ConCAD: Constraint-Aware Image-to-CAD Generation with Dual-Granularity Rewards

Chenxi Zhai; Xi Cheng; Hang Cheng; Zhicheng Guan; Mingyu Fan; Yanzhe Tang; Pingfa Feng; Long Zeng

cs.CVSep 28, 2026

ConCAD: Constraint-Aware Image-to-CAD Generation with Dual-Granularity Rewards

Authors: Chenxi Zhai, Xi Cheng, Hang Cheng, Zhicheng Guan, Mingyu Fan, Yanzhe Tang, Pingfa Feng, Long Zeng

Organizations: Tsinghua University, Shenzhen International Graduate School

Abstract

Image-to-CAD generation seeks executable parametric programs that recover both the geometry and design intent of a reference object. Existing systems are commonly evaluated by validity and shape overlap, although two solids with similar volume can encode different CAD relations. We introduce ConCAD, a constraint-aware image-to-CAD framework optimized via Group Relative Policy Optimization (GRPO) with rewards at two complementary granularities: a code-level constraint reward and an execution-level geometric reward. This complementary design disambiguates structurally distinct yet volumetrically similar shapes while ensuring valid 3D geometry. To verify that these rewards recover geometry and design intent, we introduce a B-rep geometric constraint satisfaction rate (G-CSR), which analytically extracts and evaluates geometric constraints from boundary representations. Experiments on the DeepCAD and Zero2CAD demonstrate that ConCAD achieves the best IoU and Chamfer Distance over competitive baselines, while also outperforming them on G-CSR, validating its superior recovery of both geometric fidelity and parametric design intent.

Figures & tables

Figure 1: Overview of the ConCAD training pipeline. Following supervised initialization, the policy samples $G$ candidate programs from an input image and is optimized via Group Relative Policy Optimization (GRPO), combining the pre-execution code-level constraint reward $R_{\rm con}$ with the execution-level geometry reward $R_{\rm iou}$ computed via inertia-aligned volumetric overlap on valid solids. This complementary formulation ensures both parametric design intent and 3D spatial fidelity.

Figure 2: Reconstruction examples produced by different methods on the DeepCAD test set.

Method	IR $\downarrow$	IoU $\uparrow$	CD $\downarrow$	G-CSR $\uparrow$
GPT-5.6-Sol [ 16 ]	2	0.599	0.016	0.879
DeepSeek-V4-Pro [ 21 ]	4	0.466	0.027	0.798
Qwen3-VL-2B [ 22 ]	98	0.160	0.106	0.850
CAD-Coder [ 4 ]	2	0.662	0.022	0.837
ReCAD [ 9 ]	2	0.723	0.011	0.906
Ours: SFT	1	0.717	0.011	0.903

Table 1: Results on the DeepCAD test set, evaluated using the identical test split as CAD-Coder [ 4 ] .

Reward Formulation	IR $\downarrow$	IoU $\uparrow$	CD $\downarrow$	G-CSR $\uparrow$
SFT	1	0.717	0.011	0.903
$R_{\rm exec}+R_{\rm iou}$	1	0.700	0.013	0.897
$R_{\rm exec}+R_{\rm con}$	1	0.654	0.010	0.908
$R_{\rm exec}+R_{\rm GCSR}+R_{\rm iou}$	2	0.702	0.009	0.910
$R_{\rm exec}+R_{\rm con}+R_{\rm iou}$	0	0.734	0.009	0.913

Table 3: Reward-component ablation on the DeepCAD test set.

Explore similar work

CIT-CAD: Constraint Intent Tree-based CAD Code Generation and Verification
Sep 7, 2026Yali Du, Hui Sun, San-Zhuo Xi +1Image-To-Cad Generation Solidworks
National Key Laboratory for Novel Software Technology, School of Artificial Intelligence, Nanjing University, China
CADFit: Precise Mesh-to-CAD Program Generation with Hybrid Optimization
May 2, 2026Ghadi Nehme, Eamon Whalen, Faez AhmedParametric Computer-Aided Design Programs Geometric Reconstruction
Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Siemens Digital Industries Software, Plano, TX 75024, USA
BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD
May 11, 2026Haozhe Zhang, Kaichen Liu, Miaomiao Chen +4Parametric Computer-Aided Design Programs Image-To-Code Generation
University of Virginia · University of California, San Diego · Rice University

Sep 7, 2026, cs.AIJ/K move · Enter open · S save

CIT-CAD: Constraint Intent Tree-based CAD Code Generation and Verification

Yali Du, Hui Sun, San-Zhuo Xi, Ming Li

National Key Laboratory for Novel Software Technology, School of Artificial Intelligence, Nanjing University, China

Image-To-Cad Generation Solidworks Invariant Synthesis Structural Constraints Violation

Open paper AskPDF arXiv

Natural-language Computer-Aided Design (CAD) code generation aims to turn design intent into executable and editable parametric programs. Large language models (LLMs) make this goal increasingly practical, but useful systems must preserve the construction process behind the rendered geometry. Existing benchmarks and methods mostly focus on how closely the generated CAD model matches the reference geometry, often using metrics such as Intersection over Union (IoU). Such metrics can miss errors in part decomposition, construction hierarchy, Boolean operations, sketch structure, and geometric relations. This gap calls for a representation that makes design intent explicit and lets a system check generated code against that intent. We propose CIT-CAD, a framework that infers a Constraint Intent Tree (CIT) from the input description to represent the intended entities, hierarchy, operations, and relations. The tree has two roles: it guides CAD code generation and defines expected constraints for verification. The framework extracts actual constraints from the generated program, compares them with the expected constraints, and uses mismatches to localize and repair design violations. Experiments show that the framework improves CAD generation performance, with larger gains on more complex multi-entity designs. By turning design intent into an explicit and checkable object, this work is the first attempt to move text-to-CAD generation beyond rendered-geometry matching toward construction-aware synthesis, verification, and repair.

Abstract

Explore similar work

CIT-CAD: Constraint Intent Tree-based CAD Code Generation and Verification

CADFit: Precise Mesh-to-CAD Program Generation with Hybrid Optimization

BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD