling the 3-channel position map at TSDF zero crossings. The paper states only that 'edge pixel values are propagated outward into the undefined background via dilation' (Step 3), with no fill radius or interpolation scheme. If the background position map is zero, boundary vertices will be pulled toward the origin and the reported BCD/HD gains would not hold. If it is filled with nearest-edge values, the bilinear sample across the boundary is not obviously the true 3D boundary position. This is load-bearing for the central boundary-precision claim (Tables 1, 5). Please specify the exact dilation radius and the value assigned to background pixels, and provide an ablation comparing zero-filled vs. dilated backgrounds on boundary metrics for an uncorrupted ground-truth TSDF.
  2. [Sec. 5, Limitations] The limitation section states that visible seams where neighboring charts meet 'do not affect our reconstruction and generation metrics.' This is not supported and is in tension with the use of BCD as a boundary-quality metric: chart seams create artificial open boundaries that may be included in boundary-point sampling. Please provide quantitative evidence, e.g., measure seam-induced gaps and report metrics with seams removed or repaired, or otherwise justify why BCD is insensitive to seams.
  3. [Table 1 / Sec. 4.2] GarmageNet is evaluated under the Omages-style tessellation for the main comparison, with its native official extraction reported only as an additional row. The main text says DiffGI remains superior under both settings, but the Tess. row for GarmageNet has JSD=32.61×10^-3, far worse than its official extraction (5.51×10^-3), suggesting the non-native tessellation may disadvantage GarmageNet. Since GarmageNet's representation is per-panel and not designed for Omages-style tessellation, this comparison conflates extraction method with representation. Please make the official extraction the primary comparison, or provide details on how the Omages-style tessellation was applied and justify its appropriateness.
minor comments (4)
  1. [Abstract / Sec. 3.2] The phrase 'fully differentiable mesh reconstruction' overstates the situation: the saddle-point topology choice is discrete, and differentiability holds only within a fixed topology. Please clarify.
  2. [Sec. 3.2] The complexity comparison 'O(N^2) vs O(N^3)' should state that N is the grid resolution along one dimension; as written it is ambiguous.
  3. [Table 1] On ABO, NC for DiffGI (0.83) is lower than Omages (0.89). The text attributes this to compression, but the headline claim of 'superior reconstruction fidelity' would benefit from acknowledging this trade-off explicitly in the main comparison.
  4. [References] References [13] and [14] appear to describe the same WARDROBE dataset; please consolidate or distinguish them.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: DiffGI's central claims are supported by external evaluation metrics and controlled ablations, not by construction or self-citation.

full rationale

The paper is a learned system rather than a formal derivation, and its claimed results are verified against ground-truth meshes with external metrics (CD, EMD, JSD, NC, BCD, HD, F1, P-FID/P-KID). The training losses (L_Pos, L_TSDF, L_Normal, L_KL) are optimized on source data, while the evaluation metrics are computed independently on reconstructed/generated meshes, so there is no fitted-parameter-called-prediction collapse. The Differentiable Marching Squares zero-crossing formula (Eq. 1) is a continuous linear-interpolation rule that defines subpixel positions by construction; this is a design property of the representation, not a circular prediction of an independent empirical claim. The normal-rendering loss (Eq. 3) backpropagates through the same DMS used at inference, but this is a coherent end-to-end training loop rather than a reduction to the method's own outputs. No load-bearing self-citation exists: the cited components (SD1.5, DiT, DINOv2, nvdiffrast, Omages, GarmageNet, DMTet, FlexiCubes, etc.) are external, and the authors do not cite their own prior work as the basis for the core claims. The paper also includes honest limitations (localized rounding on sharp edges, chart-seam discontinuities) and a from-scratch VAE initialization study (Table S1) showing the main gains do not depend on pretrained weights. The reader's concern about the unspecified background value of the position map during dilation (Sec. 3.1 Step 3 and Sec. 3.2) is a legitimate reproducibility/implementation detail, but it does not make any claimed result equivalent to its inputs; therefore it does not constitute circularity under the stated criteria.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The system builds on well-known machine-learning infrastructure (nvdiffrast, SD-VAE init, DINOv2, DiT) and introduces no new physical entities. It does rely on several hand-chosen constants and on unstated assumptions about the position map outside chart regions; these are the items that would need to be pinned down to reproduce the boundary-precision results.

free parameters (4)
  • TSDF truncation distance = 15 pixels
    Clamps the signed distance in the 2D UV plane; controls how much boundary information is retained after downsampling. Chosen by hand (Sec 3.1 Step 4, S8).
  • DMS regularization epsilon = 1e-5
    Added to the denominator of the linear-interpolation formula to prevent gradient explosion; shifts zero-crossing positions slightly. Eq. (1).
  • Loss weights λ_TSDF, λ_Normal, λ_KL = not reported
    Balance pixel-space losses, normal-rendering loss, and KL divergence in Eq. (2); values are not given, and the ablation results depend on them.
  • Operating resolution / latent size = 256×256×4 input, 32×32×4 latent
    Design choices that set the compression ratio and were selected to trade fidelity against efficiency; not derived.
assumptions (5)
  • domain assumption The 2D UV-plane signed distance to the chart contour is a faithful surrogate for the 3D open boundary position, including after bilinear downsampling.
    The entire subpixel-precision claim rests on the boundary in UV space corresponding to the 3D surface boundary; this is not always true under arbitrary UV packing and distortion. Invoked in Sec 3.1 Steps 1,4,5.
  • domain assumption Bilinear sampling of the position map at TSDF zero-crossings yields correct 3D coordinates, even for vertices that lie on or outside the chart mask.
    The position map is undefined in the background, yet DMS samples it at boundary vertices. The paper does not specify the background values; the reconstruction depends on this. Sec 3.2, bilinear grid sampling.
  • domain assumption nvdiffrast rasterization provides reliable normal maps for non-manifold, open-boundary meshes with two-sided surfaces.
    The normal rendering loss compares rendered normals of reconstructed and ground-truth garments; this requires a consistent notion of front/back facing on non-watertight geometry. Sec 3.3.
  • ad hoc to paper The deterministic saddle-point convention (always treating diagonal patches as separate) yields a valid topology for the target surfaces.
    In ambiguous Marching Squares cases the paper chooses one of two valid interpretations; this can split bridge-like structures and is justified only by rarity at 256×256. Sec 3.2 and S4.
  • standard math Intermediate Value Theorem and linear interpolation locate zero crossings between grid corners of opposite sign.
    Standard, non-controversial. Eq. (1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffGI: Differentiable Geometry Images for High-Fidelity Thin-Shell 3D Generation." pith.science (2026). https://pith.science/paper/QHLQIUUP

@misc{pith2026260713365,
  author       = {Pith},
  title        = {Pith review of: DiffGI: Differentiable Geometry Images for High-Fidelity Thin-Shell 3D Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QHLQIUUP}},
  note         = {Machine review of arXiv:2607.13365}
}
read the original abstract

Existing 3D generative models predominantly rely on implicit volumetric representations, which enforce watertight topology and struggle to represent thin-shell and non-manifold geometries such as garments. Geometry image-based approaches offer a surface-centric alternative, but existing methods rely on discrete binary occupancy maps whose resolution-dependent boundary encoding causes staircase artifacts and information loss upon downsampling, while surface reconstruction remains a non-differentiable post-processing step disconnected from the learning pipeline. To address this, we propose Differentiable Geometry Image (DiffGI), an end-to-end 3D-to-2D mapping framework that seamlessly integrates surface representation and geometric optimization. DiffGI replaces binary maps with a continuous 2D Truncated Signed Distance Function (TSDF), which encodes boundary position at subpixel precision within a fixed grid resolution, eliminating resolution-dependent staircase artifacts even under aggressive downsampling. Building on this continuous field, we introduce a differentiable Marching Squares algorithm based on analytical linear interpolation, allowing gradients from 3D surface losses to propagate back to the 2D latent space. Leveraging this differentiable pipeline, we train a DiffGI-VAE augmented with a geometry-aware normal rendering loss to compress complex 3D surfaces into an ultra-compact 32X32 latent space, and instantiate a transformer-based latent diffusion model with a flow-matching objective on top of this space for conditional 3D generation. Extensive experiments on garment and object datasets demonstrate that our method achieves superior reconstruction fidelity and boundary precision compared to prior geometry-image and voxel-based approaches, while requiring significantly fewer computational resources.

Figures

Figures reproduced from arXiv: 2607.13365 by the authors.

Figure 1
Figure 1. VAE reconstruction results with our TSDF-based DiffGI representation and normal rendering loss (left) versus an occupancy-based geometry image without normal loss (right), showing noticeably sharper boundaries and better preservation of thin-shell structures. 2 Related Work 2.1 Implicit and Explicit Representations for 3D Generation Recent 3D generation has primarily built on implicit field representations— signed/u… view at source ↗
Figure 2
Figure 2. Overview of the proposed DiffGI framework. The input 3D mesh is first mapped to a 2D TSDF-based DiffGI representation, which is then encoded by a DiffGI-VAE into a compact latent space. The decoder reconstructs the TSDF geometry image, from which a 3D triangle mesh is recovered via Differentiable Marching Squares. Pixel￾space losses on the TSDF and position maps, together with a geometry-aware normal rendering loss,… view at source ↗
Figure 3
Figure 3. Qualitative comparison of VAE reconstructions on the ABO and GarmageSet datasets, illustrating that our DiffGI-VAE yields sharper boundaries and better preser￾vation of thin-shell details than Omages and GarmageNet. 4.2 Reconstruction Fidelity (DiffGI-VAE) We compare DiffGI-VAE against Omages and GarmageNet on the ABO and Gar￾mageSet datasets. Omages directly converts meshes to geometry images without VAE compressio… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative ablation study of our VAE on the GarmageSet dataset. The visual comparisons demonstrate how different representations (Occ. vs. TSDF) and the nor￾mal rendering loss affect the 3D reconstruction quality [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Qualitative label-conditioned generation results on the ABO dataset. We com￾pare Omages and our DiffGI-based diffusion model, showing improved reconstruction of thin frames and open-boundary structures, with fewer staircase artifacts [PITH_FULL_IMAGE:figures/full_fig_…
Figure 6
Figure 6. Figure 6: Qualitative single-view image-to-3D comparison on GarmageSet. From left to right, each row shows the input front-view rendered normal map, the results of TRELLIS, TRELLIS.2, and GarmageNet, our DiffGI generation, and the ground-truth mesh. DiffGI preserves thin-shell d…
Figure 7
Figure 7. Figure 7: Occupancy-conditioned garment generation and local edit propagation. Given a 2D occupancy map as input, modifying only the sleeve region in 2D leads to consistent updates of the corresponding 3D geometry, demonstrating that local pattern edits are faithfully reflected …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 2 linked inside Pith

  1. [1]

    ACM Transactions on Graphics (2025)

    Binninger, A., Wiersma, R., Herholz, P., Sorkine-Hornung, O.: TetWeave: Isosur- face extraction using on-the-fly delaunay tetrahedral grids for gradient-based mesh optimization. ACM Transactions on Graphics (2025)

    2025

  2. [2]

    In: CVPR (2025)

    Boss, M., Huang, Z., Vasishta, A., Jampani, V.: SF3D: Stable fast 3D mesh re- construction with UV-unwrapping and illumination disentanglement. In: CVPR (2025)

    2025

  3. [3]

    In: Symposium on Geometry Processing (2006)

    Carr, N.A., Hoberock, J., Crane, K., Hart, J.C.: Rectangular multi-chart geometry images. In: Symposium on Geometry Processing (2006)

    2006

  4. [4]

    In: ICCV (2023)

    Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion NeRF: A unified approach to 3D generation and reconstruction. In: ICCV (2023)

    2023

  5. [5]

    In: CVPR (2022)

    Chen, W., Lin, C., Li, W., Yang, B.: 3PSDF: Three-pole signed distance function for learning surfaces with arbitrary topologies. In: CVPR (2022)

    2022

  6. [6]

    ACM Transactions on Graphics (2022)

    Chen, Z., Tagliasacchi, A., Funkhouser, T., Zhang, H.: Neural dual contouring. ACM Transactions on Graphics (2022)

    2022

  7. [7]

    ACM Transactions on Graphics (2021)

    Chen, Z., Zhang, H.: Neural marching cubes. ACM Transactions on Graphics (2021)

    2021

  8. [8]

    In: CVPR (2023)

    Cheng, Y.C., Lee, H.Y., Tulyakov, S., Schwing, A.G., Gui, L.Y.: SDFusion: Multi- modal 3D shape completion, reconstruction, and generation. In: CVPR (2023)

    2023

Show all 53 references
  1. [9]

    In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV)

    Chou, G., Bahat, Y., Heide, F.: Diffusion-SDF: Conditional generative modeling of signed distance functions. In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV). pp. 2262–2272 (2023)

    2023

  2. [10]

    In: CVPR (2022)

    Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., et al.: ABO: Dataset and benchmarks for real-world 3D object understanding. In: CVPR (2022)

    2022

  3. [11]

    In: International Con- ference on Learning Representations (ICLR) (2025)

    Elizarov, S., Rowles, C., Donné, S.: Geometry image diffusion: Fast and data- efficient text-to-3D with image-based surface representation. In: International Con- ference on Learning Representations (ICLR) (2025)

    2025

  4. [12]

    In: CVPR

    Fainstein, M., Siless, V., Iarussi, E.: DUDF: Differentiable unsigned distance fields with hyperbolic scaling. In: CVPR. pp. 4484–4493 (2024)

    2024

  5. [13]

    freshersstaff: Wardrobe vision dataset.https : / / www . kaggle . com / datasets / freshersstaff/wardrobe-vision-dataset(2025), accessed: 2025-12-01

    2025

  6. [14]

    freshersstaff: Wardrobe vision dataset.https : / / www . kaggle . com / datasets / freshersstaff/wardrobe-vision-dataset(2025), kaggle dataset. License: CC0 Public Domain. Accessed: 2025-12-01

    2025

  7. [15]

    In: ICLR (2022)

    Gu, J., Liu, L., Wang, P., Theobalt, C.: StyleNeRF: A style-based 3D-aware gen- erator for high-resolution image synthesis. In: ICLR (2022)

    2022

  8. [16]

    ACM Transactions on Graphics (2002)

    Gu, X., Gortler, S.J., Hoppe, H.: Geometry images. ACM Transactions on Graphics (2002)

    2002

  9. [17]

    In: ICLR (2024)

    Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: LRM: Large reconstruction model for single image to 3D. In: ICLR (2024)

    2024

  10. [18]

    ACM Transactions on Graphics (2023)

    Hou, F., Chen, X., Wang, W., Qin, H., He, Y.: Robust zero level-set extraction from unsigned distance fields based on double covering. ACM Transactions on Graphics (2023)

    2023

  11. [19]

    arXiv:2011.03277 (2020)

    Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primi- tives for high-performance differentiable rendering. arXiv:2011.03277 (2020)

    2011 arXiv

  12. [20]

    ACM Transactions on Graphics (TOG) (2025) DiffGI: Differentiable Geometry Images 17

    Li, S., Liu, R., Liu, C., Wang, Z., He, G., Li, Y.L., Jin, X., Wang, H.: GarmageNet: A multimodal generative framework for sewing pattern design and generic garment modeling. ACM Transactions on Graphics (TOG) (2025) DiffGI: Differentiable Geometry Images 17

    2025

  13. [21]

    CoRR (2025)

    Li, Y., Zou, Z.X., Liu, Z., Wang, D., Liang, Y., Yu, Z., Liu, X., Guo, Y.C., Liang, D., Ouyang, W., Cao, Y.P.: TripoSG: High-fidelity 3D shape synthesis using large- scale rectified flow models. CoRR (2025)

    2025

  14. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    Liao, Y., Donné, S., Geiger, A.: Deep marching cubes: Learning explicit surface representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    2018

  15. [23]

    In: ICLR (2023)

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023)

    2023

  16. [24]

    In: ICLR (2024)

    Liu, Z., Feng, Y., Xiu, Y., Liu, W., Paull, L., Black, M.J., Schölkopf, B.: Ghost on the shell: An expressive representation of general 3D shapes. In: ICLR (2024)

    2024

  17. [25]

    In: CVPR

    Long, X., Lin, C., Liu, L., Liu, Y., Wang, P., Theobalt, C., Komura, T., Wang, W.: NeuralUDF: Learning unsigned distance fields for multi-view reconstruction of surfaces with arbitrary topologies. In: CVPR. pp. 20834–20843 (2023)

    2023

  18. [26]

    ACM SIGGRAPH Computer Graphics (1987)

    Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3D surface con- struction algorithm. ACM SIGGRAPH Computer Graphics (1987)

    1987

  19. [27]

    In: CVPR

    Meng, X., Chen, W., Yang, B.: NeAT: Learning neural implicit surfaces with ar- bitrary topologies from multi-view images. In: CVPR. pp. 248–258 (2023)

    2023

  20. [28]

    Transactions on Machine Learning Research (2024)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez,P.,Haziza,D.,Massa,F.,El-Nouby,A.,etal.:DINOv2:Learningrobust visual features without supervision. Transactions on Machine Learning Research (2024)

    2024

  21. [29]

    In: CVPR (2019)

    Park,J.,Florence,P.,Straub,J.,Newcombe,R.,Lovegrove,S.:DeepSDF:Learning continuous signed distance functions for shape representation. In: CVPR (2019)

    2019

  22. [30]

    In: ICCV (2023)

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023)

    2023

  23. [31]

    Advances in neural information processing systems (2017)

    Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep hierarchical feature learn- ing on point sets in a metric space. Advances in neural information processing systems (2017)

    2017

  24. [32]

    In: CVPR (2022)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)

    2022

  25. [33]

    In: Symposium on Geometry Processing (2003)

    Sander, P.V., Wood, Z.J., Gortler, S.J., Snyder, J., Hoppe, H.: Multi-chart geom- etry images. In: Symposium on Geometry Processing (2003)

    2003

  26. [34]

    In: NeurIPS (2021)

    Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In: NeurIPS (2021)

    2021

  27. [35]

    ACM Transactions on Graphics (2023)

    Shen, T., Munkberg, J., Hasselgren, J., Yin, K., Wang, Z., Chen, W., Gojcic, Z., Fidler, S., Sharp, N., Gao, J.: Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (2023)

    2023

  28. [36]

    In: ECCV (2016)

    Sinha, A., Bai, J., Ramani, K.: Deep learning 3D shape surfaces using geometry images. In: ECCV (2016)

    2016

  29. [37]

    In: ICCV (2025)

    Son, S., Gadelha, M., Zhou, Y., Fisher, M., Xu, Z., Qiao, Y.L., Lin, M.C., Zhou, Y.: DMesh++: An efficient differentiable mesh for complex shapes. In: ICCV (2025)

    2025

  30. [38]

    In: Advances in Neural Information Processing Systems (2024)

    Son, S., Gadelha, M., Zhou, Y., Xu, Z., Lin, M.C., Zhou, Y.: DMesh: A differen- tiable mesh representation. In: Advances in Neural Information Processing Systems (2024)

    2024

  31. [39]

    Stippel, C., Mujkanovic, F., Leimkühler, T., Hermosilla, P.: Marching neurons: Ac- curatesurfaceextractionforneuralimplicitshapes.ACMTransactionsonGraphics (2025)

    2025

  32. [40]

    In: NeurIPS (2022) 18 E

    Wang, L., Chen, W., Meng, X., Yang, B., Li, J., Gao, L., et al.: HSDF: Hybrid sign and distance field for modeling surfaces with arbitrary topologies. In: NeurIPS (2022) 18 E. Shim et al

    2022

  33. [41]

    In: Pro- ceedings of the 32nd ACM International Conference on Multimedia (2024)

    Wang, R., Li, J., Zeng, D., Ma, X., Xu, Z., Zhang, J., Zhao, Q.: GenUDC: High quality 3D mesh generation with unsigned dual contouring representation. In: Pro- ceedings of the 32nd ACM International Conference on Multimedia (2024)

    2024

  34. [42]

    In: ECCV (2024)

    Wang, Z., Wang, Y., Chen, Y., Xiang, C., Chen, S., Yu, D., Li, C., Su, H., Zhu, J.: CRM: Single image to 3D textured mesh with convolutional reconstruction model. In: ECCV (2024)

    2024

  35. [43]

    arXiv preprint arXiv:2512.14692 (2025)

    Xiang, J., Chen, X., Xu, S., Wang, R., Lv, Z., Deng, Y., Zhu, H., Dong, Y., Zhao, H., Yuan, N.J., Yang, J.: Native and compact structured latents for 3D generation. arXiv preprint arXiv:2512.14692 (2025)

    2025 arXiv

  36. [44]

    In: CVPR (2025)

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3D latents for scalable and versatile 3D generation. In: CVPR (2025)

    2025

  37. [45]

    Graphical Models140, 101271 (2025)

    Xie, R., Huang, K., Luo, X., Chen, Y., Wang, L., Wang, Q., Ye, Q., Chen, W., Zheng, W., Huo, Y.: LDM: Large tensorial SDF model for textured mesh genera- tion. Graphical Models140, 101271 (2025)

    2025

  38. [46]

    CoRR (2024)

    Xu, J., Cheng, W., Gao, Y., Wang, X., Gao, S., Shan, Y.: InstantMesh: Efficient 3D mesh generation from a single image with sparse-view large reconstruction models. CoRR (2024)

    2024

  39. [47]

    In: International Conference on 3D Vision (3DV) (2025)

    Yan, X., Lee, H.H., Wan, Z., Chang, A.X.: An object is worth 64x64 pixels: Gen- erating 3D object via image diffusion. In: International Conference on 3D Vision (3DV) (2025)

    2025

  40. [48]

    In: ECCV (2024)

    Yang, H., Chen, Y., Pan, Y., Yao, T., Chen, Z., Wu, Z., Jiang, Y.G., Mei, T.: DreamMesh: Jointly manipulating and texturing triangle meshes for text-to-3D generation. In: ECCV (2024)

    2024

  41. [49]

    CoRR (2024)

    Yang, X., Shi, H., Zhang, B., Yang, F., Wang, J., Zhao, H., Liu, X., Wang, X., Lin, Q., Yu, J., et al.: Hunyuan3D 1.0: A unified framework for text-to-3D and image-to-3D generation. CoRR (2024)

    2024

  42. [50]

    ACM Transactions on Graphics (2024)

    Yu, X., Yuan, Z., Guo, Y.C., Liu, Y.T., Liu, J., Li, Y., Cao, Y.P., Liang, D., Qi, X.: TEXGen: a generative diffusion model for mesh textures. ACM Transactions on Graphics (2024)

    2024

  43. [51]

    CoRR (2025)

    Zhao, Z., Lai, Z., Lin, Q., Zhao, Y., Liu, H., Yang, S., Feng, Y., Yang, M., Zhang, S., Yang, X., et al.: Hunyuan3D 2.0: Scaling diffusion models for high resolution textured 3D assets generation. CoRR (2025)

    2025

  44. [52]

    Zheng, X., Pan, H., Wang, P., Tong, X., Liu, Y., Shum, H.: Locally attentional SDFdiffusionforcontrollable3Dshapegeneration.ACMTransactionsonGraphics (SIGGRAPH) (2023)

    2023

  45. [53]

    In: CVPR

    Zhou, J., Zhang, W., Ma, B., Shi, K., Liu, Y.S., Han, Z.: UDiFF: Generating conditional unsigned distance fields with optimal wavelet diffusion. In: CVPR. pp. 21496–21506 (2024) DiffGI: Differentiable Geometry Images 19 DiffGI: Differentiable Geometry Images for High-Fidelity ...

    2024

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.