Pith. sign in

REVIEW 3 major objections 6 minor 32 references

A vision-language model can read raster orthographic drawings and write CadQuery code that rebuilds editable 3D CAD solids more accurately than prior systems.

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 · grok-4.5

2026-07-13 05:59 UTC pith:QJQ7I7HH

load-bearing objection Clean empirical win on a real CAD pain point: orthographic rasters to executable CadQuery via SFT+geometry RL, plus an open drawing generator that actually includes hidden lines and dimensions. the 3 major comments →

arxiv 2607.08891 v1 pith:QJQ7I7HH submitted 2026-07-09 cs.CE

Ortho2CAD: 3D CAD generation from orthographic drawings using vision language models

Aditya Joglekar , Amit Regmi , Kenji Shimada , Levent Burak Kara This is my paper
classification cs.CE
keywords orthographic drawingsCadQueryvision-language models3D CAD reconstructionsupervised fine-tuningreinforcement learningparametric CADSTEP models
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.

Most design intent still travels as raster multi-view orthographic drawings, yet manufacturing and analysis need parametric, editable 3D CAD. This paper claims that a vision-language model can close that gap by emitting CadQuery programs that compile into solids. When ground-truth CadQuery exists, the model is fine-tuned by ordinary next-token supervision; when only STEP geometry exists, it is optimized by reinforcement learning whose sole reward is the intersection-over-union of the executed solid against the true solid. A new pythonOCC pipeline supplies the missing training images—first-angle projections with dashed hidden lines and bounding-box dimensions—at scale from public STEP repositories. On DeepCAD and Fusion 360 evaluation subsets the resulting model produces 100 percent valid executable code and raises mean solid IoU by more than 7 percent relative to the next-best baseline. If the approach holds, the image drawings already circulating in industry could become direct sources of editable CAD.

Core claim

Ortho2CAD establishes that conditioning a modern vision-language model on three-view dimensioned orthographic drawings, then training it with supervised token likelihood where CadQuery labels exist and with sequence-level group-relative geometric IoU rewards where they do not, yields fully executable parametric CAD whose solids match ground-truth geometry more closely than open-source, closed-source, and domain-specialized image-to-CadQuery baselines.

What carries the argument

Dual-regime training of a VLM: supervised fine-tuning on paired orthographic drawings and CadQuery when labels exist, followed by Dr. GRPO-style sequence-level reinforcement learning driven only by solid IoU rewards when labels are absent, enabled by a pythonOCC generator of standardized first-angle drawings.

Load-bearing premise

Clean synthetic first-angle drawings that show only three bounding-box dimensions and come from sketch-and-extrude parts are assumed to be a good enough proxy for the incomplete, fully dimensioned raster drawings used in real manufacturing.

What would settle it

Evaluate the released model on a held-out collection of real industrial multi-view raster drawings that contain missing views, full feature dimensioning, and operations beyond sketch-and-extrude, and check whether mean solid IoU still exceeds the best baseline while code validity remains near 100 percent.

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

If this is right

  • Where CadQuery labels exist, orthographic drawings become an effective conditioning signal for parametric CAD synthesis.
  • Where only STEP geometry exists, executable-geometry rewards can adapt a model without any target code.
  • Large orthographic drawing corpora can be generated automatically from any public STEP repository.
  • Code validity can reach 100 percent while mean solid IoU still improves over specialized and general VLMs.
  • The output remains editable CadQuery rather than meshes or voxels, so it stays usable in ordinary CAD workflows.

Where Pith is reading between the lines

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

  • The same geometry-reward loop could absorb drawings that carry full feature dimensions and section views once richer synthetic labels are available.
  • A hybrid inference path that uses the fine-tuned model for a strong first pass and a larger general model for iterative correction may close remaining IoU gaps on thin or multi-feature parts.
  • Because training data are restricted to sketch-and-extrude, new CadQuery corpora will be required before the method covers fillets, lofts, or freeform B-reps common in production.
  • Releasing the drawing generator lowers the barrier for others to test drawing-to-CAD pipelines on larger repositories such as ABC.

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

3 major / 6 minor

Summary. The paper presents Ortho2CAD, a VLM pipeline that maps rasterized first-angle orthographic drawings (three views with dashed hidden lines and three bounding-box dimensions) to executable CadQuery code and thus editable 3D CAD. When GenCAD-Code labels exist (DeepCAD), the authors supervised-fine-tune Qwen3-VL-8B-Instruct with standard cross-entropy (Eq. 1). When only STEP geometry is available (Fusion 360 Reconstruction), they initialize from the SFT model and optimize a Dr. GRPO / GSPO-style sequence-level objective with mean-centered group advantages and solid IoU reward (Eqs. 2–6). A pythonOCC generator produces the drawings at scale. On 100-example test subsets matching prior CAD-Coder protocol, the model reports 100% valid executable code and mean IoU of 0.7922 (DeepCAD SFT) and 0.5601 (Fusion RL), exceeding the listed baselines by roughly 7–8% relative IoU.

Significance. If the reported gains hold, the work supplies a practical, open pathway from the dominant industrial communication format (raster orthographic drawings) to editable parametric CAD, covering both code-supervised and code-free regimes. Strengths that should be credited include: (i) a public pythonOCC drawing generator with hidden lines and key dimensions (>150k samples), (ii) explicit dual-regime training (SFT + geometry-grounded RL) with fully stated objectives and hyperparameters, (iii) 100% code validity on the evaluation subsets, and (iv) released code. These elements make the empirical claim reproducible and useful as a baseline for drawing-to-CAD research, even though the drawings remain synthetic and sketch-and-extrude-only.

major comments (3)
  1. §3.1 and §5 correctly note that drawings carry only three bounding-box dimensions and that sources are limited to sketch-and-extrude STEP (DeepCAD / Fusion 360). The central claim of superiority is therefore scoped to this synthetic proxy. The manuscript should state more sharply in the abstract and introduction that transfer to fully dimensioned industrial drawings, incomplete views, or richer B-rep operations is untested, so that the 7% relative IoU claim is not read as a general manufacturing result.
  2. Tables 3–4 and §3.2–3.3 evaluate only the same 100-example subsets used by CAD-Coder, citing compute cost. While consistent with prior work, this leaves ranking variance and confidence intervals unreported. A major revision should either (a) report IoU mean/std over the full official test splits (or at least a substantially larger random subset) or (b) provide bootstrap/CI estimates on the 100-example sets so that the 7–8% relative gains can be assessed for statistical stability.
  3. §3.3 and Table 4: the RL model is initialized from DeepCAD SFT and then trained on Fusion geometry rewards. The large jump from Ortho2CAD DeepCAD SFT (IoU 0.3697) to Ortho2CAD RL (0.5601) is the key evidence that geometry-grounded RL compensates for missing code labels. An ablation that starts RL from the untuned Qwen3-VL backbone (or reports intermediate checkpoints) is needed to separate the contribution of the SFT warm-start from the RL objective itself; without it the load-bearing claim that RL alone enables unsupervised adaptation remains only partially supported.
minor comments (6)
  1. Abstract and §1: “average relative improvement of over 7%” aggregates DeepCAD (7.6%) and Fusion (8.1%); state the two numbers separately for clarity.
  2. §3.2: the fixed prompt is acknowledged as suboptimal; a short note on whether prompt variants were tried would help readers reproduce the SFT baseline.
  3. Figure 2 and IoU protocol: alignment/normalization follows CAD-Coder; a one-sentence restatement of the rigid alignment procedure would make the metric self-contained.
  4. Table 1: “100 test subset for final evaluation same as [3]” is clear, but the exact sampling seed or list should be released with the code for exact reproducibility.
  5. §4.2: GPT-5.2 is called “second best” on Fusion; ensure the same model version and decoding settings are fixed in the released evaluation scripts.
  6. Appendix Figures 7–8 usefully show failure modes (thin sheets, non-enclosed volumes); cross-reference them earlier in §4 when discussing qualitative results.

Circularity Check

0 steps flagged

No significant circularity: IoU and validity claims are empirical outcomes on held-out geometry, not algebraic restatements of training inputs.

full rationale

The paper's load-bearing claims are empirical (100% executable CadQuery; mean solid IoU above listed baselines on 100-example DeepCAD and Fusion 360 subsets). SFT maximizes token likelihood against external GenCAD-Code targets (Eq. 1); RL maximizes sequence-level group-relative advantages of executed IoU against held-out ground-truth STEP (Eqs. 2–6). Neither reward nor metric is defined in terms of the generated code tokens or the synthetic drawing pixels; IoU is a post-execution geometric comparison after normalization/alignment. The authors' own pythonOCC drawing generator supplies the input images, but that is ordinary synthetic-data construction: the model must still invert projections into executable programs, and baselines receive identical inputs. No uniqueness theorem, fitted constant, or self-citation is used to force the reported ranking. Scope limits (sketch-and-extrude only, three bounding-box dimensions, synthetic first-angle views) are transfer concerns, not circular reductions. Score 0 is therefore warranted.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

This is an empirical ML systems paper. Load-bearing content is training recipes, synthetic data conventions, and evaluation choices rather than unproved physical entities. Free parameters are standard optimization hyperparameters. Domain assumptions about CadQuery, IoU, and synthetic orthographic conventions carry the transfer claim; no new particles or forces are postulated.

free parameters (4)
  • SFT learning rate and schedule
    lr=1e-5, 5 epochs, batch 4 with 4 accumulation steps on Qwen3-VL-8B; chosen for stable optimization, not derived.
  • RL learning rate, epochs, group size G, max completion length L
    Table 2: 2 epochs, lr=1e-6, G=8, effective batch 64, L=7680; hand-set for sequence-level Dr. GRPO-style training.
  • Drawing generation timeout (DeepCAD)
    10-second timeout per drawing (§3.1) affects which views appear and mimics missing views; arbitrary operational cutoff.
  • Evaluation subset size (100 examples)
    Final IoU/validity reported on 100-example subsets for cost reasons; subset choice affects the headline relative gains.
axioms (5)
  • domain assumption Executable CadQuery programs are an adequate intermediate for editable parametric CAD reconstruction from drawings.
    Throughout §1–3; evaluation only scores solids after CadQuery execution to STEP.
  • domain assumption Solid IoU after normalization/alignment is a sufficient primary measure of reconstruction quality for ranking models.
    Eq. (2) and evaluation protocol following CAD-Coder [3]; invalid code assigned IoU=0.
  • domain assumption Synthetic first-angle orthographic rasters with dashed hidden edges and three bounding-box dimensions are informative enough training/eval inputs for the industrial drawing task.
    §3.1 explicitly notes incomplete feature dimensioning and sketch-extrude-only sources.
  • ad hoc to paper Mean-centered group advantages without length/variance normalization and sequence-level likelihood updates (Dr. GRPO / GSPO-style) are appropriate for sparse terminal IoU rewards.
    §3.3 design choice; justified by citation to [12,27] but not proven optimal for this task.
  • standard math Autoregressive next-token likelihood and standard cross-entropy SFT improve CadQuery generation when paired codes exist.
    Eq. (1); standard VLM fine-tuning assumption.
invented entities (2)
  • Ortho2CAD (SFT+RL VLM pipeline for orthographic→CadQuery) independent evidence
    purpose: Name the trained system that maps drawing images plus a fixed prompt to CadQuery code under two supervision regimes.
    Method artifact, not a physical postulate; independent evidence is the reported validity/IoU tables and public code claim.
  • pythonOCC orthographic drawing generator (hidden lines + key dimensions) independent evidence
    purpose: Synthesize large-scale training/eval drawings from STEP when public dimensioned drawing corpora are scarce.
    Engineering tool introduced for scale; falsifiable by running the published generator on STEP files.

pith-pipeline@v1.1.0-grok45 · 18001 in / 3558 out tokens · 36277 ms · 2026-07-13T05:59:26.415155+00:00 · methodology

0 comments
read the original abstract

Engineering design intent is often communicated through rasterized orthographic drawings. However, downstream workflows inherently require editable and parametrically defined 3D computer-aided design (CAD) models. To bridge this gap, we introduce Ortho2CAD, a vision-language model (VLM) specifically designed to translate rasterized orthographic drawings directly into editable CadQuery code, which can then be seamlessly converted into 3D CAD models. To train the model effectively, we utilize supervised fine-tuning (SFT) for instances where explicit CadQuery code labels already exist, and we apply geometry-grounded reinforcement learning (RL) to optimize the model in scenarios where ground-truth labels are absent. To enable learning at scale, we create a pythonOCC-based drawing generator that renders first-angle orthographic projections from STEP models, complete with dashed hidden lines and key dimensions. On existing datasets encompassing settings both with and without CadQuery supervision, we generate orthographic drawings and show that our model produces 100% syntactically valid code. Moreover, it achieves a 3D CAD intersection-over-union (IoU) accuracy that surpasses all baselines, with an average relative improvement of over 7% compared directly against the next best performing model. We show that leveraging VLMs with SFT and RL techniques can effectively pave the way forward for orthographic drawing to 3D CAD reconstruction. Our implementation is available at https://github.com/AdityaJoglekar/Ortho2CAD.

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

32 extracted references · 11 linked inside Pith

  1. [1]

    Com- ponent segmentation of engineering drawings using graph convolutional networks

    Zhang, W., Joseph, J., Yin, Y ., Xie, L., Furuhata, T., Ya- makawa, S., Shimada, K., and Kara, L. B., 2023. “Com- ponent segmentation of engineering drawings using graph convolutional networks”.Computers in Industry,147, p. 103885

    2023

  2. [2]

    Cad-coder: Text-to-cad generation with chain-of-thought and geometric reward

    Guan, Y ., Wang, X., Xing, X., Zhang, J., Xu, D., and Yu, Q., 2025. “Cad-coder: Text-to-cad generation with chain-of-thought and geometric reward”.arXiv preprint arXiv:2505.19713

    Pith/arXiv arXiv 2025

  3. [3]

    Cad-coder: An open-source vision-language model for computer-aided design code generation

    Doris, A. C., Alam, M. F., Heyrani Nobari, A., and Ahmed, F., 2025. “Cad-coder: An open-source vision-language model for computer-aided design code generation”. In In- ternational Design Engineering Technical Conferences and Computers and Information in Engineering Conference, V ol. 89220, American Society of Mechanical Engineers, p. V03AT03A031

    2025

  4. [4]

    cadrille: Multi- modal cad reconstruction with online reinforcement learn- ing

    Kolodiazhnyi, M., Tarasov, D., Zhemchuzhnikov, D., Nikulin, A., Zisman, I., V orontsova, A., Konushin, A., Kurenkov, V ., and Rukhovich, D., 2025. “cadrille: Multi- modal cad reconstruction with online reinforcement learn- ing”.arXiv preprint arXiv:2505.22914

    arXiv 2025

  5. [5]

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

    Wu, R., Xiao, C., and Zheng, C., 2021. “Deepcad: A deep generative network for computer-aided design models”. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 6772–6782

    2021

  6. [6]

    Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences

    Willis, K. D., Pu, Y ., Luo, J., Chu, H., Du, T., Lambourne, J. G., Solar-Lezama, A., and Matusik, W., 2021. “Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences”.ACM Transactions on Graphics (TOG),40(4), pp. 1–24

    2021

  7. [7]

    Automatic 3d cad models reconstruction from 2d orthographic drawings

    Zhang, C., Pinqui ´e, R., Polette, A., Carasi, G., De Char- nace, H., and Pernot, J.-P., 2023. “Automatic 3d cad models reconstruction from 2d orthographic drawings”.Computers & Graphics,114, pp. 179–189

    2023

  8. [8]

    FreeCAD Team, 2026. Freecad. Accessed: 2026-02-20

    2026

  9. [9]

    Reinforcement learning-based parametric cad models reconstruction from 2d orthographic drawings

    Zhang, C., Polette, A., Pinqui ´e, R., Iida, M., De Charnace, H., and Pernot, J.-P., 2025. “Reinforcement learning-based parametric cad models reconstruction from 2d orthographic drawings”.Computer-Aided Design, p. 103925

    2025

  10. [10]

    Koch, S., Matveev, A., Jiang, Z., Williams, F., Artemov, A., Burnaev, E., Alexa, M., Zorin, D., and Panozzo, D.,

  11. [11]

    Abc: A big cad model dataset for geometric deep learning

    “Abc: A big cad model dataset for geometric deep learning”. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 9601–9611

  12. [12]

    Qwen3- vl technical report

    Bai, S., Cai, Y ., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al., 2025. “Qwen3- vl technical report”.arXiv preprint arXiv:2511.21631

    Pith/arXiv arXiv 2025

  13. [13]

    Understanding r1- zero-like training: A critical perspective

    Liu, Z., Chen, C., Li, W., Qi, P., Pang, T., Du, C., Lee, W. S., and Lin, M., 2025. “Understanding r1- zero-like training: A critical perspective”.arXiv preprint arXiv:2503.20783

    Pith/arXiv arXiv 2025

  14. [14]

    Point-e: a system for generating 3d point clouds from complex prompts (2022)

    Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., and Chen, M., 2022. “Point-e: a system for generating 3d point clouds from complex prompts (2022)”.arXiv preprint arXiv:2212.08751

    Pith/arXiv arXiv 2022

  15. [15]

    T., and Mildenhall, B.,

    Poole, B., Jain, A., Barron, J. T., and Mildenhall, B.,

  16. [16]

    Dreamfusion: Text-to-3d using 2d diffusion

    “Dreamfusion: Text-to-3d using 2d diffusion”.arXiv preprint arXiv:2209.14988

    Pith/arXiv arXiv

  17. [17]

    Meshgpt: Generating triangle meshes with decoder-only transformers

    Siddiqui, Y ., Alliegro, A., Artemov, A., Tommasi, T., Siri- gatti, D., Rosov, V ., Dai, A., and Nießner, M., 2024. “Meshgpt: Generating triangle meshes with decoder-only transformers”. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 19615– 19625

    2024

  18. [18]

    Structured 3d latents for scalable and versatile 3d generation

    Xiang, J., Lv, Z., Xu, S., Deng, Y ., Wang, R., Zhang, B., Chen, D., Tong, X., and Yang, J., 2025. “Structured 3d latents for scalable and versatile 3d generation”. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 21469–21480

    2025

  19. [19]

    Brepgen: A b-rep genera- tive diffusion model with structured latent geometry

    Xu, X., Lambourne, J., Jayaraman, P., Wang, Z., Willis, K., and Furukawa, Y ., 2024. “Brepgen: A b-rep genera- tive diffusion model with structured latent geometry”.ACM Transactions on Graphics (TOG),43(4), pp. 1–14

    2024

  20. [20]

    Hola: B-rep generation us- ing a holistic latent representation

    Liu, Y ., Xu, D., Yu, X., Xu, X., Cohen-Or, D., Zhang, H., and Huang, H., 2025. “Hola: B-rep generation us- ing a holistic latent representation”.ACM Transactions on Graphics (TOG),44(4), pp. 1–25

    2025

  21. [21]

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

    Khan, M. S., Sinha, S., Sheikh, T. U., Stricker, D., Ali, S. A., and Afzal, M. Z., 2024. “Text2cad: Generating sequential cad designs from beginner-to-expert level text prompts”.Advances in Neural Information Processing Sys- tems,37, pp. 7552–7579

    2024

  22. [22]

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

    Govindarajan, P., Baldelli, D., Pathak, J., Fournier, Q., and Chandar, S., 2025. “Cadmium: Fine-tuning code lan- guage models for text-driven sequential cad design”.arXiv 11 preprint arXiv:2507.09792

    arXiv 2025

  23. [23]

    Generating cad code with vision-language models for 3d designs

    Alrashedy, K., Tambwekar, P., Zaidi, Z., Langwasser, M., Xu, W., and Gombolay, M., 2024. “Generating cad code with vision-language models for 3d designs”.arXiv preprint arXiv:2410.05340

    Pith/arXiv arXiv 2024

  24. [24]

    Image2cadseq: Computer-aided design sequence and knowledge inference from product im- ages

    Li, X., and Sha, Z., 2025. “Image2cadseq: Computer-aided design sequence and knowledge inference from product im- ages”. In International Design Engineering Technical Con- ferences and Computers and Information in Engineering Conference, V ol. 89213, American Society of Mechanical Engineers, p. V02BT02A025

    2025

  25. [25]

    Deepseekmath: Pushing the limits of mathematical reasoning in open lan- guage models

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y ., Wu, Y ., et al., 2024. “Deepseekmath: Pushing the limits of mathematical reasoning in open lan- guage models”.arXiv preprint arXiv:2402.03300

    Pith/arXiv arXiv 2024

  26. [26]

    Deepseek- r1: Incentivizing reasoning capability in llms via reinforce- ment learning

    Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma, S., Bi, X., et al., 2025. “Deepseek- r1: Incentivizing reasoning capability in llms via reinforce- ment learning”.arXiv preprint arXiv:2501.12948

    Pith/arXiv arXiv 2025

  27. [27]

    Dapo: An open- source llm reinforcement learning system at scale

    Yu, Q., Zhang, Z., Zhu, R., Yuan, Y ., Zuo, X., Yue, Y ., Dai, W., Fan, T., Liu, G., Liu, L., et al., 2025. “Dapo: An open- source llm reinforcement learning system at scale”.arXiv preprint arXiv:2503.14476

    Pith/arXiv arXiv 2025

  28. [28]

    From intent to execution: Multimodal chain-of- thought reinforcement learning for precise cad code gener- ation

    Niu, K., Yu, H., Chen, Z., Zhao, M., Fu, T., Li, B., and Xue, X., 2025. “From intent to execution: Multimodal chain-of- thought reinforcement learning for precise cad code gener- ation”.arXiv preprint arXiv:2508.10118

    Pith/arXiv arXiv 2025

  29. [29]

    Group sequence policy optimization

    Zheng, C., Liu, S., Li, M., Chen, X.-H., Yu, B., Gao, C., Dang, K., Liu, Y ., Men, R., Yang, A., et al., 2025. “Group sequence policy optimization”.arXiv preprint arXiv:2507.18071

    Pith/arXiv arXiv 2025

  30. [30]

    Introducing GPT-5.2

    OpenAI, 2025. Introducing GPT-5.2. Accessed: 2026-02- 20

    2025

  31. [31]

    Visual in- struction tuning

    Liu, H., Li, C., Wu, Q., and Lee, Y . J., 2023. “Visual in- struction tuning”.Advances in neural information process- ing systems,36, pp. 34892–34916

    2023

  32. [32]

    Cad-recode: Reverse engineering cad code from point clouds

    Rukhovich, D., Dupont, E., Mallis, D., Cherenkova, K., Kacem, A., and Aouada, D., 2025. “Cad-recode: Reverse engineering cad code from point clouds”. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pp. 9801–9811. A APPENDIX We evaluate the success and failure cases for a better un- derstanding of our Ortho2CAD model on the De...

    2025