Pith. sign in

REVIEW 4 major objections 5 minor 60 references

ParaCAD claims that encoding each B-Rep face as native parametric surface tokens—surface type, continuous parameters, and a 3D placement—lets an autoregressive transformer generate CAD solids from point clouds with higher geometric accuracy

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 19:02 UTC pith:DG73WDVZ

load-bearing objection A solid, well-engineered method paper whose abstract overstates watertightness; the validity-disaggregated metrics need fixing before I'd trust Table 1. the 4 major comments →

arxiv 2607.17093 v1 pith:DG73WDVZ submitted 2026-07-19 cs.CV

Autoregressive B-Rep Shape Generation with Parametric Surfaces

Dafei Qin , Rui Xu , Zeyu Shen , Kaichun Qiao , Hongyang Lin , Qixuan Zhang , Huaijin Pi , Lan Xu
show 3 more authors
Jingyi Yu Wenping Wang Taku Komura
This is my paper
classification cs.CV
keywords B-Rep generationparametric surfacespoint cloud conditioningautoregressive transformersurface tokenizationCAD reconstructionfinite scalar quantizationsurface–surface intersection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that a surface-centric representation—where each face is encoded by its exact surface type, continuous parameters, and a similarity transform pose—allows a decoder-only autoregressive model to generate B-Rep solids from point clouds with higher geometric precision and semantic fidelity than latent-grid or point-based baselines. The authors argue that preserving analytic primitive types and parameters, while deferring topology recovery to a deterministic geometric intersection post-processor, yields B-Reps that are directly editable and manufacturable. They support this with experiments showing lower Chamfer distance and higher type-conditioned surface F1 than existing methods, and with a validity analysis that, while imperfect, demonstrates the post-processor can assemble closed 2-manifolds from generated surfaces. A sympathetic reader would care because it addresses a core limitation of learned CAD generation: losing native surface semantics in tokenized representations.

Core claim

The central discovery is that tokenizing each B-Rep face as a fixed block of 13 tokens—a surface-type token, four tokens for the bounded UV domain, a scalar token for cone semi-angle or torus minor radius, and seven tokens for translation, Euler angles, and log-scale of the instance-level placement—makes point-cloud-conditioned B-Rep generation tractable and accurate. Freeform faces are encoded via finite scalar quantization of a latent grid, keeping the same slot structure. A decoder-only transformer predicts the token sequence from point-cloud conditioning tokens, and a geometric post-processor extends UV domains, detects tangent relations, computes surface–surface intersections, and solve

What carries the argument

The hybrid surface tokenizer. Each face becomes a 13-token block: one surface type token (plane, cylinder, cone, sphere, torus, or freeform); four tokens quantizing the bounded UV domain; one token for an additional scalar (cone semi-angle or torus minor radius); and seven tokens for translation, Euler angles, and log-scale of the instance-level similarity transform. Freeform patches use finite scalar quantization of a latent grid in the same slots. This tokenizer makes analytic surfaces exact after decoding, and the canonical coordinate system keeps sequences compact. The post-processor then recovers edges and vertices deterministically through UV extension, surface–surface intersection, po

Load-bearing premise

The whole argument rests on the geometric post-processor's ability to extend, intersect, and select generated surface patches into a closed 2-manifold; if that assembly fails, no valid B-Rep is produced even when the model's surface tokens are all correct.

What would settle it

Run the post-processor on the same generated surfaces twice, once with the point-cloud coverage score and once with random piece selection. If output validity and Chamfer distance do not degrade without the point-cloud anchor, the conditioning is not doing the claimed suppression. Also, on a test subset whose ground-truth solids have many curved intersections, compare the validity of assemblies from detokenized ground-truth surfaces against generated surfaces; if both stay near 55%, the failure sits in the assembly stage, not the generator.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Analytic primitives survive generation, so downstream CAD operations like fillet, Boolean, and offset can be applied directly to the output without primitive fitting.
  • Freeform and analytic faces share a single token structure, letting one model cover both while avoiding grid-approximation error.
  • Topology is recovered deterministically from surface intersections rather than learned, shortening token sequences and simplifying generation.
  • The surface-centric conditioning can be chained after any point-cloud-generating 3D pipeline, turning text- or image-derived clouds into B-Rep solids.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The pipeline's validity on generated surfaces (55.81%) is well below its validity on detokenized ground truth (76.60%), suggesting the post-processor is currently the weak link; improving UV-domain prediction or adding a learned piece-selection step is the most direct route to higher watertightness.
  • The tokenizer is an invertible, compact surface encoding that could serve as output space for other input modalities (sketch, image, partial scan) without retraining the geometric assembly.
  • If extended to trimmed or higher-order NURBS patches, the same divide-and-intersect assembly could cover a broader share of industrial CAD models than the five analytic primitives plus freeform approximation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes ParaCAD, an autoregressive framework for point-cloud-conditioned B-Rep generation that tokenizes each B-Rep face as a sequence of discrete tokens encoding native parametric surface type, intrinsic parameters, UV bounds, and SIM(3) placement (or a freeform FSQ latent for non-analytic patches). A decoder-only GPT generates these surface-token sequences conditioned on an input point cloud via cross-attention. A post-processing stage extends UV domains, computes surface-surface intersections, and selects a subset of the resulting pieces by solving a binary program with point-cloud coverage scores and edge-based penalties, intended to yield a closed, watertight B-Rep. The method is trained on a new 583K-instance point-cloud-to-CAD dataset derived from ABC, and evaluated against P2CAD, SEDNet+Point2CAD, HoLa, and NVDNet on 1,000 test shapes. The paper reports lower Chamfer Distance and higher surface-type-conditioned precision/recall/F1 than baselines, and also reports the ratio of topologically valid B-Rep outputs after post-processing.

Significance. If the technical claims hold, ParaCAD is a meaningful step toward preserving native surface semantics in generative B-Rep modeling, which is important for downstream CAD editing, simulation, and manufacturing. The surface-centric tokenization with explicit type tokens is a natural and sensible departure from purely grid- or latent-based surface encodings, and the deterministic geometric assembly stage is an elegant way to reduce the burden on the autoregressive model. The paper also contributes a large-scale ABC-derived point-cloud-to-B-Rep dataset and evaluates against several recently proposed baselines using their official code. However, the central claim of generating watertight B-Reps is substantially tempered by the paper's own Table 4, and the evaluation protocol does not state how invalid outputs are treated in the headline geometry/semantic metrics, making the reported advantages ambiguous until this is resolved.

major comments (4)
  1. [§3.3, Eqs. 8–10] The binary program does not enforce a closed 2-manifold, contrary to Contribution 3's phrase 'ensuring topological validity'. Equation 9 only enforces the 2-manifold condition for inner and conflict edges; Eq. 10 merely ties each boundary edge to its sole incident face, while the objective only subtracts a penalty w_b for selecting boundary edges. An open shell can therefore be feasible and, if its coverage is high enough, optimal. The paper's own Table 4 confirms this: only 55.81% of generated outputs are valid, and even detokenized ground-truth patches fail 23.4% of the time. Please either add explicit closure constraints (e.g., requiring every selected boundary edge to be matched by another selected edge/face, or solving over closed shells) or revise the contribution and abstract to claim 'topological consistency as a soft objective' rather than guaranteed validity.
  2. [§4.3.1, Table 1] The geometric similarity metric (CD) is computed by sampling from 'the generated B-Rep', but the manuscript does not state whether invalid/open/non-manifold outputs are excluded, repaired, or included. Since only 55.81% of generated outputs pass the validity check, Table 1's CD and F1 values could be averaged over a majority of invalid shells, making the reported 'faithful point-cloud alignment' and 'watertightness' comparisons misleading. Please clarify the exact protocol: are all 1,000 test shapes used regardless of validity? If invalid outputs are included, report metrics separately for valid and invalid subsets, or for the whole set with a clear statement. If they are excluded, report the number of excluded shapes and show that the comparison is still statistically meaningful.
  3. [§4.3.2, Table 3] The post-processing hyperparameters λ and w_b are tuned on 50 generated samples before evaluation. The 'Balanced' configuration gives 67.5% validity on those 50 samples, but the later full test-set validity (Table 4) is 55.81%, suggesting possible overfitting of these hyperparameters to the small tuning set. There are also no error bars, significance tests, or repeated-seed results for any of the quantitative comparisons. Please report variance over multiple runs, use a validation-derived configuration with a larger tuning set, or otherwise demonstrate that the reported gains are not the result of tuning on a small set.
  4. [Abstract and §6] The abstract states that ParaCAD 'constructs a valid B-Rep' and claims superiority in 'watertightness', while Section 6 candidly says the success rate 'needs improvement' and Table 4 shows 55.81% validity. These statements are in direct tension. The low validity is not a minor limitation; it is central to the paper's contribution. At minimum, the abstract and conclusion should be rephrased to reflect that validity is a soft objective achieved in roughly half of generated cases, and the paper should more prominently discuss the failure modes in Fig. 6 and the conditions under which post-processing succeeds or fails.
minor comments (5)
  1. [§2.3.2 / §4.3.1] The baseline name is spelled 'ParSeNet' in Related Work but 'ParseNet' in Table 1. Please unify.
  2. [Table 3, 4] Table 3's 'Best validity' 77.5% is on the 50-sample sweep set, while Table 4's generated validity is 55.81% on the full test set; the discrepancy should be explained in the text to avoid confusion.
  3. [§4.3.1] For generative baselines and ParaCAD, the best-of-k case selection (best-of-16 and best-of-4) is mentioned but not described in terms of what objective is used for selection (e.g., point-cloud distance, validity). This matters for reproducibility.
  4. [§3.3] The point-cloud-based selection in Eq. 7 discards patches with coverage below λ times the maximum. When the entire point cloud has low coverage (e.g., due to dense over-generation), this relative criterion may discard all patches or keep spurious ones. A short discussion or ablation would help.
  5. [Appendix] The main text refers to an appendix for the analytic forms of S_τ and the formal definitions of SP/SR. Since the paper is uploaded as a standalone document, consider including these definitions in the main text or clearly flagging the appendix pages.

Circularity Check

0 steps flagged

No significant circularity; central pipeline is self-contained against external benchmarks, with only a minor non-load-bearing self-citation.

full rationale

The paper's central derivation chain is an autoregressive generator that predicts native parametric surface tokens from point clouds, followed by a deterministic geometric post-processor. The tokenizer encodes surface types and parameters directly (Eq. 1), and the model is trained with a standard token-level cross-entropy loss (Eq. 5) on held-out ABC-derived shapes, evaluated against external baselines using official baseline code. The post-processing stage uses the input point cloud in the coverage-based selection (Eq. 6) and the binary program (Eqs. 8-10), but this is inherent to reconstruction, not a fitted quantity being relabeled as a prediction: the selection operates only on surface pieces already generated by the model, and the reported Chamfer distance is not identical to the coverage objective. The one self-citation, [Shen et al. 2025] for tangent detection and parameter adjustment, is by co-author Zeyu Shen, but it is not load-bearing: the paper directly measures post-processing validity on raw ground-truth patches, detokenized ground-truth patches, and generated patches (Table 4), providing independent empirical support. The candid limitation that only 55.81% of generated outputs are valid is a correctness/robustness concern, not a circularity of the derivation. No equation reduces to its input by construction, and no claim is justified solely by an unverified self-citation chain.

Axiom & Free-Parameter Ledger

8 free parameters · 7 axioms · 0 invented entities

The paper introduces no new physical entities. The main free parameters are post-processing thresholds (δ_d, δ_n, λ, w_b), tokenizer quantization ranges, and codebook size; several are tuned on small validation sets or deferred to an appendix. The axioms are standard modeling assumptions about surface families, intersection robustness, and point-coverage selection.

free parameters (8)
  • distance tolerance δ_d in coverage score (Eq. 6) = not reported
    Controls whether a surface piece is considered supported by the point cloud; value is not given in the main text.
  • normal-consistency threshold δ_n (Eq. 6) = cos(π/6) ≈ 0.866
    Hand-set threshold; directly changes which pieces survive the coverage filter.
  • relaxation factor λ (Eq. 7) = 0.6 in Balanced config
    Selected via Pareto sweep on 50 generated samples (Sec. 4.3.2); affects piece discard and final geometry.
  • boundary penalty w_b (Eq. 8) = 3.0 in Balanced config
    Selected via same Pareto sweep; balances watertightness against geometric coverage.
  • UV bound quantization codebooks (tokens s1–s4) = fixed ranges; details in appendix
    Hand-chosen ranges; discretization error propagates to surface boundaries and topological validity.
  • SIM(3) translation/rotation/log-scale quantization ranges (tokens s6–s12) = not reported
    Hand-chosen ranges; out-of-range surfaces are degraded to freeform patches.
  • tokenizer codebook size / FSQ levels = 1024
    Ablated in Table 5; 512 drops validity while 2048 gives similar performance.
  • best-of-k sampling at evaluation = best-of-16 for HoLa, best-of-4 for ParaCAD
    Generative methods are evaluated on their best sample; this favors stronger results.
axioms (7)
  • standard math SIM(3) canonicalization and Euler-angle decomposition of rotations are valid for representing surface placement.
    Section 3.1.2 relies on translation, rotation, and isotropic scale to map arbitrary surfaces into canonical coordinates.
  • domain assumption The five analytic surface families (plane, cylinder, cone, sphere, torus) plus freeform patches adequately cover ABC B-Rep faces.
    Section 3.1 restricts tokenization to these families; out-of-range analytic surfaces are converted to freeform.
  • domain assumption Surface UV extension plus tangent-relation adjustment following [Shen et al. 2025] guarantees that surface-surface intersections can be computed robustly.
    Section 3.3 depends on this to recover edges and vertices; Table 4 shows it fails on a substantial fraction of generated patches.
  • domain assumption The point-coverage score in Eq. 6 is sufficient to select the correct surface pieces.
    Section 3.3 uses distance and normal thresholds; if the input point cloud is sparse or noisy, correct pieces can be discarded.
  • domain assumption A 32×32 UV sample grid and a 4×4×3 FSQ latent faithfully represent freeform patches.
    Section 3.1.3 uses this compression; quantization loss is acknowledged in the codebook-size ablation.
  • standard math The binary integer program in Eqs. 8–10 yields a closed 2-manifold shell when a feasible selection exists.
    Section 3.3 formulates the constraints; validity results in Table 4 are the empirical check.
  • domain assumption Point clouds sampled from exact CAD geometry with OpenCASCADE are representative conditioning inputs.
    Section 4.1 constructs the dataset this way; real-world scans with noise and outliers are only briefly tested in Section 4.5.

pith-pipeline@v1.3.0-alltime-deepseek · 14923 in / 11204 out tokens · 150393 ms · 2026-08-01T19:02:33.251551+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Autoregressive B-Rep Shape Generation with Parametric Surfaces." pith.science (2026). https://pith.science/paper/DG73WDVZ

@misc{pith2026260717093,
  author       = {Pith},
  title        = {Pith review of: Autoregressive B-Rep Shape Generation with Parametric Surfaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DG73WDVZ}},
  note         = {Machine review of arXiv:2607.17093}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generative CAD modeling has broad design and application potential. Despite significant advances in Boundary Representation (B-Rep) generation, the dominant representation in CAD, existing methods largely depend on uniformly sampled point- or grid-based geometry representations, sacrificing native surface types and parameters and thereby limiting geometric fidelity and downstream usability. We present ParaCAD, an autoregressive framework for point-cloud-conditioned B-Rep generation that directly operates on native parametric surfaces. ParaCAD introduces a surface-centric tokenization that explicitly encodes each face by its exact surface type and continuous parameters, preserving the intrinsic semantics of CAD geometry. Our model first generates parametric surfaces with constrained UV domains, and then constructs a valid B-Rep by globally intersecting these surfaces to recover edges and vertices. ParaCAD places point-cloud-conditioned generation at the core of B-Rep synthesis, making it practical for user-guided reconstruction and seamless integration into existing 3D generation pipelines. Extensive experiments demonstrate that ParaCAD produces accurate B-Reps with faithful point-cloud alignment, outperforming point-based baselines in geometric precision, robustness, watertightness and downstream usability.

Figures

Figures reproduced from arXiv: 2607.17093 by Dafei Qin, Hongyang Lin, Huaijin Pi, Jingyi Yu, Kaichun Qiao, Lan Xu, Qixuan Zhang, Rui Xu, Taku Komura, Wenping Wang, Zeyu Shen.

Figure 1
Figure 1. Figure 1: ParaCAD receives point clouds as conditions and generates CAD models with native support of original parametric surfaces. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison between B-Reps representations generated by (left) grid-based surface approximation and (right) native parametric surface representation. Surfaces are shaded to highlight different types. 2 Related Works 2.1 CAD Generation via Construction Operation A substantial body of learning-based CAD generation work [Ganin et al. 2021; Guo et al. 2025; Para et al. 2021; Seff et al. 2022; Willis et al. 2021… view at source ↗
Figure 3
Figure 3. Figure 3: ParaCAD Pipeline. (Top) The surface tokenizer records and dis￾cretizes the parametric surface representation. (Middle) The decoder-only autoregressive model generates sequences of surface tokens from point￾cloud conditions. (Bottom) Geometrical operations for consolidating the final B-Rep. primitive fitting algorithm and struggle to support diverse freeform surfaces. On the direct side, ComplexGen [Guo et … view at source ↗
Figure 4
Figure 4. Figure 4: qualitatively compares our method with the baselines. P2CAD and NVDNet output meshes, while HoLa and ParaCAD generate B-Rep surfaces. Overall, ParaCAD produces consistently higher-quality reconstructions. In particular, due to limited surface￾type support, NVDNet approximates torus and freeform regions with many disconnected planar/cylindrical patches, resulting in fragmented geometry and inconsistent sema… view at source ↗
Figure 10
Figure 10. Figure 10: Model casted of aluminum alloy. B-Reps generated by ParaCAD (left) can be directly manufactured (right) using standard mechanical pro￾cess [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 5
Figure 5. Figure 5: Post-processing pipeline. (a) Surface patches translated from generated tokens. (b) Surface UV extension. (c) Surface-surface intersection and split. (d) Point cloud coverage based selection. (e) Optimization-based selection. (f) Merged [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Text conditioned B-Rep Generation. Intermediate point clouds are generated by [Zhang et al. 2024] [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 6
Figure 6. Figure 6: Typical failure cases. GT (Left) vs. Pred (Right). Top: Red rectangu￾lar: Missing local regions due to limited point-cloud coverage. Purple ellipse: Imperfectly cut surfaces become floating structures. Bottom: Hallucination of a cone surface caused by insufficient point cloud resolution. SIGGRAPH Conference Papers ’26, July 19–23, 2026, Los Angeles, CA, USA [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 11
Figure 11. Figure 11: From point clouds to manufacturable B-Reps. Our method produces native B-Rep models that can be directly rendered with metallic materials and readily integrated into downstream design and manufacturing pipelines. SIGGRAPH Conference Papers ’26, July 19–23, 2026, Los Angeles, CA, USA [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 3 linked inside Pith

  1. [1]

    Transactions on Machine Learning Research , issn=

    SolidGen: An Autoregressive Model for Direct B-rep Synthesis , author=. Transactions on Machine Learning Research , issn=. 2023 , url=

    2023

  2. [2]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Uv-net: Learning from boundary representations , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  3. [3]

    ACM Transactions on Graphics (TOG) , volume=

    Brepgen: A b-rep generative diffusion model with structured latent geometry , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  4. [4]

    arXiv preprint arXiv:2411.10848 , year=

    NeuroNURBS: Learning Efficient Surface Representations for 3D Solids , author=. arXiv preprint arXiv:2411.10848 , year=

    Pith/arXiv arXiv

  5. [5]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    DTGBrepGen: A Novel B-rep Generative Model through Decoupling Topology and Geometry , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  6. [6]

    Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    AutoBrep: Autoregressive B-Rep Generation with Unified Topology and Geometry , author=. Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    2025

  7. [7]

    ACM Transactions on Graphics (TOG) , volume=

    Hola: B-rep generation using a holistic latent representation , author=. ACM Transactions on Graphics (TOG) , volume=. 2025 , publisher=

    2025

  8. [8]

    1986 , publisher=

    Topological structures for geometric modeling (Boundary representation, manifold, radial edge structure) , author=. 1986 , publisher=

    1986

  9. [9]

    ACM Transactions on Graphics (TOG) , volume=

    Complexgen: Cad reconstruction by b-rep chain complex generation , author=. ACM Transactions on Graphics (TOG) , volume=. 2022 , publisher=

    2022

  10. [10]

    ISO 10303-42:2022: Industrial automation systems and integration --- Product data representation and exchange --- Part 42: Integrated generic resource: Geometric and topological representation , year =

    2022

  11. [11]

    Advances in neural information processing systems , volume=

    Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation , author=. Advances in neural information processing systems , volume=

  12. [12]

    The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Koch, Sebastian and Matveev, Albert and Jiang, Zhongshi and Williams, Francis and Artemov, Alexey and Burnaev, Evgeny and Alexa, Marc and Zorin, Denis and Panozzo, Daniele , title =. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

  13. [13]

    ACM Transactions on Graphics (TOG) , volume=

    BrepGPT: Autoregressive B-rep Generation with Voronoi Half-Patch , author=. ACM Transactions on Graphics (TOG) , volume=. 2025 , publisher=

    2025

  14. [14]

    Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers , pages=

    Brepdiff: Single-stage b-rep diffusion model , author=. Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers , pages=

  15. [15]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Brepnet: A topological message passing system for solid models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  16. [16]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Self-supervised representation learning for cad , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  17. [17]

    2022 , school=

    Deep learning on geometry representations , author=. 2022 , school=

    2022

  18. [18]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Deepcad: A deep generative network for computer-aided design models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  19. [19]

    International Conference on Machine Learning , pages=

    SkexGen: Autoregressive Generation of CAD Construction Sequences with Disentangled Codebooks , author =. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  20. [20]

    and Willis, Karl D.D

    Xu, Xiang and Jayaraman, Pradeep Kumar and Lambourne, Joseph G. and Willis, Karl D.D. and Furukawa, Yasutaka , title =. Proceedings of the 40th International Conference on Machine Learning , articleno =. 2023 , publisher =

    2023

  21. [21]

    Graphical Models , volume=

    CADTrans: A code tree-guided CAD generative transformer model with regularized discrete codebooks , author=. Graphical Models , volume=. 2025 , publisher=

    2025

  22. [22]

    Advances in Neural Information Processing Systems , volume=

    Text2cad: Generating sequential cad designs from beginner-to-expert level text prompts , author=. Advances in Neural Information Processing Systems , volume=

  23. [23]

    Md Ferdous Alam and Faez Ahmed , journal=. Gen. 2025 , url=

    2025

  24. [24]

    IEEE Transactions on Industrial Informatics , year=

    Img2cad: Conditioned 3-d cad model generation from single image with structured visual geometry , author=. IEEE Transactions on Industrial Informatics , year=

  25. [25]

    2025 , isbn =

    You, Yang and Uy, Mikaela Angelina and Han, Jiaqi and Thomas, Rahul and Zhang, Haotong and Du, Yi and Chen, Hansheng and Engelmann, Francis and You, Suya and Guibas, Leonidas , title =. 2025 , isbn =. doi:10.1145/3757377.3763891 , booktitle =

    arXiv 2025

  26. [26]

    SIGGRAPH Asia 2022 conference papers , pages=

    Reconstructing editable prismatic cad from rounded voxel models , author=. SIGGRAPH Asia 2022 conference papers , pages=

    2022

  27. [27]

    2024 , eprint=

    CAD-MLLM: Unifying Multimodality-Conditioned CAD Generation With MLLM , author=. 2024 , eprint=

    2024

  28. [28]

    ACM Transactions on Graphics (TOG) , volume=

    Sketch2cad: Sequential cad modeling by sketching in context , author=. ACM Transactions on Graphics (TOG) , volume=. 2020 , publisher=

    2020

  29. [29]

    arXiv preprint arXiv:2502.03997 , year=

    Cad-editor: A locate-then-infill framework with automated training data synthesis for text-based cad editing , author=. arXiv preprint arXiv:2502.03997 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2411.05823 , year=

    Flexcad: Unified and versatile controllable cad generation with fine-tuned large language models , author=. arXiv preprint arXiv:2411.05823 , year=

    Pith/arXiv arXiv

  31. [31]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Csgnet: Neural shape parser for constructive solid geometry , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  32. [32]

    ACM Transactions on Graphics (TOG) , volume=

    Inversecsg: Automatic conversion of 3d models to csg trees , author=. ACM Transactions on Graphics (TOG) , volume=. 2018 , publisher=

    2018

  33. [33]

    International Conference on Learning Representations , year=

    Learning to Infer and Execute 3D Shape Programs , author=. International Conference on Learning Representations , year=

  34. [34]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Bsp-net: Generating compact meshes via binary space partitioning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  35. [35]

    Advances in Neural Information Processing Systems , volume=

    Write, execute, assess: Program synthesis with a repl , author=. Advances in Neural Information Processing Systems , volume=

  36. [36]

    Advances in neural information processing systems , volume=

    UCSG-NET-unsupervised discovering of constructive solid geometry tree , author=. Advances in neural information processing systems , volume=

  37. [37]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Capri-net: Learning compact cad shapes with adaptive primitive assembly , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  38. [38]

    Advances in Neural Information Processing Systems , volume=

    D2CSG: Unsupervised learning of compact CSG trees with dual complements and dropouts , author=. Advances in Neural Information Processing Systems , volume=

  39. [39]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Hpnet: Deep primitive segmentation using hybrid representations , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  40. [40]

    ACM SIGGRAPH 2023 conference proceedings , pages=

    Surface and edge detection for primitive fitting of point clouds , author=. ACM SIGGRAPH 2023 conference proceedings , pages=

    2023

  41. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Point2cad: Reverse engineering cad models from 3d point clouds , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  42. [42]

    European conference on computer vision , pages=

    Parsenet: A parametric surface fitting network for 3d point clouds , author=. European conference on computer vision , pages=. 2020 , organization=

    2020

  43. [43]

    ACM Transactions on Graphics (TOG) , volume=

    Split-and-fit: Learning b-reps via structure-aware voronoi partitioning , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  44. [44]

    2nd Summit on Advances in Programming Languages (SNAPL 2017) , pages=

    Programming language tools and techniques for 3D printing , author=. 2nd Summit on Advances in Programming Languages (SNAPL 2017) , pages=. 2017 , organization=

    2017

  45. [45]

    Proceedings of the ACM on Programming Languages , volume=

    Functional programming for compiling and decompiling computer-aided design , author=. Proceedings of the ACM on Programming Languages , volume=. 2018 , publisher=

    2018

  46. [46]

    Advances in Neural Information Processing Systems , volume=

    Computer-aided design as language , author=. Advances in Neural Information Processing Systems , volume=

  47. [47]

    Advances in Neural Information Processing Systems , volume=

    Sketchgen: Generating constrained cad sketches , author=. Advances in Neural Information Processing Systems , volume=

  48. [48]

    The International Conference on Learning Representations (ICLR) , year=

    VITRUVION: A GENERATIVE MODEL OF PARAMETRIC CAD SKETCHES , author=. The International Conference on Learning Representations (ICLR) , year=

  49. [49]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Engineering sketch generation for computer-aided design , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  50. [50]

    IEEE Transactions on Visualization and Computer Graphics , year=

    Mesh2brep: B-rep reconstruction via robust primitive fitting and intersection-aware constraints , author=. IEEE Transactions on Visualization and Computer Graphics , year=

  51. [51]

    ACM Transactions on Graphics (TOG) , volume=

    Clay: A controllable large-scale generative model for creating high-quality 3d assets , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  52. [52]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    CADDreamer: CAD Object Generation from Single-view Images , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  53. [53]

    Tech, Memo , volume=

    Construcive Solid Geometry, Production Automation Project , author=. Tech, Memo , volume=. 1977 , publisher=

    1977

  54. [54]

    Proceedings of the sixth ACM symposium on Solid modeling and applications , pages=

    Partial entity structure: a compact non-manifold boundary representation based on partial topological entities , author=. Proceedings of the sixth ACM symposium on Solid modeling and applications , pages=

  55. [55]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Cad-recode: Reverse engineering cad code from point clouds , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  56. [56]

    and Ritchie, Daniel , title =

    Xu, Xianghao and Peng, Wenzhe and Cheng, Chin-Yi and Willis, Karl D.D. and Ritchie, Daniel , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2021 , pages =

    2021

  57. [57]

    The Twelfth International Conference on Learning Representations , year=

    Finite Scalar Quantization: VQ-VAE Made Simple , author=. The Twelfth International Conference on Learning Representations , year=

  58. [58]

    ACM Transactions on Graphics (TOG) , volume=

    Rfeps: Reconstructing feature-line equipped polygonal surface , author=. ACM Transactions on Graphics (TOG) , volume=. 2022 , publisher=

    2022

  59. [59]

    ACM Transactions on Graphics (TOG) , volume=

    CWF: consolidating weak features in high-quality mesh simplification , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  60. [60]

    ACM Trans

    Dong, Qiujie and Xu, Rui and Wang, Pengfei and Chen, Shuangmin and Xin, Shiqing and Jia, Xiaohong and Wang, Wenping and Tu, Changhe , title =. ACM Trans. Graph. , month = jul, articleno =. 2024 , issue_date =. doi:10.1145/3658171 , abstract =

    doi:10.1145/3658171 2024