Pith. sign in

REVIEW 3 major objections 4 minor 41 references

GraphBrep: Learning B-Rep in Graph Structure for Efficient CAD Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read GraphBrep claims that encoding a B-Rep's surface topology as an explicit adjacency matrix learned by graph diffusion removes the redundancy of tree-structured topology, cutting training time by up to 31.3% and inference time by up to…

desk verdict New adjacency-matrix learning is real and the efficiency story is credible; the maintained-quality claim is unsupported, mainly because the paper's own metrics don't measure B-Rep quality and validity sits at 56-67%. read the letter →

arxiv 2507.04765 v1 pith:NHOBUDSZ submitted 2025-07-07 cs.CV

Weilin Lai , Tie Xu , Hu Wang This is my paper
classification cs.CV
keywords B-RepgenerationCADgenerativemodelgraphdiffusionsurfaceadjacencymatrixtopologylearningefficientconditional
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

GraphBrep claims that the expensive part of direct CAD generation is not the geometry itself but how topology is encoded: representing which surfaces share edges in a tree structure forces each surface to carry a fixed maximum number of edges, most of them redundant. The paper replaces that tree with an explicit undirected adjacency matrix between surfaces, learned by a graph diffusion model conditioned on generated surface features. Because edge generation is then conditioned on the true adjacency matrix, the model only generates real edges, shortening the sequences the transformer must process. On the DeepCAD and ABC datasets this lowers training time by up to 31.3% and inference time by up to 56.3% while matching or slightly improving generation quality, including a higher fraction of watertight-valid solids.

What carries the argument

The central object is the B-Rep graph $G=(S,E,A)$, where $S$ are surfaces as nodes, $E$ are geometric edges, and $A$ is an undirected weighted adjacency matrix whose entry $A[i,j]$ counts how many edges are shared between surfaces $i$ and $j$. The mechanism that carries the argument is a graph diffusion model: a graph-transformer denoiser that recovers the adjacency matrix from noise conditioned on the denoised surface features (bounding boxes and VAE latents), with symmetry enforced by adding the predicted noise to its transpose. This explicit adjacency matrix is what lets the model condition edge generation on real connectivity instead of generating a fixed number of edges per surface, so the transformer attends over sequences of actual edges only.

What would settle it

Count, on a held-out set, how often the nearest-neighbor vertex-stitching rule reproduces the exact shared-vertex structure of the ground-truth B-Rep after generation; if that agreement falls well below the reported validity rates (66.62% on DeepCAD, 55.72% on ABC), the quality-parity claim rests on a heuristic that fails even when the watertight check passes.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a B-Rep's topology can be factored out of edge geometry and learned explicitly as a weighted undirected graph over surfaces, and doing so removes the redundancy that dominates cost in prior direct B-Rep generation. The joint distribution is decomposed as $p(G) = p(E|A,S)\,p(A|S)\,p(S)$: surfaces (nodes) are generated first, the surface adjacency matrix $A$ is generated by a graph-transformer diffusion model conditioned on the surfaces, and edges are generated only for the pairs indicated by $A$. The adjacency entries are integers counting shared edges, and the matrix is symmetrized during sampling. The consequence is that edge sequences shrink from a fixed per-surface allowance (600 on DeepCAD, 1500 on ABC during training) to the actual edge count (120 and 150), and because attention cost is quadratic in sequence length, training and inference become substantially cheaper. The reported validity of generated watertight solids also improves slightly, from 63.13% to 66.62% on DeepCAD and from 53.56% to 55.72% on ABC.

Load-bearing premise

The method assumes that nearest vertices from different generated edges can be treated as shared vertices, so the true vertex connectivity of the solid is recovered by a geometry-based nearest-neighbor rule rather than by explicit topological generation.

Editorial extensions

If this is right

  • Direct B-Rep generation can run on smaller GPU budgets: the same generation quality with up to 31.3% less training time and up to 56.3% less inference time at the tested batch sizes.
  • The explicit adjacency matrix makes topology inspectable: one can read which surfaces share edges and use that matrix as a conditioning signal for downstream editing, assembly, or repair tasks instead of recovering topology implicitly.
  • Because edge sequences scale with the true number of edges rather than a fixed per-surface maximum, the cost advantage should grow on datasets with larger models, where the quadratic attention cost of long edge sequences is most severe.
  • The category-conditional results on the Furniture dataset show the same graph decomposition supports class-conditioned generation with classifier-free guidance, so the efficiency gain does not require giving up conditional control.

Reading between the lines

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

  • The paper does not pursue it, but the explicit adjacency matrix could be paired with a learned edge-topology module instead of the nearest-neighbor stitching heuristic; the reported validity ceiling near two-thirds on DeepCAD suggests that stitching rule, not surface or edge geometry, is the current bottleneck to fully valid solids.
  • If topology is truly separable from geometry, one testable extension is to freeze the adjacency diffusion after training and re-use it as a topological prior for new surface and edge decoders, which would amortize the cost of learning connectivity across many geometric models.
  • The efficiency comparison is against one baseline at fixed model sizes; a natural stress test is to push batch size and maximum surface count further, since the quadratic-attention argument predicts the inference gap should widen as the sequences grow.
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 GraphBrep, a generative model for B-Rep CAD models that represents surface topology explicitly as an undirected weighted adjacency matrix between surfaces, learned by a graph diffusion model conditioned on surface features. The joint distribution is decomposed as p(E|A,S)p(A|S)p(S), and edge generation is conditioned on the pair of surfaces sharing each edge. The authors claim this compact representation reduces redundant edge information, lowering training and inference costs while maintaining generation quality relative to BrepGen and DeepCAD. Experiments are reported on DeepCAD, ABC, and Furniture datasets, with quantitative distribution and CAD metrics, efficiency comparisons, and conditional generation results.

Significance. If the quality-parity claim holds, the efficiency gain is practically significant: the edge sequence length drops from 600/1500 (BrepGen) to 120/150 (GraphBrep) on DeepCAD/ABC, with reported training and inference time reductions up to 31.3% and 56.3%, respectively, consistent with quadratic attention cost. The explicit adjacency-matrix representation is a principled alternative to tree-structured topology and the probability decomposition in Eq. (3) is exact. However, the quality half of the central claim is not yet established: the paper relies on point-cloud metrics that it concedes in Section 5 do not reflect B-Rep quality, and the nearest-vertex heuristic for edge topology yields watertight-validity rates of only 66.62% (DeepCAD) and 55.72% (ABC), only a few points above BrepGen. The efficiency analysis is the strongest contribution; the topology-validity evidence needs substantial strengthening.

major comments (3)
  1. [Section 3.1, final paragraph] The within-surface edge topology is resolved by the deterministic rule that 'the nearest vertices from different edges are regarded as shared vertices.' This is a post-processing heuristic, not part of the generative model, and nothing enforces that the matched vertices form a valid 2-manifold loop structure. The validity rates in Table 1 (66.62% on DeepCAD, 55.72% on ABC) suggest this heuristic fails in a large fraction of samples, and the absolute gains over BrepGen are only 3.49 and 2.16 percentage points. The paper should directly evaluate the fidelity of this reconstruction, for example by comparing the recovered edge-vertex connectivity against ground-truth B-Reps on a validation subset, or by ablate the heuristic with a learned edge-topology module.
  2. [Section 5, Conclusion and Limitations] The authors state that 'existing evaluation metrics borrowed from point cloud generation fail to reflect the quality and distribution of B-Rep models accurately.' Since the quantitative quality-parity claim in Table 1 rests on COV, MMD, and JSD computed from surface point clouds, those numbers cannot by themselves support the conclusion of 'high-quality CAD generation.' The paper should either adopt B-Rep-specific metrics (for example topology-aware validity or edge-connectivity accuracy) or provide a systematic analysis of topological correctness beyond the watertight check to substantiate the parity claim.
  3. [Table 1] The main quantitative comparison reports only averages over 20 runs without standard deviations or significance tests. The observed differences are small (e.g., MMD 0.93 vs 0.96 on DeepCAD; Validity 66.62% vs 63.13%), so it is unclear whether the claimed improvements or parity are statistically meaningful. Please report error bars, confidence intervals, or pairwise significance tests for the headline comparisons on both datasets.
minor comments (4)
  1. [Section 4.3] The ablation paragraph repeats the phrase 'limits of 50 and 80 on the ABC dataset' three times and refers to 'Table 3' where the results appear in Table 2; this should be corrected.
  2. [Section 3.2.2, Eq. (8)] Equation (8) appears incomplete: 'q = N(...)' lacks a subscript on q and the conditioning argument on x_{t-1}; it seems intended to be the same as Eq. (7) and should be fixed or removed.
  3. [Section 4.2.3] The sentence 'the edge number per surface is set to 20, and the number of surfaces is 30' would be clearer if it explicitly stated whether these are training-time or inference-time maximums.
  4. [Figures 4–6] Figure 5 is referenced before Figure 4 in the adjacency-matrix denoising paragraph, and the figure captions do not clearly indicate the data flow of the conditioning variables; renumbering or reordering would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the efficiency and quality claims rest on external baselines and measured costs, while the noted edge-topology heuristic is an empirical limitation rather than a self-referential derivation.

full rationale

The paper's derivation chain is self-contained against external benchmarks. The key decomposition in Eq. (3), p(G) = p(E|A,S) p(A|S) p(S), is a standard probability factorization, not a restatement of any predicted quantity. The adjacency matrix is generated by a graph diffusion model conditioned on surface features, and the edge distribution is then conditioned on the sampled adjacency matrix; this is a modeling choice, not a fitted constant renamed as a prediction. The efficiency claim is supported by direct timing comparisons with BrepGen and by the reported reduction in edge sequence length (600/1500 down to 120/150 on DeepCAD/ABC), which is a consequence of the explicit graph representation rather than of any fitted parameter. Quality is assessed against external baselines using COV, MMD, JSD, Novelty, Uniqueness, and Validity; even where validity is modest (66.62% on DeepCAD and 55.72% on ABC), those numbers are measured outcomes, not inputs to the method. The manuscript itself flags its own limitations: Section 5 states that existing point-cloud metrics 'fail to reflect the quality and distribution of B-Rep models accurately,' and Section 3.1 resolves within-surface edge topology with the heuristic that 'the nearest vertices from different edges are regarded as shared vertices.' Both are genuine empirical weaknesses and should be weighed as correctness risks, but neither makes the central claim true by definition. No load-bearing self-citation chain appears, and the external components adopted (UV-Net discretization, DiT-style denoisers, DiGress graph transformer) are standard building blocks, not unverified results imported from the authors' own prior work. Overall, the derivation is not circular.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The ledger is light. The method introduces no new physical entities. It inherits standard diffusion and graph-transformer machinery, plus a domain assumption about seam-split B-Reps and an ad hoc nearest-vertex heuristic for edge topology. The two free parameters are inference-time max surface count and the duplicate-box threshold, both of which affect the reported quality metrics.

free parameters (2)
  • max_surface_number_inference = 30 for DeepCAD, 50/80 for ABC and Furniture
    Hand-picked per dataset. Table 2 shows quality metrics vary substantially with this choice: raising DeepCAD max surface count from 30 to 50 drops COV from 74.06 to 66.47, so the reported parity depends on this tuning.
  • duplicate_bbox_threshold = 0.08
    Threshold in Section 4.1.3 used to remove duplicate sampled surface B-boxes; hand-chosen and affects how many surfaces are passed to the downstream latent and adjacency stages.
assumptions (4)
  • domain assumption Closed faces are split along seams so that B-Rep contains no loops and the surface graph is undirected
    Adopted from SolidGen in Section 3.1. If violated for a generated shape, the adjacency matrix cannot represent the topology.
  • ad hoc to paper Nearest-vertex matching between edge endpoints reconstructs valid shared vertices
    Section 3.1 states 'the nearest vertices from different edges are regarded as shared vertices'. If this heuristic fails, the B-Rep is invalid; validity of about 66% on DeepCAD suggests frequent failure.
  • domain assumption UV-Net uniform grid discretization with M=N=R=32 captures enough geometric detail
    Section 3.1 inherits this from prior work; it limits the fidelity of curved surfaces but is the accepted standard in the subfield.
  • standard math Standard DDPM, DiT, and DiGress equations hold as stated, including the Bayes posterior and variance schedule
    Section 3.2.2 uses standard diffusion results, though the equations are garbled in the extracted text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GraphBrep: Learning B-Rep in Graph Structure for Efficient CAD Generation." pith.science (2026). https://pith.science/paper/NHOBUDSZ

@misc{pith2026250704765,
  author       = {Pith},
  title        = {Pith review of: GraphBrep: Learning B-Rep in Graph Structure for Efficient CAD Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NHOBUDSZ}},
  note         = {Machine review of arXiv:2507.04765}
}
read the original abstract

Direct B-Rep generation is increasingly important in CAD workflows, eliminating costly modeling sequence data and supporting complex features. A key challenge is modeling joint distribution of the misaligned geometry and topology. Existing methods tend to implicitly embed topology into the geometric features of edges. Although this integration ensures feature alignment, it also causes edge geometry to carry more redundant structural information compared to the original B-Rep, leading to significantly higher computational cost. To reduce redundancy, we propose GraphBrep, a B-Rep generation model that explicitly represents and learns compact topology. Following the original structure of B-Rep, we construct an undirected weighted graph to represent surface topology. A graph diffusion model is employed to learn topology conditioned on surface features, serving as the basis for determining connectivity between primitive surfaces. The explicit representation ensures a compact data structure, effectively reducing computational cost during both training and inference. Experiments on two large-scale unconditional datasets and one category-conditional dataset demonstrate the proposed method significantly reduces training and inference times (up to 31.3% and 56.3% for given datasets, respectively) while maintaining high-quality CAD generation compared with SOTA.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 11 canonical work pages

  1. [1]

    J. Ho, A. Jain, P. Abbeel, Denoising Diffusion Probabilistic Models, in: Adv. Neural Inf. Process. Syst., Curran Associates, Inc., 2020: pp. 6840 –6851. https://proceedings.neurips.cc/paper/2020/hash/4c5bcfec8584af0d967f1ab101 79ca4b-Abstract.html (accessed October 28, 2024)

    work page 2020

  2. [2]

    Podell, Z

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, R. Rombach, SDXL: Improving Latent Diffusion Models for High -Resolution Image Synthesis, (2023). https://doi.org/10.48550/arXiv.2307.01952

    arXiv doi:10.48550/arxiv.2307.01952 2023

  3. [3]

    J. Jo, S. Lee, S.J. Hwang, Score -based Generative Modeling of Graphs via the System of Stochastic Differential Equations, in: 2022. https://www.semanticscholar.org/paper/Score-based-Generative-Modeling-of- Graphs-via-the-Jo-Lee/3312af738432a817e0d3d913dca87d7445a85696 (accessed October 27, 2024)

    work page 2022

  4. [4]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High -Resolution Image Synthesis with Latent Diffusion Models, (2022). https://doi.org/10.48550/arXiv.2112.10752

    arXiv doi:10.48550/arxiv.2112.10752 2022

  5. [5]

    Ramesh, M

    A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. V oss, A. Radford, M. Chen, I. Sutskever, Zero -Shot Text -to-Image Generation, (2021). https://doi.org/10.48550/arXiv.2102.12092

    arXiv doi:10.48550/arxiv.2102.12092 2021

  6. [6]

    Peebles, S

    W. Peebles, S. Xie, Scalable Diffusion Models with Transformers, (2023). https://doi.org/10.48550/arXiv.2212.09748

    arXiv doi:10.48550/arxiv.2212.09748 2023

  7. [7]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M. -A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, G. Lample, LLaMA: Open and Efficient Foundation Language Models, (2023). https://doi.org/10.48550/arXiv.2302.13971

    arXiv doi:10.48550/arxiv.2302.13971 2023

  8. [8]

    Achiam, S

    OpenAI, J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F.L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, R. Avila, I. Babuschkin, S. Balaji, V . Balcom, P. Baltescu, H. Bao, M. Bavarian, J. Belgum, I. Bello, J. Berdine, G. Bernadett -Shapiro, C. Berner, L. Bogdonoff, O. Boiko, M. Boyd, A.-L. Brakman, G. Brockman, T. Brooks, M. Brunda...

    arXiv doi:10.48550/arxiv.2303.08774 2024

Show all 41 references
  1. [9]

    DeepSeek-AI, D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, X. Zhang, X. Yu, Y . Wu, Z.F. Wu, Z. Gou, Z. Shao, Z. Li, Z. Gao, A. Liu, B. Xue, B. Wang, B. Wu, B. Feng, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, D. Dai, D. Chen, D. Ji, E. Li...

    2025 doi arXiv

  2. [10]

    Qwen, A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, X. ...

    2025 doi arXiv

  3. [11]

    Blattmann, T

    A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitch, M. Kilian, D. Lorenz, Y . Levi, Z. English, V . V oleti, A. Letts, V . Jampani, R. Rombach, Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets, (2023). https://doi.org/10.48550/arXiv.2311.15127

    2023 doi arXiv

  4. [12]

    W. Kong, Q. Tian, Z. Zhang, R. Min, Z. Dai, J. Zhou, J. Xiong, X. Li, B. Wu, J. Zhang, K. Wu, Q. Lin, J. Yuan, Y . Long, A. Wang, A. Wang, C. Li, D. Huang, F. Yang, H. Tan, H. Wang, J. Song, J. Bai, J. Wu, J. Xue, J. Wang, K. Wang, M. Liu, P. Li, S. Li, W. Wang, W. Yu, X. Deng...

    2025 doi arXiv

  5. [13]

    Polyak, A

    A. Polyak, A. Zohar, A. Brown, A. Tjandra, A. Sinha, A. Lee, A. Vyas, B. Shi, C.-Y . Ma, C.-Y . Chuang, D. Yan, D. Choudhary, D. Wang, G. Sethi, G. Pang, H. Ma, I. Misra, J. Hou, J. Wang, K. Jagadeesh, K. Li, L. Zhang, M. Singh, M. Williamson, M. Le, M. Yu, M.K. Singh, P. Zhan...

    2025 doi arXiv

  6. [14]

    Mildenhall, P.P

    B. Mildenhall, P.P. Srinivasan, M. Tancik, J.T. Barron, R. Ramamoorthi, R. Ng, NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis, (2020). https://doi.org/10.48550/arXiv.2003.08934

    2020 doi arXiv

  7. [15]

    Poole, A

    B. Poole, A. Jain, J.T. Barron, B. Mildenhall, DreamFusion: Text -to-3D using 2D Diffusion, (2022). https://doi.org/10.48550/arXiv.2209.14988

    2022 doi arXiv

  8. [16]

    Nichol, H

    A. Nichol, H. Jun, P. Dhariwal, P. Mishkin, M. Chen, Point -E: A System for Generating 3D Point Clouds from Complex Prompts, (2022). https://doi.org/10.48550/arXiv.2212.08751

    2022 doi arXiv

  9. [17]

    M. Chen, R. Shapovalov, I. Laina, T. Monnier, J. Wang, D. Novotny, A. Vedaldi, PartGen: Part -level 3D Generation and Reconstruction with Multi -View Diffusion Models, (2024). https://doi.org/10.48550/arXiv.2412.18608

    2024 doi arXiv

  10. [18]

    Z. Zhao, Z. Lai, Q. Lin, Y . Zhao, H. Liu, S. Yang, Y . Feng, M. Yang, S. Zhang, X. Yang, H. Shi, S. Liu, J. Wu, Y . Lian, F. Yang, R. Tang, Z. He, X. Wang, J. Liu, X. Zuo, Z. Chen, B. Lei, H. Weng, J. Xu, Y . Zhu, X. Liu, L. Xu, C. Hu, S. Yang, S. Zhang, Y . Liu, T. Huang, L....

    2025 doi arXiv

  11. [19]

    M. Mews, A. Aynetdinov, V . Schiller, P. Eisert, A. Akbik, Don’t Mesh with Me: Generating Constructive Solid Geometry Instead of Meshes by Fine -Tuning a Code-Generation LLM, (2024). https://doi.org/10.48550/arXiv.2411.15279

    2024 doi arXiv

  12. [20]

    V . Liu, J. Vermeulen, G. Fitzmaurice, J. Matejka, 3DALL-E: Integrating Text- to-Image AI in 3D Design Workflows, (2023). https://doi.org/10.48550/arXiv.2210.11603

    2023 doi arXiv

  13. [21]

    Yavartanoo, S

    M. Yavartanoo, S. Hong, R. Neshatavar, K.M. Lee, Text2CAD: Text to 3D CAD Generation via Technical Drawings, (2024). https://doi.org/10.48550/arXiv.2411.06206

    2024 doi arXiv

  14. [22]

    X. Xu, J.G. Lambourne, P.K. Jayaraman, Z. Wang, K.D.D. Willis, Y . Furukawa, BrepGen: A B -rep Generative Diffusion Model with Structured Latent Geometry, (2024). https://doi.org/10.48550/arXiv.2401.15563

    2024 doi arXiv

  15. [23]

    Jayaraman, J.G

    P.K. Jayaraman, J.G. Lambourne, N. Desai, K.D.D. Willis, A. Sanghi, N.J.W. Morris, SolidGen: An Autoregressive Model for Direct B-rep Synthesis, (2023). https://doi.org/10.48550/arXiv.2203.13944

    2023 doi arXiv

  16. [24]

    Xu, K.D.D

    X. Xu, K.D.D. Willis, J.G. Lambourne, C. -Y . Cheng, P.K. Jayaraman, Y . Furukawa, SkexGen: Autoregressive Generation of CAD Construction Sequences with Disentangled Codebooks, (2022). https://doi.org/10.48550/arXiv.2207.04632

    2022 doi arXiv

  17. [25]

    X. Xu, P.K. Jayaraman, J.G. Lambourne, K.D.D. Willis, Y . Furukawa, Hierarchical Neural Coding for Controllable CAD Model Generation, (2023). https://doi.org/10.48550/arXiv.2307.00149

    2023 doi arXiv

  18. [26]

    R. Wu, C. Xiao, C. Zheng, DeepCAD: A Deep Generative Network for Computer-Aided Design Models, (2021). https://doi.org/10.48550/arXiv.2105.09492

    2021 doi arXiv

  19. [27]

    Ganin, S

    Y . Ganin, S. Bartunov, Y . Li, E. Keller, S. Saliceti, Computer-Aided Design as Language, in: Adv. Neural Inf. Process. Syst., Curran Associates, Inc., 2021: pp. 5885–5897. https://proceedings.neurips.cc/paper/2021/hash/2e92962c0b6996add9517e424 2ea9bdc-Abstract.html (accesse...

    2021

  20. [28]

    Zhang, S

    Z. Zhang, S. Sun, W. Wang, D. Cai, J. Bian, FlexCAD: Unified and Versatile Controllable CAD Generation with Fine-tuned Large Language Models, (2025). 37 https://doi.org/10.48550/arXiv.2411.05823

    2025 doi arXiv

  21. [29]

    A. Seff, W. Zhou, N. Richardson, R.P. Adams, Vitruvion: A Generative Model of Parametric CAD Sketches, (2022). https://doi.org/10.48550/arXiv.2109.14124

    2022 doi arXiv

  22. [30]

    Ganin, S

    Y . Ganin, S. Bartunov, Y . Li, E. Keller, S. Saliceti, Computer-aided design as language, in: Proc. 35th Int. Conf. Neural Inf. Process. Syst., Curran Associates Inc., Red Hook, NY , USA, 2024: pp. 5885–5897

    2024

  23. [31]

    W. Ma, S. Chen, Y . Lou, X. Li, X. Zhou, Draw Step by Step: Reconstructing CAD Construction Sequences from Point Clouds via Multimodal Diffusion, in: 2024 IEEECVF Conf. Comput. Vis. Pattern Recognit. CVPR, IEEE, Seattle, WA, USA, 2024: pp. 27144 –27153. https://doi.org/10.1109...

    2024

  24. [32]

    W. Para, S. Bhat, P. Guerrero, T. Kelly, N. Mitra, L.J. Guibas, P. Wonka, SketchGen: Generating Constrained CAD Sketches, in: Adv. Neural Inf. Process. Syst., Curran Associates, Inc., 2021: pp. 5077 –5088. https://proceedings.neurips.cc/paper/2021/hash/28891cb4ab421830acc36b1f...

    2021

  25. [33]

    H. Guo, S. Liu, H. Pan, Y . Liu, X. Tong, B. Guo, ComplexGen: CAD reconstruction by B-rep chain complex generation, ACM Trans Graph 41 (2022) 129:1-129:18. https://doi.org/10.1145/3528223.3530078

    2022

  26. [34]

    C. Niu, Y . Song, J. Song, S. Zhao, A. Grover, S. Ermon, Permutation Invariant Graph Generation via Score -Based Generative Modeling, in: Proc. Twenty Third Int. Conf. Artif. Intell. Stat., PMLR, 2020: pp. 4474 –4484. https://proceedings.mlr.press/v108/niu20a.html (accessed Oc...

    2020

  27. [35]

    Huang, L

    H. Huang, L. Sun, B. Du, Y . Fu, W. Lv, GraphGDP: Generative Diffusion Processes for Permutation Invariant Graph Generation, in: IEEE Computer Society, 2022: pp. 201–210. https://doi.org/10.1109/ICDM54844.2022.00030

    2022

  28. [36]

    Vignac, I

    C. Vignac, I. Krawczuk, A. Siraudin, B. Wang, V . Cevher, P. Frossard, DiGress: Discrete Denoising diffusion for graph generation, (2023). https://doi.org/10.48550/arXiv.2209.14734

    2023 doi arXiv

  29. [37]

    X. Chen, Y . Li, A. Zhang, L.-P. Liu, NVDiff: Graph Generation through the Diffusion of Node Vectors, (2023). http://arxiv.org/abs/2211.10794 (accessed October 27, 2024)

    2023 arXiv

  30. [38]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High -Resolution Image Synthesis with Latent Diffusion Models, in: IEEE Computer Society, 38 2022: pp. 10674–10685. https://doi.org/10.1109/CVPR52688.2022.01042

    2022

  31. [39]

    Jayaraman, A

    P.K. Jayaraman, A. Sanghi, J.G. Lambourne, K.D.D. Willis, T. Davies, H. Shayani, N. Morris, UV-Net: Learning from Boundary Representations, (2021). https://doi.org/10.48550/arXiv.2006.10211

    2021 doi arXiv

  32. [40]

    S. Koch, A. Matveev, Z. Jiang, F. Williams, A. Artemov, E. Burnaev, M. Alexa, D. Zorin, D. Panozzo, ABC: A Big CAD Model Dataset For Geometric Deep Learning, (2019). https://doi.org/10.48550/arXiv.1812.06216

    2019 doi arXiv

  33. [41]

    Kingma, M

    D.P. Kingma, M. Welling, Auto -Encoding Variational Bayes, (2022). https://doi.org/10.48550/arXiv.1312.6114. 39

    2022 doi arXiv

Pith tools

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