CADFS: A Big CAD Program Dataset and Framework for Computer-Aided Design with Large Language Models
TLDR
Introduces CADFS, a dataset of 450k real-world CAD models with 15 operations and a FeatureScript-based framework for text-to-CAD and image-to-CAD generation using VLMs.
Reasoning
The paper presents a large-scale dataset and a novel representation (FeatureScript) that enables complex CAD generation beyond sketch-extrude, achieving SOTA results. Strengths include real-world data and multimodal annotations; weaknesses are the lack of explicit limitations and potential reliance on a proprietary format.
Read-first score
Read-first score 63.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 102.
Field roles
Rank sensitivity
Stability: volatile; rank range: 16.
Keyword Scores
Deep Analysis
Innovations
- Data-centric framework CADFS enabling large vision-language models to generate complex CAD design histories
- FeatureScript-based representation that supports 15 modeling operations beyond sketch-extrude
- Dataset of 450k real-world CAD models with clean, executable FeatureScript programs and multimodal annotations
- Pipeline for reconstructing clean FeatureScript programs from real-world CAD data
- Fine-tuning a VLM on this representation to achieve state-of-the-art text-conditioned CAD generation and image-based reconstruction
Methodology
CADFS is a data-centric framework that uses a FeatureScript-based representation to capture complex CAD design histories. A reconstruction pipeline produces clean, executable FeatureScript programs and multimodal annotations from 450k real-world CAD models spanning 15 operations. A vision-language model is fine-tuned on this dataset for text-conditioned generation and image-based reconstruction.
Key Results
Fine-tuning a VLM on the CADFS dataset achieves state-of-the-art performance in text-conditioned CAD generation and image-based reconstruction, yielding more accurate, diverse, and feature-rich designs. Ablations show that the FeatureScript representation, extended operation set, and representation-aligned textual descriptions each significantly improve performance.