Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data
TLDR
Zero-to-CAD uses an LLM agent to synthesize millions of executable CAD programs without real data, enabling image-to-CAD reconstruction.
Reasoning
The paper presents a novel agentic framework for generating large-scale synthetic CAD construction sequences, addressing data scarcity. Its strengths include scalability and demonstrated utility for downstream tasks, but it lacks real-world validation and may have limitations from purely synthetic data.
Read-first score
Read-first score 65.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 98.
Field roles
Rank sensitivity
Stability: volatile; rank range: 10.
Keyword Scores
Deep Analysis
Innovations
- Agentic search framework using LLM in a feedback-driven CAD environment to synthesize executable CAD programs
- Synthesis of approximately one million executable, readable, editable CAD sequences without real construction-history data
- Release of a curated subset of 100,000 high-quality models selected for geometric diversity
- Fine-tuning a vision-language model on synthetic data to reconstruct editable CAD programs from multi-view images, outperforming GPT-5.2
Methodology
The framework frames CAD program synthesis as an agentic search problem, embedding a large language model (LLM) within a feedback-driven CAD environment that iteratively generates, executes, and validates code using tools and documentation lookup to promote geometric validity and operation diversity. The resulting synthetic dataset is used to fine-tune a vision-language model for CAD program reconstruction from multi-view images.
Key Results
The agentic approach synthesizes approximately one million executable CAD sequences covering diverse operations beyond sketch-and-extrude. A vision-language model fine-tuned on this synthetic data outperforms strong baselines including GPT-5.2 on reconstructing editable CAD programs from multi-view images.