NURBGen: High-Fidelity Text-to-CAD Generation through LLM-Driven NURBS Modeling

TL;DR

NURBGen tackles the challenge of generating high-fidelity editable CAD models from natural language by casting text-to-CAD as a structured language-generation task. It fine-tunes an LLM to produce JSON-encoded NURBS surface parameters, enabling direct conversion to BRep geometry, and introduces partABC, a large-scale dataset of part-level CAD components with NURBS annotations and captions. A hybrid representation that combines untrimmed NURBS with analytic primitives addresses trimming artifacts while improving token efficiency. Experimental results show superior geometric fidelity, dimensional accuracy, and caption quality, highlighting the practical potential of NURBS-based text-to-CAD generation and providing a resource for future research.

Abstract

Generating editable 3D CAD models from natural language remains challenging, as existing text-to-CAD systems either produce meshes or rely on scarce design-history data. We present NURBGen, the first framework to generate high-fidelity 3D CAD models directly from text using Non-Uniform Rational B-Splines (NURBS). To achieve this, we fine-tune a large language model (LLM) to translate free-form texts into JSON representations containing NURBS surface parameters (\textit{i.e}, control points, knot vectors, degrees, and rational weights) which can be directly converted into BRep format using Python. We further propose a hybrid representation that combines untrimmed NURBS with analytic primitives to handle trimmed surfaces and degenerate regions more robustly, while reducing token complexity. Additionally, we introduce partABC, a curated subset of the ABC dataset consisting of individual CAD components, annotated with detailed captions using an automated annotation pipeline. NURBGen demonstrates strong performance on diverse prompts, surpassing prior methods in geometric fidelity and dimensional accuracy, as confirmed by expert evaluations. Code and dataset will be released publicly.

Figures (13)

• Figure 1: Text-to-CAD generation results from NURBGen, showcasing reconstructed CAD models from text prompts.
• Figure 2: Overview of our partABC dataset, data preparation and fine-tuning pipeline. Left: We extract part-level CAD models from the ABC dataset by decomposing CAD assemblies into individual components. Middle: Each part is represented using a hybrid format—faces are encoded as untrimmed NURBS surfaces, with analytic primitives used where NURBS fitting fails. We also generate high-quality captions using InternVL3-13B with a metadata-guided annotation pipeline. Right: We fine-tune Qwen3-4B to map text captions to structured hybrid CAD representations, which can be directly converted to BRep models.
• Figure 3: Our proposed hybrid representation. Left: Untrimmed NURBS surfaces introduce artifacts in hole-like or thin regions. Right We resolve this by substituting their NURB representation with analytic curves (e.g., lines, circles) for improved geometric fidelity.
• Figure 4: Qualitative comparison of reconstructed CAD models from text prompts. From top to bottom, we show generations from GPT-4o, DeepCAD, Text2CAD, and our proposed NURBGen. NURBGen consistently produces more detailed and structurally coherent results, with higher fidelity to the input prompt and fewer geometric artifacts compared to baselines.
• Figure 5: Captions from partABC dataset generated using our captioning pipeline.