SadilKhan/NURBGen

NURBGen

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

Muhammad Usama* · Mohammad Sadil Khan* · Didier Stricker · Muhammad Zeshan Afzal

*equally contributing first authors

---

Model Details

Property	Value
Base model	Qwen/Qwen3-4B
Adapter type	LoRA
Fine-tuning framework	ms-swift
Checkpoint step	180,000

---

How to Use

Requirements

pip install ms-swift transformers peft torch

Single Prompt (ms-swift)

from swift.llm import PtEngine, RequestConfig, InferRequest

engine = PtEngine(
    "Qwen/Qwen3-4B",
    adapters=["SadilKhan/NURBGen"],
    use_hf=True,
)

request_config = RequestConfig(max_tokens=8192, temperature=0.3)

response = engine.infer(
    [InferRequest(messages=[{"role": "user", "content": "Generate NURBS for the following: Design a small table with rounded edges and tapered legs. Include four dowel pins along one side for assembly. The table has chamfers at specific corners and fillets on its underside for smooth transitions. Dimensions: length 23.75 mm, width 70.00 mm, height 27.50 mm."}])],
    request_config=request_config,
)

print(response[0].choices[0].message.content)

---

Single Prompt (HuggingFace / PEFT)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_id = "Qwen/Qwen3-4B"
adapter_id    = "SadilKhan/NURBGen"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model     = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

prompt = "Generate NURBS for the following: Design a small table with rounded edges and tapered legs. Include four dowel pins along one side for assembly. The table has chamfers at specific corners and fillets on its underside for smooth transitions. Dimensions: length 23.75 mm, width 70.00 mm, height 27.50 mm."
messages = [{"role": "user", "content": prompt}]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=8192,
        temperature=0.3,
        do_sample=False,
    )

result = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(result)

Output Format

Each result is saved as {uid}.json:

UID    : smooth_curved_surface
PROMPT : A smooth curved surface with 6 control points
------------------------------------------------------------
<generated NURBS representation>

---

Citation

If you use NURBGen in your research, please cite:

@inproceedings{usama2026nurbgen,
  title={NURBGen: High-Fidelity Text-to-CAD Generation through LLM-Driven NURBS Modeling},
  author={Usama, Muhammad and Khan, Mohammad Sadil and Stricker, Didier and Afzal, Muhammad Zeshan},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  number={12},
  pages={9603--9611},
  year={2026}
}