AAAI 2026 · Text-to-CAD Generation

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

The first framework to generate industry-standard NURBS surfaces directly from text prompts, producing editable, parametric CAD models convertible to STEP format.

Muhammad Usama^1,2,3,* · Mohammad Sadil Khan^1,2,3,*,† · Didier Stricker^1,2 · Muhammad Zeshan Afzal^1,3

¹DFKI Kaiserslautern ²RPTU Kaiserslautern-Landau ³MindGarage

* Equally Contributing First Authors † Corresponding author

arXiv GitHub partABC Dataset Poster

SCROLL

/ 01 — Overview

Welcome to NURBGen

First NURBS-Based Text-to-CAD

First framework to generate industry-standard NURBS representations directly from text prompts using LLMs.

partABC Dataset

300k part-level CAD models with NURBS annotations and automatically generated high-quality captions (~85% accuracy).

Hybrid Representation

Combines untrimmed NURBS with analytic primitives — lines, arcs, circles — for geometric robustness and token efficiency.

State-of-the-Art Performance

Surpasses all baselines with 64.1% human preference and lowest invalidity ratio of 0.018.

/ 02 — Method

Architecture & Pipeline

A two-stage pipeline: data preparation with automated NURBS extraction and captioning, followed by LLM fine-tuning for structured CAD generation.

📐

NURBS Extraction

BRep models are normalized and converted to untrimmed NURBS via pythonOCC. Control points, knot vectors, degrees, and weights are extracted for each face.

🔤

Caption Generation

Multi-view renders (6 viewpoints) fed into InternVL3-13B with geometric metadata guidance: dimensions, holes, volume, surface area — producing shape-centric captions.

🧠

LLM Fine-tuning

Qwen3-4B fine-tuned with LoRA (rank 64, α=128) on 180k steps across 4×H200 GPUs. Maps text captions to structured JSON NURBS representations.

Figure 2: Overview of partABC dataset, data preparation and fine-tuning pipeline.

Hybrid Representation

Not all surfaces can be robustly represented by untrimmed NURBS. Thin regions around holes or fillets often introduce geometric artifacts. We detect degenerate faces using Chamfer Distance comparison (CD ≤ ε, where ε = 6×10⁻⁴) and fall back to analytic primitives.

70%

NURBS surfaces

30%

Analytic primitives

Analytic primitives include:

Geom_Circle Geom_Line Geom_BSpline Geom_Arc Geom_Ellipse

NURBS face representation (JSON)

// Face encoded as NURBS surface
{
  "face_0": {
    "poles": [[[0.1285, 0.9999, 0.4857]...]],
    "u_knots": [1.5707, 3.1415, 4.7123],
    "v_knots": [-0.1714, 0.1428],
    "u_mults": [3, 2, 3],
    "v_mults": [2, 2],
    "u_degree": 2,
    "v_degree": 1,
    "weights": [[1.0, 2]...]
  },
  // Degenerate face → analytic fallback
  "face_1": {
    "type": "Geom_Circle",
    "center": [0.282, 0.083, 0.416],
    "radius": 0.051984,
    "normal": [-1.0, 0.0, 0.0]
  }
}
          

Model

Qwen3-4B + LoRA

rank 64, α=128

Training

180k steps

4×H200 GPUs · 3 days

Context Window

8192 / 14k tokens

train / inference

Throughput

~800 tokens/s

NVIDIA RTX 3090

Optimizer

AdamW

lr=5×10⁻⁵, linear warm-up

/ 03 — Evaluation

Performance Results

NURBGen significantly outperforms all baselines in both geometric fidelity and visual alignment with text prompts.

64.1%

Human Preference
(Top-1)

User eval · 1k samples

61.6%

GPT-4o Preference
(5k samples)

Multi-view comparison

0.018

Invalidity Ratio
(Lowest)

Valid BRep structures

4.43

Chamfer Distance
(×10²)

8192-pt normalized

Human Preference Study (%)

5 CAD designers chose the best reconstruction per prompt. Higher is better.

NURBGen

64.1%

Text2CAD

26.1%

DeepCAD

5.6%

GPT-4o

1.5%

Method	User (1k) ↑	GPT-4o ↑
Undecided	2.7	3.2
GPT-4o	1.5	1.9
DeepCAD	5.6	6.1
Text2CAD	26.1	27.2
NURBGen (Ours) SOTA	64.1	61.6

↑ Higher is better. Human evaluation: 5 CAD designers via majority vote on 1k samples. GPT-4o: 2×2 multi-view grid evaluation on 5k samples.

Method	IR ↓	CD ↓	HD ↓	JSD ↓	MMD ↓
GPT-4o	0.17	7.20	0.36	72.87	4.17
DeepCAD	0.32	10.28	0.45	89.77	4.43
Text2CAD	0.05	9.66	0.42	85.27	4.54
NURBGen (Ours) SOTA	0.018	4.43	0.25	57.94	2.14

CD, JSD, MMD multiplied by 10². ↓ Lower is better. Evaluated on 7,500 test samples using 8,192 uniformly sampled points normalized to unit cube. IR: Invalidity Ratio, CD: Chamfer Distance, HD: Hausdorff Distance, JSD: Jensen-Shannon Divergence, MMD: Maximum Mean Discrepancy.

/ 04 — Dataset

The partABC Dataset

A large-scale curated dataset of 300k part-level CAD models with NURBS annotations and high-quality automatically generated captions.

300k Part Models

~85% Caption Accuracy

3 Complexity Tiers

Derived from the ABC dataset (1M CAD models), we decompose assembly-level designs into individual part-level components using PythonOCC, generating 3M instances. A complexity-aware filtering strategy retains 300k geometrically diverse parts.

COMPLEXITY DISTRIBUTION

Simple

10%

Moderate

50%

Complex

40%

Annotation Pipeline

01 4 orthographic renders (512×512) via Blender + Freestyle edges

02 Metadata extraction: dimensions, volume, surface area, hole count via Euler characteristic

03 InternVL3-13B generates shape-centric captions with precise geometric details

Figure 5: Sample captions generated by our automated annotation pipeline using InternVL3-13B.

/ 05 — Gallery

Generated CAD Models

High-fidelity 3D CAD models generated by NURBGen from text descriptions, converted to industry-standard BRep format.

TEXT PROMPT Socket head cap screw with a large countersunk washer. Features a hexagonal socket drive and a cylindrical threaded shank. Dimensions: length 92.96 mm, width 79.38 mm, height 43.66 mm. Ensure smooth curvature at transitions.

TEXT PROMPT Design a rectangular plate with dimensions 330.20 mm x 233.40 mm x 6.00 mm. Include two square through-holes near each end.

TEXT PROMPT Cylindrical bushing with flanges on both ends, featuring a central hollow bore. Symmetrical geometry, suitable for insertion, guiding, or alignment in mechanical assemblies.

TEXT PROMPT A spherical object with subtle indentations on its surface, exhibiting uniform dimensions of 2.33 mm in length, width, and height. The sphere has minor linear markings near the base.

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

Welcome to NURBGen

Architecture & Pipeline

Hybrid Representation

Performance Results

Human Preference Study (%)

The partABC Dataset

Generated CAD Models

Qualitative Comparison

Citation

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

Welcome to NURBGen

Architecture & Pipeline

Hybrid Representation

Performance Results

Human Preference Study (%)

The partABC Dataset

Generated CAD Models

Qualitative Comparison

Citation

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