Vision2CAD: A Visual Agent Harness for Explicit Geometry Referencing and Localization in Parametric CAD Modeling
Xi Cheng, Chenxi Zhai, Hang Cheng, Mingyu Fan, Pingfa Feng, Long Zeng
cs.CV · cs.AI
Quick verdict
Vision2CAD claims top image-to-CAD accuracy and editability via ID-based referencing; evidence is solid but CR-filtered and code is unreleased.
Summary
Vision2CAD is a training-free visual agent that converts a single image into a native parametric CAD model (a FreeCAD document) by coupling a vision-language model (gpt-5.6-sol) with deterministic CAD-kernel operations. It targets three failure modes of prior work: selecting existing geometry, interpreting sketch-plane local coordinates, and constraining sketches to projected external geometry. Mechanisms include an ID-tagged view plus geometry table for explicit entity selection, a bridge converting VLM-predicted view coordinates into sketch coordinates, and projected-edge samples that seed external constraints; depth-image registration, DINOv3-based retrieval, and preview verification complete the loop. The authors also build GERD: 2,700 Onshape parts with per-step alignment of commands, geometry states, and Body/Face/Edge/Vertex IDs. On GERD-EVL (100 test parts) it achieves 100% compilation rate, mIoU 0.4733 (+11.1% over the strongest baseline, which covers only 18/100 parts) and CD 0.0596 (−17.3%); on a 100-part DeepCAD subset, mIoU 0.6987 (+5.6%) and CD 0.0320 (−41.8%). After a uniform 10% extrusion-depth increase it retains 100% CR with the best metrics; six ablations show each component contributes, most from explicit geometry referencing. Limitations: headline comparisons are conditional on unequal success subsets, no significance testing or run variance is reported, DeepCAD evaluation covers only 100 parts, input is single-view, pattern operations are unsupported, and code/data are only promised.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An ID-based referencing interface, a deterministic view-to-sketch coordinate bridge, and projected-edge constraints enable explicit, stable feature dependencies in VLM-generated parametric CAD within a defined operation/constraint set.
- On GERD-EVL (100 test parts) Vision2CAD achieves 100% CR, mIoU 0.4733 and CD 0.0596, improving over the strongest baseline per metric by 11.1% (mIoU) and 17.3% (CD).
- On a 100-part DeepCAD subset it improves mIoU by 5.6% and reduces CD by 41.8% versus the strongest baseline per metric.
- After a uniform 10% increase in extrusion depths, Vision2CAD retains 100% CR and the best geometric metrics on both datasets, indicating preserved parametric dependencies.
- GERD is claimed to be the first dataset aligning commands, geometry states, and Body/Face/Edge/Vertex IDs at every modeling step (2,700 Onshape parts).
Strengths
Clean problem decomposition mapped one-to-one onto mechanisms
Sec. 3 isolates three concrete failure modes (geometry referencing, sketch-plane local coordinates, projected-external constraints), and Sec. 4 answers each with a named component (EGR, LCB, PGR), making the ablation design in Sec. 5.5 directly interpretable.
Editability-centric evaluation is the right instinct
The 10% extrusion-depth protocol (Sec. 5.3) honestly argues reference accuracy against a single GT history is not measurable and instead probes dependency preservation indirectly; Tables 6-7 show the predicted dissociation — static CR stays 100% across all ablations while editing CR falls to 89% (w/o PGR) and 90% (w/o LCB).
GERD fills a real data gap
Deterministic kernel conversion removes a whole class of VLM errors
Disclosed conditional-metric caveat in the static comparison
Weaknesses
Headline margins are survivor-biased
CR criterion is near-tautological for Vision2CAD
Narrow editability probe vs 'proved' claim
No statistical rigor
Reproducibility currently not possible
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Vision2CAD (arXiv 2609.22688) targets image-to-parametric-CAD generation with explicit feature dependencies. The authors identify three failure modes of prior art (Sec. 3): selecting existing faces/edges/vertices as references, exposing sketch-plane local coordinate systems, and constraining sketches to projected external geometry. The system answers each with an ID-tagged render plus geometry table, a deterministic view-to-sketch coordinate bridge, and projected-edge sampling feeding an ordered constraint solver; depth-image registration (Sec. 4.1.1), DINOv3-based retrieval (Sec. 4.1.2), and preview-based acceptance (Sec. 4.3) complete the agent loop. A second contribution is GERD, 2,700 Onshape parts with per-step alignment of commands, geometry states, and Body/Face/Edge/Vertex IDs. Headline claims: best mIoU/CD on GERD-EVL and a 100-part DeepCAD subset (Tables 2-3), 100% compilation retention under a uniform 10% extrusion-depth edit (Tables 4-5), and component contributions via six ablations (Tables 6-7).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Vision2CAD?
A training-free agent framework that turns a single product image into an editable parametric CAD model (a FreeCAD document). A vision-language model chooses operations and references geometry by explicit IDs shown in tagged renders and a geometry table, while deterministic CAD-kernel code handles coordinate conversion and sketch constraints.
What is the GERD dataset?
GERD (Geometry Explicit Reference Dataset) contains 2,700 publicly available Onshape parts split 2,500/100/100 for training/validation/test. Every modeling step records the command, the resulting geometry, and aligned Body/Face/Edge/Vertex IDs, enabling research on explicit geometry references in CAD generation.
How much better is Vision2CAD than prior methods?
On its GERD-EVL test split: mIoU 0.4733 vs the best baseline's 0.4261 (+11.1%, though that baseline compiled on only 18 of 100 parts) and Chamfer distance 0.0596 vs 0.0721 (−17.3%). On a 100-part DeepCAD subset: mIoU +5.6% and Chamfer distance −41.8% versus the strongest baseline per metric.
Is code or data available for Vision2CAD?
Not yet. The paper states data and code will be made available 'upon publication' with the repository link still a placeholder. Replication also requires a proprietary VLM (gpt-5.6-sol), FreeCAD 1.1, and unpublished prompt templates.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-11 are available only as captions and fragments; the visual quality of tag images, registration results, and previews could not be assessed.
- Figure 11's extracted content is garbled (misplaced 'Compile Failed' labels), so qualitative comparisons were not verifiable.
- Prompt templates P(1)-P(3) and any supplementary material are not in the extracted text; prompt sensitivity could not be reviewed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.