IterCAD: Iterative Program Repair for CAD Code Generation from Orthographic Views

Yuchuan Wu; Ke Niu; Haiyang Yu; Zhuofan Chen; Xiangyang Xue; Bin Li

cs.CVAug 25, 2026

IterCAD: Iterative Program Repair for CAD Code Generation from Orthographic Views

Authors: Yuchuan Wu, Ke Niu, Haiyang Yu, Zhuofan Chen, Xiangyang Xue, Bin Li

Organizations: Fudan University Shanghai, China

Abstract

Generating executable parametric CAD code from dimension-annotated orthographic drawings is a challenging task requiring geometric understanding, procedural reasoning, and precise numerical prediction. Existing vision-language approaches typically formulate this problem as one-shot generation, preventing the model from inspecting intermediate CAD results and correcting early mistakes, often leading to non-executable code or geometrically inconsistent outputs. In this paper, we propose IterCAD, an iterative framework that reformulates orthographic-view-to-CAD generation as a progressive program repair process. Instead of predicting the final CAD code in a single pass, IterCAD repeatedly analyzes the current CAD result, reasons about its discrepancy with the target views, and explicitly decides whether to REVISE the code or STOP the refinement process. To make iterative repair learnable, we further construct IterCAD-RS, a structured revise-or-stop supervision set containing both repairable intermediate CAD states and already-correct states, and develop a three-stage training strategy for initial generation, revision learning, and multi-turn RL optimization. By closing the loop between visual understanding, geometric verification, and code refinement, IterCAD progressively corrects structural and parametric errors. Experiments on CADExpert show that IterCAD consistently improves code executability and geometric fidelity over strong one-shot baselines.

Figures & tables

Figure 1. Comparison of one-shot CAD generation and IterCAD . While one-shot methods often fail to produce correct CAD programs, IterCAD progressively improves the program through iterative state assessment and revision. Comparison between one-shot CAD generation methods and IterCAD. One-shot methods may produce execution failures or incorrect geometry, while IterCAD repeatedly assesses and revises the current CAD program until obtaining a correct result.

Structural Family	Perturbation Target	Representative Operations	Typical Resulting Errors
Body-level	Sketch plane	Change XY / YZ / XZ workplane	Wrong global orientation or projection mismatch
Body-level	Base profile geometry	Modify primitive type, polygon side count, or profile scale	Incorrect outer silhouette or overall geometry
Body-level	Extrusion parameter	Modify extrusion height or depth	Incorrect thickness or global dimensions
cut	Inner profile geometry	Change inner profile type or size	Wrong cavity shape or incorrect inner dimensions
cut	Modifier existence	Remove the cut operation	Missing inner structure
hole_pattern	Hole multiplicity	Increase or decrease the number of holes	Incorrect hole count

Table 1. Perturbation taxonomy used to construct IterCAD-RS . Starting from a correct structured CAD state ( ShapeSpec ), we inject one or two geometric perturbations to synthesize executable but geometrically inconsistent intermediate states for revise-or-stop supervision.

Figure 2. Overview of IterCAD . Stage I learns first-turn initial draft generation from dimension-annotated orthographic views. Stage II learns structured revise-or-stop behavior from intermediate CAD states. Stage III further optimizes the full multi-turn repair process with GRPO over sampled repair trajectories, yielding a unified iterative policy for initial generation, revision, and adaptive stopping.

Model	IoU (%) $\uparrow$	Mean CD $\downarrow$	Med CD $\downarrow$	Exec. (%) $\uparrow$
LLaVA-1.5 ( Liu et al., 2024 )	0.73	29.36	8.14	4.79
Phi-3.5-Vision ( Abdin et al., 2024 )	3.44	27.92	7.75	8.50
InternVL2.5 ( Chen et al., 2024 )	11.98	22.27	7.36	23.92
Qwen2.5-VL ( Bai et al., 2025b )	19.21	20.68	6.64	31.27
InternVL3 ( Zhu et al., 2025 )	24.59	21.40	6.62	29.68
Gemini2.5 Pro ( Comanici et al., 2025 )	30.86	14.13	5.97	39.40

Table 2. Main results on CADExpert. IterCAD-Q3VL and IterCAD-Q3.5 denote our method instantiated with Qwen3-VL-8B-Instruct and Qwen3.5-9B, respectively. Best results are shown in bold , and second-best results are underlined .

Base model: Qwen3.5-9B
S1	S2	S3	MT	IoU (%) $\uparrow$	Mean CD $\downarrow$	Med CD $\downarrow$	Exec. (%) $\uparrow$
$\checkmark$	$\times$	$\times$	$\times$	57.42	3.5819	0.1939	96.39
$\checkmark$	$\checkmark$	$\times$	$\times$	54.46	3.4506	0.3520	93.65
$\checkmark$	$\checkmark$	$\times$	$\checkmark$	60.40	2.8102	0.1590	93.37
$\checkmark$	$\times$	$\checkmark$	$\checkmark$	64.26	2.6319	0.1448	94.28
$\checkmark$	$\checkmark$	$\checkmark$	$\times$	89.91	0.7492	0.1193	98.51

Table 3. Ablation study of IterCAD on CADExpert. S1, S2, and S3 denote Stages I, II, and III, respectively, and MT denotes multi-turn inference. Best results are shown in bold .

Figure 3. Iterative repair behavior on CADExpert. (a) Round-wise IoU. (b) Cumulative stopping ratios by refinement turn. (c) Average inference time under one-turn and multi-turn settings. Three plots compare IterCAD across refinement turns. The first shows increasing IoU, the second shows that most samples stop within the early turns, and the third compares one-turn and multi-turn inference time for the two backbones.

Parametric CAD Modeling

Parametric CAD modeling represents 3D designs as editable, constraint-driven construction histories and develops methods to generate, reconstruct, and revise them.

See all 81 papers

IterCAD: An Iterative Multimodal Agent for Visually-Grounded CAD Generation and Editing
Jun 11, 2026Tao Hu, Jiaxin Ai, Licheng Wen +12Parametric CAD Modeling Text-to-CAD Generation
Shanghai Artificial Intelligence Laboratory · University of Science and Technology of China · Shanghai Innovation Institute +4
CADReasoner: Iterative Program Editing for CAD Reverse Engineering
Feb 18, 2026Soslan Kabisov, Vsevolod Kirichuk, Andrey Volkov +5Parametric CAD Modeling 3D Reconstruction
Lomonosov Moscow State University · Universit´e Paris Dauphine · Innopolis University +1
CIT-CAD: Constraint Intent Tree-based CAD Code Generation and Verification
Sep 7, 2026Yali Du, Hui Sun, San-Zhuo Xi +1Parametric CAD Modeling Text-to-CAD Generation
National Key Laboratory for Novel Software Technology, School of Artificial Intelligence, Nanjing University, China

Jun 11, 2026J/K move · Enter open · S save

IterCAD: An Iterative Multimodal Agent for Visually-Grounded CAD Generation and Editing

Tao Hu, Jiaxin Ai, Licheng Wen, Xueheng Li+11

Shanghai Artificial Intelligence Laboratory · University of Science and Technology of China · Shanghai Innovation Institute+4

Parametric CAD Modeling Text-to-CAD Generation Multimodal Agents Code Generation Benchmarks

Open paper AskPDF arXiv

Computer-Aided Design is pivotal in modern manufacturing, yet existing automated methods predominantly rely on open-loop, one-shot generation, creating a mismatch with iterative real-world practices. In this paper, we present IterCAD, a unified multimodal agent framework for closed-loop, interactive CAD generation and editing. We formulate the task as a multi-turn interaction between a multimodal agent and an executable CAD sandbox, covering three tasks: Drawing-to-Code, Text-to-Code, and Interactive Editing. To support this, we develop a data synthesis pipeline incorporating advanced industrial manufacturing features to generate standard-compliant multi-view engineering drawings, complex code-editing tasks, and high-fidelity interaction trajectories. We optimize the agent via progressive SFT followed by geometry-aware reinforcement learning with viable-prefix masking to enhance code executability and geometric fidelity. Finally, we introduce the IterCAD-Bench evaluation suite and propose the Chamfer Distance Tolerance-Recall (CD-TR) curve alongside its AUC-TR metric, establishing a survivor-bias-free standard that unifies code validity and geometric precision. Extensive experiments demonstrate that IterCAD achieves highly competitive performance across multiple benchmarks, significantly outperforming existing approaches in both code executability and geometric precision, while exhibiting superior capabilities in closed-loop iterative refinement.

Abstract

Parametric CAD Modeling

Explore similar work

IterCAD: An Iterative Multimodal Agent for Visually-Grounded CAD Generation and Editing

CADReasoner: Iterative Program Editing for CAD Reverse Engineering

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