Pith. sign in

REVIEW 4 major objections 4 minor 77 references

DreamCAD shows that editable CAD geometry can be learned from raw point supervision alone.

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

2026-08-02 18:41 UTC pith:3XQFFTQ3

load-bearing objection DreamCAD's point-supervised Bézier patch generation is a real advance, but the 'editable in standard CAD' claim is undercut by the paper's own admission that the STEP files lack BRep topology. the 4 major comments →

arxiv 2603.05607 v2 pith:3XQFFTQ3 submitted 2026-03-05 cs.CV cs.AI

DreamCAD: Scaling Multi-modal CAD Generation using Differentiable Parametric Surfaces

Mohammad Sadil Khan , Muhammad Usama , Rolandos Alexandros Potamias , Didier Stricker , Muhammad Zeshan Afzal , Jiankang Deng , Ismail Elezi This is my paper
classification cs.CV cs.AI
keywords CAD generationBézier surfacesdifferentiable tessellationpoint-level supervisiontext-to-CADBRep topologyflow matchingCAD captioning
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.

The paper's central claim is that CAD generation can be scaled by decoupling geometry from topology: instead of training on design histories or BRep annotations, DreamCAD represents each shape as a set of C0-continuous rational Bézier patches, tessellates them differentiably, and supervises the result with point clouds alone. This lets it train on over a million unannotated meshes from public datasets and generate editable surface geometry from text, images, or point clouds. The authors report state-of-the-art reconstruction accuracy on ABC and Objaverse, with Chamfer Distance reduced by up to 70 percent and more than 75 percent user preference, plus a new one-million-caption dataset for text-to-CAD. A sympathetic reader would take away that scalable, editable CAD generation no longer depends on expensive CAD-specific annotations.

Core claim

DreamCAD learns, from point-level supervision only, to output CAD geometry as a set of bicubic rational Bézier patches with shared boundary control points, which guarantees C0 continuity between adjacent patches. A sparse-voxel VAE encodes a mesh into structured latents; a decoder refines control points and weights; differentiable tessellation converts the patches to meshes so a Chamfer loss can compare against the raw point cloud. Conditional generation uses two flow-matching transformers in a coarse-to-fine scheme — first a voxel grid, then per-voxel latent features, then the final parametric surface. The exported STEP files do not carry full BRep topology, but the authors show the patch g

What carries the argument

The load-bearing object is the bicubic rational Bézier patch, defined by a 4x4 grid of control points and non-negative weights, evaluated with Bernstein basis functions. Because the patch is differentiable in its parameters, uniform tessellation into triangles permits end-to-end gradient flow from a Chamfer loss. C0 continuity is enforced structurally rather than by optimization: the initial surface comes from flood-filling a sparse voxel grid, each quad becomes a patch with unit weights, and adjacent patches literally share boundary control points, so the decoder's deformation updates are averaged across the shared point.

Load-bearing premise

The claim that the exported STEP files are directly editable in standard CAD software presumes that a set of Bézier patches without full BRep topology can be edited normally, whereas mainstream CAD tools often treat such imports as dumb solids.

What would settle it

Open a DreamCAD-exported STEP file in a mainstream CAD package and attempt a routine edit—move a hole, change a fillet, or run a solidity and watertight check. If the model fails the solidity check or cannot be edited with ordinary feature tools, the paper's editability claim does not hold for standard workflows, even if the geometry looks correct.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Text-to-CAD can scale beyond small design-history datasets by training on unannotated meshes, opening up millions of existing 3D models as training data.
  • Image- and point-conditioned CAD reconstruction improves Chamfer Distance by up to 70 percent and achieves zero or near-zero invalidity ratios in the reported benchmarks.
  • A compact parametric patch output can serve as a geometric prior for downstream topology recovery, as demonstrated by the paper's 99.2% valid-model result after fine-tuning a language model.
  • Numerically constrained features such as hole counts and gear-tooth counts can be preserved when the text-to-image stage is fine-tuned on the newly released caption dataset.
  • The coarse-to-fine generation pipeline runs in roughly 15 seconds for image and point inputs and 30 seconds for text inputs, making interactive CAD prototyping plausible.

Where Pith is reading between the lines

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

  • If the STEP editability gap is filled by post-processing topology recovery, DreamCAD's geometry stage could plug into existing CAD pipelines as an automatic reverse-engineering front end.
  • The decoupling principle suggests future generative models could treat topology as a separate learned stage, potentially extending editable outputs to assemblies and freeform objects beyond traditional mechanical parts.
  • CADCap-1M's metadata-augmented captions, which include part names, hole counts, and dimension ratios, could be reused to train other multimodal CAD models, not just DreamCAD.
  • Because training relies on Chamfer loss over sampled points and a 32^3 voxel grid, very thin or high-curvature features may define the practical boundary of the approach; a targeted test on such shapes would clarify the limits.

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 / 4 minor

Summary. DreamCAD proposes a two-stage pipeline for CAD generation: a VAE that encodes sparse voxels with local visual features into structured latents and decodes them into C0-continuous rational Bézier patches via differentiable tessellation, trained with Chamfer loss on 1.3M unannotated meshes; and coarse-to-fine flow-matching generators conditioned on text, image, or point clouds. The paper also releases CADCap-1M, 1M+ GPT-5 captions, and reports large improvements over design-history and UV-based baselines on ABC/Objaverse, plus a topology-recovery experiment.

Significance. The contribution is potentially important: if the geometric reconstruction is robust, it opens large-scale training for parametric CAD geometry without BRep annotation. The C0-continuity-by-construction via shared boundary control points is clean, and the ablations on regularizers and voxel resolution are useful. The paper is also transparent about an important limitation in §5.4 ('outputs still lack complete CAD topology'). The main unresolved issues concern the claimed STEP editability and the fairness of the SOTA comparisons; these affect the paper's framing rather than the soundness of the core geometric mechanism.

major comments (4)
  1. [Abstract / §1] The abstract and introduction claim the outputs are 'exportable as STEP files and editable in standard CAD software.' However, §1 immediately states the STEP files 'do not contain industry-standard CAD topology,' and §5.4 repeats that they 'still lack complete CAD topology.' Without face-edge-vertex connectivity, imported STEP patches typically form a 'dumb solid' in mainstream CAD kernels, where feature-based editing is unavailable. 'Editable via control point and weight manipulation' is not demonstrated in any named CAD package. This is a load-bearing overclaim: please either remove or substantially qualify the editability claim, or provide a concrete demonstration (e.g., opening the exported STEP in FreeCAD/SolidWorks, modifying a control point, and showing the resulting feature tree).
  2. [§5.1, Table 2] The SOTA claim is weakened by unequal training data. DreamCAD trains on 1.3M meshes, while DeepCAD is trained on DeepCAD-170K, BRepDiff on the ABC subset, and other baselines on much smaller corpora. The baselines paragraph acknowledges a 'data-scale disparity' and promises additional results on the DeepCAD dataset in the supplementary, but the supplementary material contains no such comparison. Without a matched-scale or same-data experiment, the reported CD reductions (up to 70%) cannot be attributed to the architecture. Please report the promised DeepCAD-scale results or explicitly frame the comparison as 'at our training scale.'
  3. [§5.2 / §4.2] CADCap-1M quality is evaluated by GPT-5 judging captions that GPT-5 itself generated (98.31% correct), and the text-to-CAD evaluation in §5.1 also uses GPT-5 as the judge for a model whose text-to-image stage was fine-tuned on GPT-5-generated captions. This introduces circularity for the caption-quality and text-to-CAD preference claims. The user studies mitigate the concern, but they are reported only as aggregate percentages (75–95%) with no breakdown or statistical detail. Please report per-task user study results and, if possible, have an independent judge (or at least a different VLM) re-evaluate a random subset.
  4. [Table 2, IR column] DreamCAD's Invalidity Ratio is reported as 0.00 on all rows, but DreamCAD outputs are always differentiable Bézier patches rather than BRep solids; the IR metric, defined as 'fraction of non-convertible BRep outputs,' is not applied to DreamCAD in the same sense as to the baselines. The zero IR is therefore trivially achieved and should not be presented as comparable to BRep-generation baselines. Please clarify this in the metric definition and in the discussion of Table 2.
minor comments (4)
  1. [Supplementary §10.A] The text says 'we adopt the same weight initialization scheme and KL weighting as in Trellis [65],' but reference [65] is cited as Structured 3D Latents (SLAT), not Trellis. Please correct the name to avoid confusion.
  2. [§5.3, Table 3a] The ablation reports that adding G1+Laplacian increases CD from 0.0210 to 0.0259 while reducing Laplacian loss. The text says this 'maintains strong geometric accuracy,' but no variance or statistical significance is reported. A brief comment on the CD/Lap trade-off would help.
  3. [§5.1, Baselines] The statement 'we additionally report results on the DeepCAD dataset in the supplementary' has no corresponding section in the provided supplementary; either add the table or remove the sentence.
  4. [Figure 9] The figure labels the topology-recovery model as 'Qwen3' while the text (§5.4) says Qwen3-4B; please make the naming consistent.

Circularity Check

2 steps flagged

Core geometry pipeline is self-contained; only mild, non-load-bearing circularity in GPT-5-based caption and preference evaluation.

specific steps
  1. other [Section 4.2 (CADCap-1M Dataset) and Section 5.2 (Caption Quality)]
    "We leverage GPT-5 to generate high-quality captions for 1M+ models from existing large-scale CAD datasets. ... Overall, 95.8% (user) and 98.31% (GPT-5) of captions are judged correct."

    GPT-5 both generates the captions in CADCap-1M and is the judge that rates those captions as correct. The 98.31% GPT-5 correctness figure is therefore a self-consistency score rather than an independent measurement. The user-study figure (95.8%) is independent and corroborates the claim, so this circularity affects only the GPT-5 portion of the caption-quality evaluation and is not load-bearing for the main geometric contributions.

  2. other [Section 5.1 (Multimodal Generation Evaluation, Metrics paragraph)]
    "For text- and image-to-CAD, visual alignment is measured on 5k and 1k samples through GPT-5 [63] and user studies by 14 CAD-experts, respectively. In both settings, GPT-5 and human evaluators are shown multi-view renderings ... and asked to select the model best matching the input (text or image)."

    The text prompts used for Text2CAD are drawn from the same GPT-5-based captioning pipeline as CADCap-1M, and the same model family (GPT-5) is then used to judge which reconstruction best matches the prompt. This makes the GPT-5 preference metric partly a measure of consistency with GPT-5's own captioning style rather than purely external alignment. The concurrently reported user preference (e.g., 85.40 and 83.48 on ABC/Objaverse) is independent and mitigates this, so the circularity is minor and not load-bearing for the central geometric claims.

full rationale

The paper's central derivation is not circular. DreamCAD's parametric surface generation is supervised by point-level Chamfer loss (Eq. 5) against external ground-truth points sampled from unannotated meshes, and C0 continuity is enforced structurally by sharing boundary control points and averaging boundary updates rather than by fitting the evaluation metric. The reported CD/HD/F1 numbers are computed against held-out ground-truth shapes, so the geometric benchmark is externally grounded. The topology-recovery experiment in Section 5.4 uses DreamCAD's own output as input, but it is explicitly presented as a downstream application and as future work, not as evidence for the generative model itself. NURBGen is cited as a prior method/toolkit for the downstream recovery step, not as a uniqueness theorem or as justification for DreamCAD's design, so no self-citation chain is load-bearing. The only identifiable circularity is in the GPT-5-based evaluation: GPT-5 generated CADCap-1M captions and also judged them correct, and GPT-5 serves as the judge for text-to-CAD alignment on prompts produced by the same captioning pipeline. These are self-consistency measurements rather than independent validation. However, the paper also reports user studies with 14 CAD experts (95.8% caption correctness and 75-85% preference), which independently support the same conclusions, so the circularity is mild and non-load-bearing. The STEP/BRep editability overclaim identified by the reader is a correctness/validity risk, not a circularity, and does not affect this score.

Axiom & Free-Parameter Ledger

6 free parameters · 8 axioms · 0 invented entities

The method introduces no new physical entities (no new particles, forces, or conserved quantities). It relies on a set of hand-chosen hyperparameters and dataset-filtering rules that affect the benchmark numbers. The most consequential domain assumption is that C0-continuous patches without full BRep topology count as 'editable CAD,' which the authors themselves qualify in Section 5.4. The GPT-5 evaluation introduces mild circularity for the caption-quality claim, and the dataset filtering rules act as free parameters for the training distribution.

free parameters (6)
  • Voxel grid resolution (32^3) = 32
    Chosen as quality-efficiency trade-off; ablation shows 48/64 improve CD by only 1-5% but triple/quadruple patch count.
  • Loss weights lambda_cd, lambda_g1, lambda_lp = 1e2, 5e-3, 1
    Empirically set; no sensitivity analysis in the paper.
  • Tessellation resolution schedule = (4,4) to (16,16)
    Sigmoid-scheduled during training; chosen without full ablation.
  • Number of control points per patch (4x4) = 4x4 bicubic
    Design choice for the Bezier representation; not systematically varied.
  • Latent dimension z_i = 8
    Chosen following SLAT/Trellis; not ablated.
  • Dataset filtering thresholds (99% cuboids/cylinders, face/edge/vertex counts) = faces>=5, vertices>=10, edges>=10
    Hand-crafted rules to remove trivial models; effectively free parameters that define the training/testing distribution and affect benchmark comparability.
axioms (8)
  • domain assumption Rational Bezier surfaces can represent a sufficient variety of CAD geometry for the target shapes
    The method restricts to bicubic rational Bezier patches; complex CAD surfaces with exact conic sections or trimmed patches may not be faithfully represented. Assumed throughout, especially in Section 3-4.
  • domain assumption Chamfer distance between uniformly sampled surface points is a sufficient training signal for producing editable CAD surfaces
    Used as the only geometric loss (Eq. 5). It does not explicitly penalize self-intersections or non-manifold geometry; regularization is added ad hoc to mitigate visible artifacts.
  • domain assumption C0 continuity via shared boundary control points suffices for the claimed editability and practical CAD use
    The paper claims the output is 'editable in standard CAD software' but later admits the STEP files lack industry-standard topology. This assumption is load-bearing for the practical contribution and is weakened in Section 5.4.
  • domain assumption GPT-5-generated captions from rendered views plus metadata are high-quality and reliable
    CADCap-1M is built using GPT-5 captions; caption quality is evaluated partly by GPT-5 itself, which introduces circularity. User studies provide some independent support.
  • standard math Flood-fill internal quad removal on a 32^3 voxel grid yields a surface quad topology appropriate for representing the object
    The initial surface patch layout is determined by this algorithm; its assumptions about voxel connectivity and topology are standard but may fail for thin features or holes at the voxel resolution.
  • domain assumption The flow-matching framework (following SLAT/Trellis) can generate structured latents for this CAD domain
    The conditional generation pipeline essentially applies SLAT's flow-matching to CAD shapes; this assumes the latent space learned by the VAE is smooth and amenable to flow matching, which is not separately proven.
  • domain assumption DINOv2 embeddings and rendered normal images provide useful local geometric cues
    The VAE encoder fuses these features; their effectiveness for CAD geometry is an empirical assumption, not derived.
  • standard math OpenCascade conversion of BRep to meshes is faithful and does not introduce significant error
    All training meshes from BRep datasets are generated via OpenCascade tessellation; the quality of this conversion is assumed to be adequate for supervision.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DreamCAD: Scaling Multi-modal CAD Generation using Differentiable Parametric Surfaces." pith.science (2026). https://pith.science/paper/3XQFFTQ3

@misc{pith2026260305607,
  author       = {Pith},
  title        = {Pith review of: DreamCAD: Scaling Multi-modal CAD Generation using Differentiable Parametric Surfaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3XQFFTQ3}},
  note         = {Machine review of arXiv:2603.05607}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Computer-Aided Design (CAD) relies on structured and editable geometric representations, yet existing generative methods are constrained by small annotated datasets with explicit design histories or boundary representation (BRep) labels. Meanwhile, millions of unannotated 3D meshes remain untapped, limiting progress in scalable CAD generation. To address this, we propose DreamCAD, a multi-modal generative framework that directly produces editable BReps from point-level supervision, without CAD-specific annotations. DreamCAD represents each BRep as a set of parametric patches (e.g., B\'ezier surfaces) and uses a differentiable tessellation method to generate meshes. This enables large-scale training on 3D datasets while reconstructing connected and editable surfaces. Furthermore, we introduce CADCap-1M, the largest CAD captioning dataset to date, with 1M+ descriptions generated using GPT-5 for advancing text-to-CAD research. DreamCAD achieves state-of-the-art performance on ABC and Objaverse benchmarks across text, image, and point modalities, improving geometric fidelity and surpassing 75% user preference. Code and dataset will be publicly available.

Figures

Figures reproduced from arXiv: 2603.05607 by Didier Stricker, Ismail Elezi, Jiankang Deng, Mohammad Sadil Khan, Muhammad Usama, Muhammad Zeshan Afzal, Rolandos Alexandros Potamias.

Figure 1
Figure 1. Figure 1: Our proposed DreamCAD (left) is a multimodal generative framework that can reconstruct editable CAD models from text, images, and point clouds using parametric patches. CADCap-1M (right) provides 1M+ GPT-5–generated captions. Abstract Multimodal CAD generation faces a fundamental scala￾bility challenge. Design-history methods are confined to small annotated datasets, while BRep topology is discrete and non… view at source ↗
Figure 2
Figure 2. Figure 2: Bezier surface representation and differentiable tessella- ´ tion. MARVEL-40M+ [53] have driven progress in text-to-3D generation, the CAD domain remains limited by multi￾modal data scarcity. Existing CAD datasets, such as Text2CAD-160K [21], contain design histories with text captions but are small in scale. Moreover, large-scale CAD datasets such as ABC-1M [23] and Automate-440K [19] lack textual or visu… view at source ↗
Figure 3
Figure 3. Figure 3: DreamCAD Overview: (A). Sparse Transformer VAE takes as input mesh, generates active voxels vi with local features fi, from DINOv2 [46] embeddings, normal images, and SDF values and encodes it to generate structured latents zi. These are then decoded into parametric (rational bezier) surfaces and optimized using Chamfer loss. ´ (B). Initial C 0 -continuous Parametric Surface generation from sparse voxels v… view at source ↗
Figure 4
Figure 4. Figure 4: Examples of metadata-augmented captions from [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on Point2CAD (Top-Right), Image2CAD (Bottom-Left) and Text2CAD (Right) tasks. For each task, the first four examples are from the ABC dataset, while the last two from Objaverse dataset. ✗ indicates invalid models. and PointNet++ [48] for point clouds. We use pretrained weights for DINOv2 while training PointNet++ jointly with the Flow Transformer models. Text-to-CAD Generation: While… view at source ↗
Figure 6
Figure 6. Figure 6: Examples showing VAE reconstructions for different [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of captions produced by GPT-5 with and [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of images generated from the pretrained [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: (Left) Topology recovery from patch-based outputs to hybrid NURBS CAD representation [55]. (Right) For multimodal inputs, DreamCAD reconstructions and the corresponding recovered topology. 5.3. Ablation Study Regularizations: We analyze the impact of G1 and Laplacian regularizers on VAE reconstructions. We train the VAE from scratch on 300K samples and evaluating on 15K ABC test shapes (Table 3a and [PITH… view at source ↗
Figure 10
Figure 10. Figure 10: Wordcloud of captions from CADCap-1M 9. More on CADCap-1M For GPT-5–based captioning, we processed over 1M sam￾ples using the batch API (batch size 2k) over three weeks at a total cost of $1800. For metadata augmentation, approxi￾mately 20%, 39%, and 46% of samples in ABC, Automate, and Fusion360 contain part names. In [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Distribution of patch sizes in training datasets. [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Examples of generated captions for [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Examples of generated captions for different types of [PITH_FULL_IMAGE:figures/full_fig_p015_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Examples of generated captions for letter-based (top) and engraving-style (bottom) CAD models 16 [PITH_FULL_IMAGE:figures/full_fig_p016_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: More qualitative results for the point-to-CAD reconstruction task using DreamCAD. Each row shows input point clouds (top) and the corresponding reconstructed CAD models (bottom) generated by DreamCAD across a wide range of shapes, including mechanical components, furniture, utensils, and free-form surfaces. As seen in the examples, DreamCAD successfully recovers clean parametric geometry from input point … view at source ↗
Figure 16
Figure 16. Figure 16: More qualitative results for the image-to-CAD generation task using DreamCAD. Each row shows the ground-truth reference image (top) and the CAD reconstruction produced by DreamCAD (bottom) across a wide range of object categories, including furniture, mechanical parts, consumer products, and free-form designs. As seen in the examples, DreamCAD consistently recovers accurate 3D geometry from a single image… view at source ↗
Figure 17
Figure 17. Figure 17: More qualitative text-to-CAD results on the test set. Each row shows an input text description (bottom) and the corresponding CAD geometry generated by DreamCAD (top). DreamCAD successfully reconstructs shapes from complex text prompts, ranging from mechanical parts (gear assemblies, flange plates, couplings), tools (pyramidal axe head), and structural elements (DIN rail, stepped platforms), to free-form … view at source ↗
Figure 18
Figure 18. Figure 18: Text-to-CAD generation results on GPT-generated prompts. Each example shows a text prompt produced by GPT-5 (bottom) and the corresponding CAD reconstruction (top) by DreamCAD. As shown, DreamCAD generalizes beyond dataset-style prompts and reliably interprets free-form, open-vocabulary instructions which includes consumer electronics (smartphone, computer mouse), household items (frying pan, stool, water… 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

77 extracted references · 5 linked inside Pith

  1. [1]

    Brep boundary and junction detection for cad reverse engi- neering

    Sk Aziz Ali, Mohammad Sadil Khan, and Didier Stricker. Brep boundary and junction detection for cad reverse engi- neering. InICMI. IEEE, 2024. 3, 14

    2024

  2. [2]

    Autodesk Fusion 360: Integrated cad, cam, and cae software.https://www.autodesk.com/ products/fusion-360/, 2024

    Autodesk Inc. Autodesk Fusion 360: Integrated cad, cam, and cae software.https://www.autodesk.com/ products/fusion-360/, 2024. 3

    2024

  3. [3]

    Blender – a 3d modelling and rendering package.https://www.blender.org/,

    Blender Online Community. Blender – a 3d modelling and rendering package.https://www.blender.org/,

  4. [4]

    Shapenet: An information-rich 3d model repository.arXiv:1512.03012,

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

    Pith/arXiv arXiv

  5. [5]

    Abo: Dataset and benchmarks for real-world 3d object understand- ing

    Jasmine Collins, Shubham Goel, Kenan Deng, Achlesh- war Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understand- ing. InCVPR, 2022. 2

    2022

  6. [6]

    Manuel Contero, David P ´erez-L´opez, Pedro Company, and Jorge D. Camba. A quantitative analysis of parametric cad model complexity and its relationship to perceived modeling complexity.Advanced Engineering Informatics, 56, 2023. 2

    2023

  7. [7]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. InCVPR, 2023. 2, 3, 7

    2023

  8. [8]

    On the g1 continu- ity of piecewise b ´ezier surfaces: a review with new results

    Wen-Hui Du and Francis JM Schmitt. On the g1 continu- ity of piecewise b ´ezier surfaces: a review with new results. Computer-Aided Design, 1990. 5

    1990

  9. [9]

    Transcad: A hi- erarchical transformer for cad sequence inference from point clouds

    Elona Dupont, Kseniya Cherenkova, Dimitrios Mallis, Gleb Gusev, Anis Kacem, and Djamila Aouada. Transcad: A hi- erarchical transformer for cad sequence inference from point clouds. InECCV, 2024. 3

    2024

  10. [10]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InICML, 2024. 2, 6

    2024

  11. [11]

    Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 5th edition, 2001

    Gerald Farin.Curves and surfaces for CAGD: a practical guide. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 5th edition, 2001. 2

    2001

  12. [12]

    Foley, Andries van Dam, Steven K

    James D. Foley, Andries van Dam, Steven K. Feiner, and John F. Hughes.Computer graphics: principles and practice (2nd ed.). Addison-Wesley Longman Publishing Co., Inc., USA, 1990. 2

    1990

  13. [13]

    3d-future: 3d fur- niture shape with texture.IJCV, 2021

    Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d fur- niture shape with texture.IJCV, 2021. 2, 5, 6

    2021

  14. [14]

    Cadmium: Fine-tuning code language models for text-driven sequential cad design

    Prashant Govindarajan, Davide Baldelli, Jay Pathak, Quentin Fournier, and Sarath Chandar. Cadmium: Fine-tuning code language models for text-driven sequential cad design. arXiv:2507.09792, 2025. 3

    arXiv 2025

  15. [15]

    Complexgen: Cad reconstruction by b-rep chain complex generation.SIGGRAPH, 2022

    Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. Complexgen: Cad reconstruction by b-rep chain complex generation.SIGGRAPH, 2022. 2, 3, 14

    2022

  16. [16]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 7, 9

    2022

  17. [17]

    Future prospects of computer-aided design (cad) – a review from the perspective of artificial intelligence (ai), extended reality, and 3d printing.Results in Engineering, 2022

    Bonsa Hunde and Abraham Woldeyohannes. Future prospects of computer-aided design (cad) – a review from the perspective of artificial intelligence (ai), extended reality, and 3d printing.Results in Engineering, 2022. 2

    2022

  18. [18]

    Lam- bourne, Karl D.D

    Pradeep Kumar Jayaraman, Aditya Sanghi, Joseph G. Lam- bourne, Karl D.D. Willis, Thomas Davies, Hooman Shayani, and Nigel Morris. Uv-net: Learning from boundary repre- sentations. InCVPR, 2021. 3

    2021

  19. [19]

    Automate: A dataset and learning approach for automatic mating of cad assem- blies.SIGGRAPH, 2021

    Benjamin Jones, Dalton Hildreth, Duowen Chen, Ilya Baran, Vladimir G Kim, and Adriana Schulz. Automate: A dataset and learning approach for automatic mating of cad assem- blies.SIGGRAPH, 2021. 2, 3, 5, 6

    2021

  20. [20]

    Cad-signet: Cad language inference from point clouds us- ing layer-wise sketch instance guided attention

    Mohammad Sadil Khan, Elona Dupont, Sk Aziz Ali, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Cad-signet: Cad language inference from point clouds us- ing layer-wise sketch instance guided attention. InCVPR,

  21. [21]

    Text2cad: Generating sequential CAD designs from beginner-to-expert level text prompts

    Mohammad Sadil Khan, Sankalp Sinha, Sheikh Talha Ud- din, Didier Stricker, Sk Aziz Ali, and Muhammad Zeshan Afzal. Text2cad: Generating sequential CAD designs from beginner-to-expert level text prompts. InNeurIPS, 2024. 2, 3, 5, 7

    2024

  22. [22]

    Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal naviga- tion

    Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal naviga- tion. InCVPR, 2024. 2

    2024

  23. [23]

    Abc: A big cad model dataset for geometric deep learning

    Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. InCVPR, 2019. 2, 3, 5, 6, 7

    2019

  24. [24]

    cadrille: Multi-modal cad reconstruc- tion with online reinforcement learning.arXiv:2505.22914,

    Maksim Kolodiazhnyi, Denis Tarasov, Dmitrii Zhem- chuzhnikov, Alexander Nikulin, Ilya Zisman, Anna V orontsova, Anton Konushin, Vladislav Kurenkov, and 10 Danila Rukhovich. cadrille: Multi-modal cad reconstruc- tion with online reinforcement learning.arXiv:2505.22914,

    arXiv

  25. [25]

    Flux.https://github.com/ black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/ black-forest-labs/flux, 2024. 6

    2024

  26. [26]

    Lambourne, Karl D.D

    Joseph G. Lambourne, Karl D.D. Willis, Pradeep Kumar Jayaraman, Aditya Sanghi, Peter Meltzer, and Hooman Shayani. Brepnet: A topological message passing system for solid models. InCVPR, 2021. 2

    2021

  27. [27]

    Brepdiff: Single-stage b-rep diffusion model

    Mingi Lee, Dongsu Zhang, Cl ´ement Jambon, and Young Min Kim. Brepdiff: Single-stage b-rep diffusion model. InProceedings of the Special Interest Group on Com- puter Graphics and Interactive Techniques Conference Con- ference Papers, 2025. 3, 7

    2025

  28. [28]

    Applications of computer-aided design tech- nology in research, engineering and industry.Computer Aided Design: Technology, Types and Practical Applica- tions, 2012

    Jeremy Li. Applications of computer-aided design tech- nology in research, engineering and industry.Computer Aided Design: Technology, Types and Practical Applica- tions, 2012. 1

    2012

  29. [29]

    Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model

    Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2024. 6

    2024

  30. [30]

    Dtgbrepgen: A novel b-rep generative model through decoupling topology and ge- ometry

    Jing Li, Yihang Fu, and Falai Chen. Dtgbrepgen: A novel b-rep generative model through decoupling topology and ge- ometry. InCVPR, 2025. 3

    2025

  31. [31]

    Cad-llama: Leveraging large language models for computer-aided design parametric 3d model generation

    Jiahao Li, Weijian Ma, Xueyang Li, Yunzhong Lou, Guichun Zhou, and Xiangdong Zhou. Cad-llama: Leveraging large language models for computer-aided design parametric 3d model generation. InCVPR, 2025. 3

    2025

  32. [32]

    Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations

    Pu Li, Jianwei Guo, Xiaopeng Zhang, and Dong-Ming Yan. Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations. InCVPR, 2023. 3

    2023

  33. [33]

    Step1x-3d: Towards high-fidelity and controllable gen- eration of textured 3d assets.arXiv:2505.07747, 2025

    Weiyu Li, Xuanyang Zhang, Zheng Sun, Di Qi, Hao Li, Wei Cheng, Weiwei Cai, Shihao Wu, Jiarui Liu, Zihao Wang, et al. Step1x-3d: Towards high-fidelity and controllable gen- eration of textured 3d assets.arXiv:2505.07747, 2025. 4

    Pith/arXiv arXiv 2025

  34. [34]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maxim- ilian Nickel, and Matthew Le. Flow matching for generative modeling. InICLR, 2023. 5

    2023

  35. [35]

    Point2cad: Reverse engineering cad models from 3d point clouds

    Yujia Liu, Anton Obukhov, Jan Dirk Wegner, and Konrad Schindler. Point2cad: Reverse engineering cad models from 3d point clouds. InCVPR, 2024. 2, 3, 14

    2024

  36. [36]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InICLR, 2019. 6

    2019

  37. [37]

    Scalable 3d captioning with pretrained models

    Tiange Luo, Chris Rockwell, Honglak Lee, and Justin John- son. Scalable 3d captioning with pretrained models. In NeurIPS, 2023. 3, 5

    2023

  38. [38]

    Cadinstruct: A multimodal dataset for natural language-guided cad program synthesis

    Chaofan Lv and Jinsong Bao. Cadinstruct: A multimodal dataset for natural language-guided cad program synthesis. Computer-Aided Design, 2025. 3

    2025

  39. [39]

    Sharp challenge 2023: Solving cad history and parameters recovery from point clouds and 3d scans

    Dimitrios Mallis, Ali Sk Aziz, Elona Dupont, Kseniya Cherenkova, Ahmet Serdar Karadeniz, Mohammad Sadil Khan, Anis Kacem, Gleb Gusev, and Djamila Aouada. Sharp challenge 2023: Solving cad history and parameters recovery from point clouds and 3d scans. overview, datasets, metrics, and baselines. InICCVw, 2023. 2

    2023

  40. [40]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    B Mildenhall, PP Srinivasan, M Tancik, JT Barron, R Ra- mamoorthi, and R Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. InECCV, 2020. 2

    2020

  41. [41]

    Princeton university press,

    John Willard Milnor and David W Weaver.Topology from the differentiable viewpoint. Princeton university press,

  42. [42]

    Laplacian mesh optimization

    Andrew Nealen, Takeo Igarashi, Olga Sorkine, and Marc Alexa. Laplacian mesh optimization. InProceedings of the 4th International Conference on Computer Graphics and Interactive Techniques in Australasia and Southeast Asia,

  43. [43]

    Onshape: Cloud-based cad platform.https: //www.onshape.com/, 2024

    Onshape Inc. Onshape: Cloud-based cad platform.https: //www.onshape.com/, 2024. 3

    2024

  44. [44]

    Opencascade technology (occt)

    Open Cascade SAS. Opencascade technology (occt). https : / / www . opencascade . com / open - cascade-technology/, 2024. 2

    2024

  45. [45]

    Gpt-5.https://openai.com/gpt- 5/,

    OpenAI. Gpt-5.https://openai.com/gpt- 5/,

  46. [46]

    Maxime Oquab, Timoth ´ee Darcet, Theo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Rus- sell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang- Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nico- las Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patri...

    Pith/arXiv arXiv 2023

  47. [47]

    Deepsdf: Learning con- tinuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning con- tinuous signed distance functions for shape representation. InCVPR, 2019. 2

    2019

  48. [48]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space.NeurIPS, 2017

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space.NeurIPS, 2017. 6

    2017

  49. [49]

    Extrudenet: Unsupervised inverse sketch- and-extrude for shape parsing

    Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, and Junzhe Zhang. Extrudenet: Unsupervised inverse sketch- and-extrude for shape parsing. InECCV, 2022. 3

    2022

  50. [50]

    Cad-recode: Reverse engineering cad code from point clouds

    Danila Rukhovich, Elona Dupont, Dimitrios Mallis, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Cad-recode: Reverse engineering cad code from point clouds. InICCV,

  51. [51]

    Parsenet: A parametric surface fitting network for 3d point clouds

    Gopal Sharma, Difan Liu, Subhransu Maji, Evangelos Kalogerakis, Siddhartha Chaudhuri, and Radom ´ır M ˇech. Parsenet: A parametric surface fitting network for 3d point clouds. InECCV, 2020. 3, 14

    2020

  52. [52]

    Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials

    Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahen- dra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and David Novotny. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. In NeurIPS, 2024. 3, 6

    2024

  53. [53]

    Marvel-40m+: Multi-level visual elabo- ration for high-fidelity text-to-3d content creation

    Sankalp Sinha, Mohammad Sadil Khan, Muhammad Us- ama, Shino Sam, Didier Stricker, Sk Aziz Ali, and Muham- mad Zeshan Afzal. Marvel-40m+: Multi-level visual elabo- ration for high-fidelity text-to-3d content creation. InCVPR,

  54. [54]

    Stefan Stojanov, Anh Thai, and James M. Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. InCVPR, 2021. 2, 6

    2021

  55. [55]

    Nurbgen: High- fidelity text-to-cad generation through llm-driven nurbs modeling

    Muhammad Usama, Mohammad Sadil Khan, Didier Stricker, and Muhammad Zeshan Afzal. Nurbgen: High- fidelity text-to-cad generation through llm-driven nurbs modeling. InAAAI, 2026. 2, 3, 7, 9, 14

    2026

  56. [56]

    Point2cyl: Reverse engineering 3d objects from point clouds to extrusion cylinders

    Mikaela Angelina Uy, Yen yu Chang, Minhyuk Sung, Purvi Goel, Joseph Lambourne, Tolga Birdal, and Leonidas Guibas. Point2cyl: Reverse engineering 3d objects from point clouds to extrusion cylinders. InCVPR, 2022. 3

    2022

  57. [57]

    Text- to-cad generation through infusing visual feedback in large language models

    Ruiyu Wang, Yu Yuan, Shizhao Sun, and Jiang Bian. Text- to-cad generation through infusing visual feedback in large language models. InICML, 2025. 7

    2025

  58. [58]

    Cad-gpt: Synthesising cad construction sequence with spatial reasoning-enhanced mul- timodal llms

    Siyu Wang, Cailian Chen, Xinyi Le, Qimin Xu, Lei Xu, Yanzhou Zhang, and Jie Yang. Cad-gpt: Synthesising cad construction sequence with spatial reasoning-enhanced mul- timodal llms. InAAAI, 2025. 3

    2025

  59. [59]

    3d representation methods: A survey

    Zhengren Wang. 3d representation methods: A survey. arXiv:2410.06475, 2024. 2

    Pith/arXiv arXiv 2024

  60. [60]

    Karl D. D. Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G. Lambourne, Armando Solar-Lezama, and Wojciech Matusik. Fusion 360 gallery: A dataset and en- vironment for programmatic cad construction from human design sequences.SIGGRAPH, 2021. 2, 3, 5

    2021

  61. [61]

    Differentiable rendering of parametric geometry.SIGGRAPH, 2023

    Markus Worchel and Marc Alexa. Differentiable rendering of parametric geometry.SIGGRAPH, 2023. 3, 4

    2023

  62. [62]

    Deepcad: A deep generative network for computer-aided design models

    Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. InICCV, 2021. 2, 3, 7

    2021

  63. [63]

    Gpt- 4v(ision) is a human-aligned evaluator for text-to-3d genera- tion

    Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt- 4v(ision) is a human-aligned evaluator for text-to-3d genera- tion. InCVPR, 2024. 7

    2024

  64. [64]

    3d shapenets: A deep representation for volumetric shapes

    Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Lin- guang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 5

    2015

  65. [65]

    Structured 3d latents for scalable and versatile 3d gen- eration

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d gen- eration. InCVPR, 2025. 2, 4, 5, 14

    2025

  66. [66]

    Text-to-cadquery: A new paradigm for cad generation with scalable large model ca- pabilities.arXiv:2505.06507, 2025

    Haoyang Xie and Feng Ju. Text-to-cadquery: A new paradigm for cad generation with scalable large model ca- pabilities.arXiv:2505.06507, 2025. 3, 7

    Pith/arXiv arXiv 2025

  67. [67]

    Brepgen: A b-rep generative diffusion model with structured latent geometry.SIGGRAPH, 2024

    Xiang Xu, Joseph Lambourne, Pradeep Jayaraman, Zhengqing Wang, Karl Willis, and Yasutaka Furukawa. Brepgen: A b-rep generative diffusion model with structured latent geometry.SIGGRAPH, 2024. 2, 3

    2024

  68. [68]

    Qwen3 technical report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. InArxiv,

  69. [69]

    Img2cad: Re- verse engineering 3d cad models from images through vlm- assisted conditional factorization.arXiv:2408.01437, 2024

    Yang You, Mikaela Angelina Uy, Jiaqi Han, Rahul Thomas, Haotong Zhang, Yi Du, Hansheng Chen, Francis Engel- mann, Suya You, and Leonidas Guibas. Img2cad: Re- verse engineering 3d cad models from images through vlm- assisted conditional factorization.arXiv:2408.01437, 2024. 2

    arXiv 2024

  70. [70]

    Clay: A controllable large-scale generative model for creating high-quality 3d assets.SIGGRAPH, 2024

    Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets.SIGGRAPH, 2024. 3, 6

    2024

  71. [71]

    Hunyuan3d 2.0: Scaling diffu- sion models for high resolution textured 3d assets generation

    Zibo Zhao, Zeqiang Lai, Qingxiang Lin, Yunfei Zhao, Haolin Liu, Shuhui Yang, Yifei Feng, Mingxin Yang, Sheng Zhang, Xianghui Yang, et al. Hunyuan3d 2.0: Scaling diffu- sion models for high resolution textured 3d assets generation. CoRR, 2025. 2, 4

    2025

  72. [72]

    Cadparser: a learning approach of sequence modeling for b-rep cad

    Shengdi Zhou, Tianyi Tang, and Bin Zhou. Cadparser: a learning approach of sequence modeling for b-rep cad. In IJCAI, 2023. 2, 3, 5

    2023

  73. [73]

    Intelligent cad 2.0.Visual Informatics, 2024

    Qiang Zou, Yingcai Wu, Zhenyu Liu, Weiwei Xu, and Shum- ing Gao. Intelligent cad 2.0.Visual Informatics, 2024. 1 12 : Scaling Multi-modal CAD Generation using Differentiable Parametric Surfaces Supplementary Material

    2024

  74. [74]

    Visual Feature Generation for Sparse V oxels:We first normalize each mesh into the range[−0.5,0.5] 3 before vox- elization

    More on Data Preparation A. Visual Feature Generation for Sparse V oxels:We first normalize each mesh into the range[−0.5,0.5] 3 before vox- elization. To generate150multi-view images, we use three complementary camera trajectories that jointly provide full coverage of the object. (1)Azimuth sweep:50images are rendered by rotating the camera around the ob...

  75. [75]

    Stem-shaped connec- tor . . . six leaf-like fins

    More on CADCap-1M For GPT-5–based captioning, we processed over 1M sam- ples using the batch API (batch size 2k) over three weeks at a total cost of$1800. For metadata augmentation, approxi- mately 20%, 39%, and 46% of samples in ABC, Automate, and Fusion360 contain part names. In Table 4, we provide summary statistics of the CADCap- 1M captions. The mean...

  76. [76]

    Parameter counts for all components of DreamCAD

    More Experimental Results V AE Sparse Structure Coarse Fine-Grained Image Point Image Point 71M 133M 354M 280M 400M 325M Table 5. Parameter counts for all components of DreamCAD. A. Training Details:For V AE training, we adopt the same weight initialization scheme and KL weighting as in Trel- lis [65]. The output layer predicting the deformation vector is...

  77. [77]

    Prusa I3 MK3 MMU

    Discussion on Future Research Direction DreamCAD addresses the first and one of most challeng- ing stages of scalable CAD generation: learning accurate geometric reconstruction from large-scale unstructured 3D data without CAD-specific annotations. A natural question is whetherpatch-based parametric surfaces are sufficient for real industrial workflows. W...

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.