Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data

1. Problem Statement

The field of generative Computer-Aided Design (CAD) faces a critical data scarcity. While large-scale 3D datasets (e.g., ABC, Objaverse) exist, they primarily consist of static Boundary Representations (B-Reps) or meshes, which lack the construction history (parametric recipe) essential for editing, constraint management, and design intent.

• Existing Limitations: Current datasets with construction sequences (e.g., DeepCAD, Fusion 360 Gallery) are small, limited to simple "sketch-and-extrude" workflows, and lack complex operations like fillets, chamfers, Boolean unions, and lofts.
• The Gap: There is no large-scale, diverse, and executable dataset of CAD programs that includes rich operation vocabularies and is readable by humans. Existing synthetic methods (e.g., CAD-Recode) often produce code that is syntactically valid but semantically shallow (hard-coded values, coordinate chains) and lacks the structural diversity of human-designed models.

2. Methodology: Zero-to-CAD Pipeline

The authors propose Zero-to-CAD, a scalable framework that synthesizes approximately one million executable CAD sequences from scratch without using any real-world CAD files. The core innovation is framing CAD generation as an agentic search problem.

A. Agentic Synthesis Loop

The system embeds a Large Language Model (LLM) within a feedback-driven CAD environment (using the CadQuery Python library). The agent operates in a closed loop:

1. Generation: The LLM generates CadQuery code based on a part description.
2. Execution & Validation: The code is executed in an isolated subprocess. If it fails, the agent receives structured error feedback (syntax errors, topology issues, export failures).
3. Tool Use: Upon failure, the agent can:
   • Lookup Documentation: Retrieve specific API syntax from CadQuery documentation (using TF-IDF and regex).
   • Reason & Repair: Analyze the error, consult docs, and revise the code.
4. Iteration: This process repeats (up to 10 turns per attempt) until the code executes successfully and passes geometric validation.

B. Two-Stage Generation Protocol

To ensure diversity and scale, the pipeline uses a two-stage approach:

1. Catalog Generation: The LLM generates a diverse catalog of part descriptions (e.g., "mounting bracket," "gear") organized into 65 predefined mechanical categories.
2. Code Generation: For each description, the agent generates code. To prevent mode collapse, the system uses:
   • Reference Snippets: Providing templates for complex categories to guide structure without enforcing parameter copying.
   • Prompt Variability: Injecting randomness to encourage diverse design solutions.

C. Validation Framework

A rigorous three-stage filter ensures only high-quality data is released:

1. Code Execution: Checks for syntax errors and runtime exceptions.
2. Geometric Validation: Ensures topological validity (no self-intersections), connectivity (single solid body), minimum complexity (≥7 faces), and positive volume.
3. Export Validation: Verifies successful export to both STL and STEP formats to catch subtle geometric defects.

D. Dataset Curation

• Total Scale: ~1 million generated sequences (999,633 accepted).
• Curated Subset: A 100,000-model subset selected via k-means clustering on DINOv3 visual embeddings to ensure maximum geometric diversity.
• Readability: Unlike coordinate-chain transpilation, Zero-to-CAD programs use named parameters (e.g., plate_thickness, fillet_radius) and logical construction steps, making them human-readable and editable.

3. Key Contributions

1. Zero-to-CAD Dataset: The first million-scale dataset of executable, interpretable CAD construction sequences with broad operation coverage (Booleans, fillets, chamfers, lofts, sweeps, shells).
2. Agentic Synthesis Framework: A novel pipeline demonstrating that LLMs, when grounded in a CAD environment with tool access and documentation, can self-correct to produce valid, complex designs without real training data.
3. Readability & Editability: The generated code mimics human engineering practices (named variables, logical flow), bridging the gap between synthetic generation and practical usability.
4. Bootstrapping Capability: Proof that synthetic data can bootstrap sequence generation models, enabling them to reconstruct editable CAD from images without ever seeing real construction histories.

4. Experimental Results

A. Dataset Quality & Diversity

• Comparison: Compared against CAD-Recode and ABC.
• Metrics: Zero-to-CAD achieves a lower Fréchet distance to the real-world ABC dataset (0.164 vs. 0.268) and higher k-ball coverage.
• Complexity: The mean face count (46.2) closely matches real-world parts (ABC: 50.7), whereas CAD-Recode is significantly simpler (16.4).
• Validity: 0% of Zero-to-CAD parts are disconnected multi-body solids or below the 7-face threshold, compared to 56.3% and 12.9% respectively for CAD-Recode.

B. Image-to-Sequence Reconstruction (Bootstrapping)

The authors fine-tuned a Qwen3-VL-2B Vision-Language Model (VLM) on the Zero-to-CAD dataset to reconstruct editable CAD code from 8 multi-view images.

• In-Distribution Performance: The fine-tuned model achieved 82.1% success rate with a mean IoU of 0.747 on the Zero-to-CAD test set.
• Baseline Comparison: It significantly outperformed the base Qwen model (6.6% success) and GPT-5.2 (72.2% success, 0.485 IoU).
• Out-of-Distribution (OOD) Generalization: When tested on the real-world ABC dataset (human-designed), the model maintained a 61.0% success rate and mean IoU of 0.377, outperforming GPT-5.2 variants in geometric fidelity (IoU and Chamfer Distance), despite being trained only on synthetic data.

5. Significance

• Solving the Data Bottleneck: Zero-to-CAD demonstrates that the scarcity of editable CAD data can be overcome through scalable agentic synthesis, eliminating the need for proprietary data extraction.
• Bridging Geometry and Intent: By providing parametric, readable code, the dataset enables the next generation of CAD AI to not just generate shapes, but generate designs that engineers can modify.
• Feasibility of Synthetic Supervision: The results prove that LLM priors about mechanical design, when validated by execution environments, can be converted into high-quality training data. This allows smaller, specialized models (2B parameters) to outperform massive frontier models (GPT-5.2) on specific CAD tasks, offering a cost-effective path for industrial AI adoption.
• Resource Release: The authors release the full dataset, the curated subset, precomputed embeddings, and the fine-tuned model, providing a vital resource for the research community.