HoLa: B-Rep Generation using a Holistic Latent Representation

Daniel Cohen-Or; Duoteng Xu; Hao Zhang; Hui Huang; Xiang Xu; Xingyao Yu; Yilin Liu

REVIEW 3 major objections 5 minor 18 references

HoLa: B-Rep Generation using a Holistic Latent Representation

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read This paper introduces HoLa, a latent representation defined only over surface primitives that encodes an entire B-Rep model, and shows that a single diffusion model built on it generates valid CAD models at 82.68% on DeepCAD, nearly…

desk verdict A genuinely new B-Rep representation that mostly delivers; the validity gains are real and externally checked, but the intersection classifier's loss equations and missing positive-class metrics need fixing before the core mechanism is fully convincing. read the letter →

arxiv 2504.14257 v3 pith:KQ7BOP2Y submitted 2025-04-19 cs.GR cs.CV

Yilin Liu , Duoteng Xu , Xingyao Yu , Xiang Xu , Daniel Cohen-Or , Hao Zhang , Hui Huang This is my paper · ORCID

classification cs.GRcs.CV

keywords boundaryrepresentationB-RepgenerationholisticlatentspaceneuralintersectiondiffusionmodelCADtopologylearningconditional3D

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

HoLa claims that an entire B-Rep model—surface and curve geometry, vertices, and the discrete adjacency topology—can be stored in a single latent space defined only over the surface primitives. The paper's key move is to treat topology as geometry: because a B-Rep curve is necessarily the intersection of two surfaces, a learned neural intersection module can recover each curve and the surface-to-curve adjacency from pairs of surface latent vectors, so no separate curve, vertex, or topology latents are needed. On this latent space the authors train one diffusion model that generates B-Reps unconditionally or from point clouds, single/multi-view images, sketches, and text prompts. If the claim holds, it removes the multi-step pipelines of prior B-Rep generators, cuts redundancies and incoherences, and raises the validity of generated models on DeepCAD from roughly 50% (DeepCAD and BRepGen) to 82.68%. The same representation also improves point-conditioned reconstruction over fitting-based baselines and remains robust to noisy, cropped, or sparse point clouds.

What carries the argument

The holistic latent (HoLa) space is a per-surface VAE latent, each surface latent having a small spatial resolution (2x2) and feature dimension 8 so that orientation information survives pooling. The load-bearing component is the neural intersection module, a cross-attention network applied to ordered pairs of surface latents that outputs (i) a binary intersection prediction and (ii) a curve feature for the shared intersection; a positional encoding of surface order makes the recovered half-curve orientation consistent, since swapping the surface pair reverses the curve direction. The decoder reconstructs surfaces from the surface latents and curves from the recovered curve features, and these are sewn into watertight B-Reps via a B-spline fitting and wire-loop post-process. A single latent diffusion model then generates or denoises the full surface-latent set conditioned on a 256-dimensional vector extracted from noise, images, point clouds, sketches, or text.

What would settle it

Run the trained intersection classifier on B-Reps with coplanar adjacent faces or open sheet bodies and check whether each ground-truth shared curve is recovered uniquely and whether the binary intersection predictions on positive pairs have high precision and recall; a low recall on intersecting surface pairs would show that topology is not fully recoverable from surface latents alone. A second, cleaner check is to measure the Euler characteristic of the wireframe graph of generated models: for watertight genus-g solids it must satisfy V - E + F = 2 - 2g, so systematic violations would refute the claim that topology is encoded consistently.

Watch

Extended reading notes

Core claim

The central claim is that the topological connection between two surfaces is intrinsically tied to the geometry of their intersecting curve, and therefore a full B-Rep can be encoded and decoded using surface latents alone. The paper builds a variational autoencoder whose encoder fuses surface and curve features through a graph neural network and self-attention into per-surface latent vectors; the decoder then applies a neural intersection module to each pair of surface latents, predicting whether they intersect and producing the feature of the shared half-curve from the pair. The loss combines surface and curve reconstruction, a binary intersection classification, and KL regularization. The authors report that this holistic latent supports a single latent diffusion model that, on the DeepCAD benchmark, produces 82.68% valid watertight B-Reps versus 47.74% for BRepGen and 50.82% for DeepCAD, with improved coverage and distributional metrics; on ABC the valid ratio is 60.46% versus 32.68% for BRepGen. The failures they document are inconsistent surface primitives that lead to inaccurate trimming and non-watertight bodies, plus noise from padding variable-length surface sets to a fixed maximum.

Load-bearing premise

Every curve in a solid model of the kind used here must be recoverable from exactly two supporting surface primitives; this holds for closed solids whose neighboring faces are not coplanar, but it fails for adjacent faces lying in the same plane, for open shells, for thin sheet shapes, and for edges where more than two faces meet, since two surfaces then do not define a unique shared curve.

Editorial extensions

If this is right

One diffusion model replaces the separate surface, curve, and vertex generators of prior pipelines; changing the input modality only retrains the diffusion model, not the VAE.
On the filtered DeepCAD benchmark, unconditional validity rises to 82.68% (versus 47.74% for BRepGen and 50.82% for DeepCAD), while coverage increases to 78.87% and MMD and JSD improve.
Point-conditioned generation reaches 98.23% validity when 32 test-time samples are selected by Chamfer distance, and it outperforms fitting-based pipelines (HPNet+Point2CAD, SEDNet+Point2CAD) on topology F-scores.
The same latent handles text, single-view, multi-view, and sketch conditioning, with multi-view inputs giving the best Chamfer distance among image conditions.
Ablations show that the spatial resolution of the surface latent and the half-curve orientation are necessary: removing them drops validity from 94.18% to 79.62% while degrading precision and recall.

Reading between the lines

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

If the intersection-recoverability premise generalizes, the same pairwise-intersection inductive bias could be applied to assembly or joint representations, where each connection is also a lower-order geometric feature shared by exactly two components.
A direct stress test would evaluate the intersection classifier separately on coplanar adjacent faces and on open shells; if curves there are not unique, the representation needs an explicit manifoldness guard rather than just the closed-solid cases in DeepCAD.
Since validity is measured through OpenCascade's checker, part of the reported gain could be sensitive to the checker's tolerance; an independent topological audit of generated B-Reps (for example, the Euler characteristic of the wireframe graph) would pin down how much of the validity gap is topology consistency versus geometric fitting.
The per-surface latent still carries a fixed maximum primitive count through padding; compressing the whole B-Rep to a single global latent, as the paper suggests as future work, would remove that noise source and may further improve validity.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Two things to know. First, this is a real step forward for B-Rep generation: encoding an entire B-Rep into a latent space defined only over surface primitives, then recovering curves and topology with a learned surface-pair intersection module, is genuinely new and it works. The validity jump from ~48% to 82.68% on DeepCAD, checked externally by OpenCascade, is the kind of result that makes people in the field sit up. Second, the paper is mostly honest and the protocol is disclosed, but the load-bearing intersection classifier is under-reported, and one of its loss equations is misprinted.

What's actually new: prior work (BRepGen, SolidGen, ComplexGen, NVDNet) either uses separate latents for surfaces/curves/vertices, autoregressive steps, or explicit curve/vertex encoding. HoLa's bet is that topology can be reformulated as geometry: two surfaces in a closed solid intersect in a curve, so if you can predict which surface pairs intersect and reconstruct the shared curve from their latents, you don't need a separate curve latent. The ablations support this — removing self-attention, spatial resolution, or half-edge orientation degrades validity and reconstruction, which is exactly what you'd expect if the latent is really carrying the directional info needed to recover edges. The point-cloud-conditioned results are also solid, with quantitative comparisons against HPNet/SEDNet+Point2CAD and NVDNet, and the robustness tests on noisy scans are a nice extra.

Soft spots, in proportion. (1) Eqs. 6-7 define the intersection loss using T_SC, the surface-to-curve adjacency, while the classifier operates on surface pairs. As printed, the indices don't line up. This is likely a notation error, but it means the actual training objective of the intersection module is not specified in the paper. A referee should ask for the corrected equation and, more importantly, positive-class precision/recall. The paper notes balanced sampling, so 99.99% accuracy isn't vacuous, but recall on true intersections is the number that verifies the central premise. (2) Text- and image-conditioned results are qualitative only, with no comparison to Text2CAD or Img2CAD even though baselines exist. That's a smaller gap, but it matters for the claim of being the first to do these conditions. (3) The evaluation distribution is filtered to 7-30 surfaces; the paper reports BRepGen's post-filter validity but not its own pre-filter number, so the headline comparison is slightly asymmetric. The filter is disclosed, though, so this is a minor issue. (4) No code or data.

These are all addressable. The central mechanism — recovering edges from surface latents in closed solids — holds up in the evaluated setting. The coplanar-adjacent-face limitation is real but not relevant to the DeepCAD/ABC solids used here.

Who it's for: anyone working on CAD representation learning or generative modeling of B-Reps. It deserves a serious referee; I'd send it out and expect it to come back with minor-to-major revisions focusing on Eqs. 6-7 and the intersection-classifier metrics. I would cite it.

Referee Report

3 major / 5 minor

Summary. The paper introduces HoLa, a holistic latent representation for B-Rep CAD models. The representation encodes a full B-Rep (surfaces, curves, vertices, and topological relations) into a latent space defined only over surface primitives. The central design is a neural intersection module that, given a pair of surface latents, predicts whether the surfaces intersect and recovers the geometry of the shared curve; curves, vertices, and topology are then decoded from these surface-pair predictions. On top of this VAE, the authors train a single latent diffusion model for unconditional generation and for conditional generation from point clouds, images, sketches, and text. On the DeepCAD dataset the method reports an unconditional validity of 82.68% versus 47.74% for BRepGen, and improved topology/geometry F-scores in point-conditioned generation. The paper also presents ablations, robustness experiments with imperfect point clouds, and qualitative results for text-, image-, and sketch-conditioned generation.

Significance. If the central claim holds, this is a substantial advance for B-Rep generation: replacing multi-step, primitive-by-primitive generation with a single unified latent space would reduce training complexity and improve generated model validity. The paper's evaluation has notable strengths: validity is checked externally via OpenCascade, the point-conditioned comparison uses disclosed protocol details, the single-run version of the method is reported alongside the multi-run version, and the ablations support the importance of the half-edge structure and spatial resolution in the latent. The claimed validity gain (82.68% vs 47.74%) is large and is supported by an external checker rather than only self-reported metrics. However, the load-bearing mechanism of the paper—the neural intersection module—is not directly evaluated in a way that establishes its precision/recall, and the training objective for this module is not written consistently in the paper. These gaps are fixable and do not invalidate the approach, but they must be addressed before the central claims can be accepted.

major comments (3)

[Sec. 4.2, Eqs. (6)-(7)] The intersection-classifier loss is not well-formed as printed. T_SC is defined in Sec. 3 as the m-by-n surface-to-curve adjacency matrix, but in Eqs. (6)-(7) it is used as a label for a surface pair (i,j), and the double sum runs over n curves rather than over surface pairs. In addition, I_bc is defined as a classifier on R^16, yet it is called with two surface latent vectors (z_s^(i), z_s^(j)). As written, the objective is dimensionally inconsistent and the actual training target of the intersection classifier is unspecified. Please correct the loss by introducing an explicit surface-pair adjacency matrix (or deriving it from T_SC) and by defining the classifier input as the fused feature produced by the cross-attention module.
[Sec. 5.5] The only direct evidence about the intersection classifier is the statement that intersection accuracy 'remains consistently high (99.99%)'. This number is uninterpretable without reporting precision and recall on the positive (intersecting) class, because intersecting pairs are rare among all surface pairs and a trivial always-negative classifier can achieve almost 100% accuracy on the natural distribution. The paper also does not state whether the 99.99% figure is computed on the balanced training distribution or on the natural inference distribution. Since the decoder's curve and topology recovery depends entirely on this classifier, please report the confusion matrix (or precision/recall) on both balanced and natural pair distributions, and, ideally, the effect of the classifier threshold on final topology F-scores and validity.
[Sec. 1, Sec. 3] The central prior that 'any curve in a B-Rep must be the intersection between two surface primitives' is not valid for all B-Reps it is claimed to cover. Adjacent coplanar faces (smooth edges), tangent-surface contacts, open shells, and non-manifold configurations do not determine a unique shared curve from the two supporting surfaces alone. The paper does not restrict its claim to closed solids with transverse face intersections, nor does it quantify how often the test-set curves violate the premise. Please either narrow the representational claim to the closed-solid, non-coplanar, non-tangent setting, or provide a dataset-scale analysis: report the fraction of DeepCAD curves for which the two adjacent faces have distinct, non-tangent supporting surfaces, and report topology FE/EV recall separately on the curves that satisfy the premise versus those that do not.

minor comments (5)

[Fig. 1 caption] The caption reports '84% vs. ≈50%' for unconditional validity, while the abstract and Table 1 report 82.68% vs 47.74%; please reconcile these numbers.
[Sec. 5.2, Table 2] The Ours32 variant samples 32 models per input point cloud and selects by Chamfer distance to the input, which is a best-of-N test-time augmentation not available to the deterministic baselines. This is disclosed in the text, but the table would benefit from a footnote stating that the comparison is not single-sample-equivalent.
[Sec. 5.3] The text-conditioned results are entirely qualitative; given that Text2CAD provides a baseline and a quantitative protocol, reporting at least one quantitative metric (e.g., the baseline's own evaluation measure) would strengthen the claim that the method produces 'more plausible models' from text.
[Sec. 4.2, Eq. (6)] There appears to be a typographical error in the first term of Eq. (6), where 'I_bc(S_i(i), z(j)_s)' mixes a surface feature with a surface latent; this should be part of the correction to the well-formed objective described in the major comment, not a separate issue.
[General] The paper does not state whether code or trained models will be released. Given the heavy training cost (about 1M iterations for the VAE and 1M for the LDM) and the importance of the exact intersection-module training protocol, a reproducibility statement would be valuable.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; the central representation claim is externally benchmarked, with one non-load-bearing self-citation.

full rationale

The central claim—that a full B-Rep can be encoded into a surface-only latent because curve geometry and topology can be recovered from surface-pair latents—is not circular. The paper's prior (Sec. 1: 'any curve in a B-Rep must be the intersection between two surface primitives') is stated as a geometric observation, not as a result derived from the method. The latent is trained with reconstruction losses over surface and curve geometry plus a supervised intersection-classification loss (Eqs. 4-8), and generated B-Reps are validated externally against OpenCascade's BRep validity checker and against ground-truth Chamfer distances and topology F-scores (Secs. 5.1-5.2). The Ours32 test-time Chamfer selection is disclosed alongside the single-run Ours result, so the headline validity numbers are not hidden selection artifacts. The only same-group self-citation is NVDNet [Liu et al. 2024a], used as a comparison baseline in point-conditioned generation; it is not load-bearing for the representation or generation claims. Two correctness concerns are noted but are not circularity: (i) Sec. 5.5 reports intersection accuracy as 'consistently high (99.99%)' without positive-class precision/recall or the distribution on which it was computed; and (ii) Eqs. 6-7 index the inter-loss by the surface-to-curve matrix T_SC (m x n) while the classifier inputs are surface pairs, so the written loss is dimensionally inconsistent and the actual classifier objective is under-specified. These concerns weaken the empirical support for the topology-recovery premise, but they do not make the derivation reduce to its own inputs.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

Central claim rests on three families of choices: hand-set loss weights and latent dimensions, a fixed evaluation distribution defined by the 7 to 30 surface filter, and the core geometric prior that edges are recoverable from incident surface geometries. The paper provides no formal guarantee for the geometric prior and does not release code, so the empirical validity numbers are the main support. The loss equations for the intersection classifier contain an indexing inconsistency, which weakens the written specification of the training signal.

free parameters (5)

Loss weights w1, w2, w3 = 1, 1e-1, 1e-6
Hand-set weights in Eq. (4) balance reconstruction, intersection classification, and KL regularization; the training behavior of the VAE is sensitive to them.
Maximum number of surface primitives M = Not stated explicitly; dataset filtered to 7 to 30 surfaces
Latents are padded by random repetition to a fixed length M (Sec. 4.3); this choice defines the diffusion model input size and affects validity.
Evaluation distribution filter = 7 to 30 surfaces per model
Filter in Sec. 5.1 removes simple and over-complex models; the headline 82.68 percent validity is defined on this filtered distribution.
OpenCascade validity precision = 0.1
Threshold of the B-Rep validity checker used for all validity numbers in Table 1.
Test-time sampling count for conditional generation = 1 for base, 32 for Ours32
Validity on point-conditioned generation rises from 83.87 percent to 98.23 percent when 32 samples are generated and the best Chamfer-distance match is selected (Table 2).

assumptions (4)

domain assumption Every curve in a B-Rep is the intersection of exactly two surface primitives, and its geometry is determined by those two surfaces.
Stated in Section 1 as the key observation. True for closed manifold solids with non-coplanar adjacent faces; false for coplanar face splits, open shells, and non-manifold edges. The entire surface-only latent design depends on it.
ad hoc to paper A neural intersection module can learn to recover curve geometry and intersection labels from pairs of surface latents.
Assumed learnability of the inductive bias; supported indirectly by final validity and topology scores, but not directly verified due to missing class-separated accuracy.
domain assumption DeepCAD and ABC, after duplicate removal and the 7-to-30-surface filter, are representative B-Rep generation benchmarks.
The evaluation domain defines the reported validity numbers and does not establish generalization to assemblies, freeform surfaces, or sheet bodies.
domain assumption OpenCascade's validity check at precision 0.1 is a fair external criterion and is applied identically to baselines.
Validity is defined by this checker in Sec. 5.1; it is the same external oracle used for all methods in Table 1.

invented entities (1)

HoLa holistic latent space with neural intersection recovery independent evidence
purpose: Encodes full B-Rep geometry and topology in per-surface latent vectors so a single diffusion model can generate valid CAD models from diverse inputs.
The representation makes a falsifiable claim: surface latents alone must decode into valid, watertight B-Reps. OpenCascade validity, topology F-scores, and failure cases (Figs. 7, 14, 17) provide external evidence.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HoLa: B-Rep Generation using a Holistic Latent Representation." pith.science (2026). https://pith.science/paper/KQ7BOP2Y

@misc{pith2026250414257,
  author       = {Pith},
  title        = {Pith review of: HoLa: B-Rep Generation using a Holistic Latent Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KQ7BOP2Y}},
  note         = {Machine review of arXiv:2504.14257}
}

abstract

We introduce a novel representation for learning and generating Computer-Aided Design (CAD) models in the form of $\textit{boundary representations}$ (B-Reps). Our representation unifies the continuous geometric properties of B-Rep primitives in different orders (e.g., surfaces and curves) and their discrete topological relations in a $\textit{holistic latent}$ (HoLa) space. This is based on the simple observation that the topological connection between two surfaces is intrinsically tied to the geometry of their intersecting curve. Such a prior allows us to reformulate topology learning in B-Reps as a geometric reconstruction problem in Euclidean space. Specifically, we eliminate the presence of curves, vertices, and all the topological connections in the latent space by learning to distinguish and derive curve geometries from a pair of surface primitives via a neural intersection network. To this end, our holistic latent space is only defined on surfaces but encodes a full B-Rep model, including the geometry of surfaces, curves, vertices, and their topological relations. Our compact and holistic latent space facilitates the design of a first diffusion-based generator to take on a large variety of inputs including point clouds, single/multi-view images, 2D sketches, and text prompts. Our method significantly reduces ambiguities, redundancies, and incoherences among the generated B-Rep primitives, as well as training complexities inherent in prior multi-step B-Rep learning pipelines, while achieving greatly improved validity rate over current state of the art: 82% vs. $\approx$50%.

Figures

Figures reproduced from arXiv: 2504.14257 by the authors.

**Figure 1.** We learn a holistic latent (HoLa) space for generating boundary representations (B-Reps), which facilitates the design of our first diffusion-based generator to take on a large variety of inputs. Conditional generation results of 3D B-Reps are shown, in two views (bottom), for text, point cloud, multi-view image, and 2D sketch inputs. Unconditional B-Reps generation with HoLa exhibits greatly improved validity over … view at source ↗

**Figure 2.** Disjoint vs. holistic latent B-Reps. Previous neural B-Rep models all learn separate latent spaces and their respective decoders/generators for surface, curve, vertex primitives, even their relations. In contrast, our HoLa representation is learned to blend the geometry and topology of all primitives into one unified latent space, for B-Rep generation. widely used for 3D representation, researchers are turning to mo… view at source ↗

**Figure 3.** A B-Rep model that consists of surfaces, curves, and their topological connections is encoded into a holistic latent space via a feature fusion model. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (23 more)

**Figure 4.** Figure 4: Given a GT B-Rep model, we use convolutional, self-attention and cross-attention layers to fuse the features of surfaces and curves and their connectivity [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Given a pair of sampled surface latent 𝑧𝑠 from our holistic latent space, we use a neural intersection module to identify and recover the intersected curve feature. Then, the decoder network takes the sampled and recovered features to reconstruct the input B-Rep model …

**Figure 6.** Figure 6: Based on the learned latent representation [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: Qualitative results of unconditional generation on DeepCAD dataset by a) DeepCAD, b) BRepGen, c) ours. Models are [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

**Figure 8.** Figure 8: Qualitative results of our unconditional generation result as complexity increases. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

**Figure 9.** Figure 9: Qualitative results of our unconditional generation on ABC dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

**Figure 10.** Figure 10: Novelty analysis of unconditional generation in terms of the light [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]

**Figure 11.** Figure 11: Qualitative results of point-conditioned generation on DeepCAD test set by a) HPNet+Point2CAD, b) SEDNet+Point2CAD, c) ours. Compared to the [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

**Figure 12.** Figure 12: Qualitative results of point-conditioned generation on imperfect point clouds, including a) random cropping, b) adding 2% Gaussian noise, c) masking [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]

**Figure 13.** Figure 13: Qualitative results of point-conditioned generation on scanned point clouds from the structured light scanner. The scanned point cloud usually [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]

**Figure 15.** Figure 15: Relation between the number of runs and the valid ratio of the [PITH_FULL_IMAGE:figures/full_fig_p014_15.png]

**Figure 16.** Figure 16: Qualitative results of text-based conditioned generation. Annotations come from the test set of Text2CAD [Khan et al. 2024]. [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]

**Figure 17.** Figure 17: Failure cases of txt-based conditioned generation. Primitive numbers, detailed shape description and inconsistent regularization, are the main reasons [PITH_FULL_IMAGE:figures/full_fig_p015_17.png]

**Figure 18.** Figure 18: Qualitative results of text-based conditioned generation using [PITH_FULL_IMAGE:figures/full_fig_p016_18.png]

**Figure 19.** Figure 19: Qualitative results of multi-view conditioned generation. [PITH_FULL_IMAGE:figures/full_fig_p017_19.png]

**Figure 20.** Figure 20: Qualitative results of single-view conditioned generation. [PITH_FULL_IMAGE:figures/full_fig_p018_20.png]

**Figure 21.** Figure 21: Qualitative results of sketch-based conditioned generation. [PITH_FULL_IMAGE:figures/full_fig_p019_21.png]

**Figure 22.** Figure 22: Multiple sampled generation results from the same single-view and sketch-based input. [PITH_FULL_IMAGE:figures/full_fig_p020_22.png]

**Figure 23.** Figure 23: Additional results of text-conditioned generation. [PITH_FULL_IMAGE:figures/full_fig_p023_23.png]

**Figure 24.** Figure 24: Additional results of point-conditioned generation. [PITH_FULL_IMAGE:figures/full_fig_p023_24.png]

**Figure 25.** Figure 25: Additional results of text-conditioned generation. [PITH_FULL_IMAGE:figures/full_fig_p024_25.png]

**Figure 26.** Figure 26: Additional results of single-view conditioned generation. [PITH_FULL_IMAGE:figures/full_fig_p024_26.png]

**Figure 27.** Figure 27: Additional results of sketch-conditioned generation. [PITH_FULL_IMAGE:figures/full_fig_p025_27.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 14 canonical work pages

[1]

Publication date: August 2025

2025
[4]

Advanced Engineering Informatics 56 (2023), 101970

A quantitative analysis of parametric CAD model complexity and its relationship to perceived modeling complexity. Advanced Engineering Informatics 56 (2023), 101970. Steven A. Coons

work page 2023
[7]

SolidGen: An Autoregressive Model for Direct B-rep Synthesis. Trans. on Machine Learning Research (2023). Pradeep Kumar Jayaraman, Aditya Sanghi, Joseph G. Lambourne, Karl D. D. Willis, Thomas Davies, Hooman Shayani, and Nigel J. W. Morris

work page 2023
[10]

on Graphics (Proc

Def: Deep Estimation of Sharp Geometric Features in 3D Shapes.ACM Trans. on Graphics (Proc. SIGGRAPH) 41, 4 (2022), 108:1–108:22. Chandrakana Nandi, Anat Caspi, Dan Grossman, and Zachary Tatlock

work page 2022
[11]

Functional programming for compiling and decompiling computer-aided design. Proc. ACM on Programming Languages 2 (2018), 99:1–99:31. OpenCascade

work page 2018
[12]

DINOv2: Learning Robust Visual Features without Supervision. Trans. on Machine Learning Research (2024). Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas

work page 2024
[13]

Computer Graphics Forum 42, 2 (2023), 545–568

Neurosymbolic Models for Computer Graphics. Computer Graphics Forum 42, 2 (2023), 545–568. Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji

work page 2023
[14]

Computer Graphics Forum 41, 2 (2022), 641–676

Neural Fields in Visual Computing and Beyond. Computer Graphics Forum 41, 2 (2022), 641–676. Jingwei Xu, Chenyu Wang, Zibo Zhao, Wen Liu, Yi Ma, and Shenghua Gao. 2024b. CAD-MLLM: Unifying Multimodality-Conditioned CAD Generation With MLLM. arXiv:2411.04954 Xiang Xu, Pradeep Kumar Jayaraman, Joseph George Lambourne, Karl D. D. Willis, and Yasutaka Furukawa

arXiv 2022

Show all 18 references

[16]

arXiv:2408.01437 Fenggen Yu, Qimin Chen, Maham Tanveer, Ali Mahdavi Amiri, and Hao Zhang

Img2CAD: Reverse Engineering 3D CAD Models from Images through VLM-Assisted Conditional Factorization. arXiv:2408.01437 Fenggen Yu, Qimin Chen, Maham Tanveer, Ali Mahdavi Amiri, and Hao Zhang
[17]

ACM Trans

CLAY: A Controllable Large-scale Gener- ative Model for Creating High-quality 3D Assets. ACM Trans. on Graphics (Proc. SIGGRAPH) 43, 4 (2024), 120:1–120:20. Shengdi Zhou, Tianyi Tang, and Bin Zhou

2024
[202]

Xiang Xu, Joseph G

38443–38461. Xiang Xu, Joseph G. Lambourne, Pradeep Kumar Jayaraman, Zhengqing Wang, Karl D. D. Willis, and Yasutaka Furukawa. 2024a. BrepGen: A B-rep Generative Diffusion Model with Structured Latent Geometry. ACM Trans. on Graphics (Proc. SIGGRAPH) 43, 4 (2024), 119:1–119:14...

2024
[2003]

Tianrun Chen, Chunan Yu, Yuanqi Hu, Jing Li, Tao Xu, Runlong Cao, Lanyun Zhu, Ying Zang, Yong Zhang, Zejian Li, and Linyun Sun

On Visual Similarity Based 3D Model Retrieval.Computer Graphics Forum 22, 3 (2003), 223–232. Tianrun Chen, Chunan Yu, Yuanqi Hu, Jing Li, Tao Xu, Runlong Cao, Lanyun Zhu, Ying Zang, Yong Zhang, Zejian Li, and Linyun Sun

2003
[2018]

TOG 37, 6 (2018),

InverseCSG: automatic conversion of 3D models to CSG trees. TOG 37, 6 (2018),

2018
[2020]

ACM Trans

Sketch2CAD: sequential CAD modeling by sketching in context. ACM Trans. on Graphics (Proc. SIGGRAPH Asia) 39, 6 (2020), 164:1–164:14. Pu Li, Jianwei Guo, Xiaopeng Zhang, and Dong-Ming Yan. 2023a. SECAD-Net: Self- Supervised CAD Reconstruction by Learning Sketch-Extrude Operati...

2020
[2021]

ACM Trans

AutoMate: a dataset and learning approach for automatic mating of CAD assemblies. ACM Trans. on Graphics 40, 6 (2021), 227:1–227:18. Benjamin T. Jones, Michael Hu, Milin Kodnongbua, Vladimir G. Kim, and Adriana Schulz

2021
[2022]

ACM Trans

Com- plexGen: CAD reconstruction by B-rep chain complex generation. ACM Trans. on Graphics (Proc. SIGGRAPH) 41, 4 (2022), 129:1–129:18. Jonathan Ho, Ajay Jain, and Pieter Abbeel

2022
[2023]

HG- CAD: Hierarchical Graph Learning for Material Prediction and Recommendation in Computer-Aided Design. J. Computing and Information Science in Engineering 24, 1 (2023), 011007. Shaked Brody, Uri Alon, and Eran Yahav

2023
[2024]

arXiv:2410.03417 Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang

Img2CAD: Conditioned 3D CAD Model Generation from Single Image with Structured Visual Geometry. arXiv:2410.03417 Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang

arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Publication date: August 2025

2025

[2] [4]

Advanced Engineering Informatics 56 (2023), 101970

A quantitative analysis of parametric CAD model complexity and its relationship to perceived modeling complexity. Advanced Engineering Informatics 56 (2023), 101970. Steven A. Coons

work page 2023

[3] [7]

SolidGen: An Autoregressive Model for Direct B-rep Synthesis. Trans. on Machine Learning Research (2023). Pradeep Kumar Jayaraman, Aditya Sanghi, Joseph G. Lambourne, Karl D. D. Willis, Thomas Davies, Hooman Shayani, and Nigel J. W. Morris

work page 2023

[4] [10]

on Graphics (Proc

Def: Deep Estimation of Sharp Geometric Features in 3D Shapes.ACM Trans. on Graphics (Proc. SIGGRAPH) 41, 4 (2022), 108:1–108:22. Chandrakana Nandi, Anat Caspi, Dan Grossman, and Zachary Tatlock

work page 2022

[5] [11]

Functional programming for compiling and decompiling computer-aided design. Proc. ACM on Programming Languages 2 (2018), 99:1–99:31. OpenCascade

work page 2018

[6] [12]

DINOv2: Learning Robust Visual Features without Supervision. Trans. on Machine Learning Research (2024). Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas

work page 2024

[7] [13]

Computer Graphics Forum 42, 2 (2023), 545–568

Neurosymbolic Models for Computer Graphics. Computer Graphics Forum 42, 2 (2023), 545–568. Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji

work page 2023

[8] [14]

Computer Graphics Forum 41, 2 (2022), 641–676

Neural Fields in Visual Computing and Beyond. Computer Graphics Forum 41, 2 (2022), 641–676. Jingwei Xu, Chenyu Wang, Zibo Zhao, Wen Liu, Yi Ma, and Shenghua Gao. 2024b. CAD-MLLM: Unifying Multimodality-Conditioned CAD Generation With MLLM. arXiv:2411.04954 Xiang Xu, Pradeep Kumar Jayaraman, Joseph George Lambourne, Karl D. D. Willis, and Yasutaka Furukawa

arXiv 2022