CAD-Coder: An Open-Source Vision-Language Model for Computer-Aided Design Code Generation
TLDR
CAD-Coder is an open-source VLM fine-tuned to generate editable CAD code from images, outperforming GPT-4.5 and Qwen2.5-VL-72B.
Reasoning
The paper introduces a novel dataset and demonstrates strong results with 100% valid syntax and high 3D similarity, but generalizability claims are limited to 'some signs' and real-world evaluation details are sparse in the abstract.
Read-first score
Read-first score 65.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 99.
Field roles
Rank sensitivity
Stability: volatile; rank range: 14.
Keyword Scores
Deep Analysis
Innovations
- CAD-Coder: an open-source Vision-Language Model fine-tuned to generate editable CAD code (CadQuery Python) directly from visual input.
- GenCAD-Code dataset: a novel collection of over 163k CAD-model image and code pairs for training.
- Demonstration of generalizability to real-world images and CAD operations unseen during fine-tuning.
Methodology
CAD-Coder is a VLM fine-tuned on the GenCAD-Code dataset of 163k image–code pairs to output CadQuery Python code from images. It is evaluated against baselines GPT-4.5 and Qwen2.5-VL-72B using valid syntax rate and 3D solid similarity metrics.
Key Results
CAD-Coder achieves 100% valid syntax rate and the highest 3D solid similarity accuracy, outperforming GPT-4.5 and Qwen2.5-VL-72B, and shows initial generalizability to real-world images and unseen CAD operations.
Limitations
- Generalizability is only partial, described as 'some signs' of success on real-world images and unseen operations, indicating limited robustness outside the training distribution.