Zero-to-CAD — Qwen3-VL-2B
A vision-language model fine-tuned to reconstruct executable CAD programs from multi-view images.
Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data
Mohammadmehdi Ataei, Farzaneh Askari, Kamal Rahimi Malekshan, Pradeep Kumar Jayaraman
Autodesk Research
Related Resources
| Resource | Link |
|---|---|
| 📄 Paper | Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data |
| 📦 Zero-to-CAD 1M (full dataset) | ADSKAILab/Zero-To-CAD-1m |
| 📦 Zero-to-CAD 100K (curated subset) | ADSKAILab/Zero-To-CAD-100k |
| 🤖 Fine-tuned Model (this model) | You are here |
| 🗂️ Collection | ADSKAILab/Zero-To-CAD |
Model Description
This model is a fully fine-tuned Qwen3-VL-2B-Instruct that takes 8 rendered views of a 3D shape (4 front, 4 rear at 256×256) and generates executable CadQuery Python code that reproduces the geometry.
The model was trained entirely on synthetic data from Zero-to-CAD 1M (979,633 training samples) — no real-world CAD files were used.
Key Results
| Benchmark | Success Rate | Mean IoU | Median IoU | P90 IoU |
|---|---|---|---|---|
| Zero-to-CAD test | 82.1% | 0.747 | 0.847 | 0.999 |
| ABC (out-of-distribution) | 61.0% | 0.377 | 0.303 | 0.854 |
Comparison with Baselines
| Model | Zero-to-CAD Success | Zero-to-CAD Mean IoU | ABC Success | ABC Mean IoU |
|---|---|---|---|---|
| This model | 82.1% | 0.747 | 61.0% | 0.377 |
| GPT-5.2 High | 72.2% | 0.485 | 66.2% | 0.344 |
| GPT-5.2 Medium | 71.1% | 0.495 | 62.6% | 0.346 |
| Qwen3-VL-2B (base) | 6.6% | 0.184 | 5.4% | 0.131 |
Quick Start
Inference
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from datasets import load_dataset
from PIL import Image
import io
model_name = "ADSKAILab/Zero-To-CAD-Qwen3-VL-2B"
model = Qwen3VLForConditionalGeneration.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_name)
# Load 8 rendered views from the dataset
ds = load_dataset("ADSKAILab/Zero-To-CAD-1m", split="train", streaming=True)
sample = next(iter(ds))
views = [
Image.open(io.BytesIO(sample[f"image_{i}"])) if isinstance(sample[f"image_{i}"], bytes)
else sample[f"image_{i}"]
for i in range(8)
]
# Or load 8 views from local files:
# views = [Image.open(f"view_{i}.png") for i in range(8)]
messages = [
{
"role": "system",
"content": "You are a CAD code assistant. Given multiple rendered views of a 3D shape, generate clean, well-structured CadQuery Python code that accurately reproduces the geometry."
},
{
"role": "user",
"content": [
*[{"type": "image", "image": view} for view in views],
{"type": "text", "text": "Generate CadQuery code for this shape."}
]
}
]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=text, images=views, return_tensors="pt").to(model.device)
output_ids = model.generate(**inputs, max_new_tokens=4096)
output_text = processor.batch_decode(output_ids[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0]
print(output_text)
Execute the generated code
import cadquery as cq
exec(output_text)
# `result` contains the reconstructed CadQuery solid
# Export
cq.exporters.export(result, "output.step")
cq.exporters.export(result, "output.stl")
Training Details
| Hyperparameter | Value |
|---|---|
| Base model | Qwen3-VL-2B-Instruct |
| Training mode | Full fine-tuning |
| Max sequence length | 4,096 tokens |
| Optimizer | AdamW |
| Learning rate | 1 × 10⁻⁴ |
| Weight decay | 0.0 |
| LR scheduler | Cosine |
| Warmup ratio | 0.03 |
| Attention dropout | 0.1 |
| GPUs | 16 × NVIDIA H100 80GB |
| Per-GPU batch size | 1 |
| Effective batch size | 16 |
| Epochs | 3 |
| Precision | bfloat16 |
| Distributed strategy | DDP |
Evaluation Protocol
- Metric: Voxelized IoU at 64³ resolution between generated and ground-truth solids
- Rotational alignment: Maximum IoU over 45° rotation increments
- Success rate: Percentage of generations producing valid, executable CadQuery code
Intended Uses
- Image-to-CAD reconstruction — reconstruct editable parametric CAD from rendered views
- Research baseline — starting point for Image-to-Sequence CAD generation research
- Integration — combine with rendering pipelines for end-to-end 3D reconstruction
Limitations
- Trained on synthetic data only; may struggle with photorealistic or noisy inputs
- Expects 8 clean rendered views at 256×256 — other configurations are untested
- Outputs CadQuery code only; other CAD formats require post-processing
- Complex multi-part assemblies may exceed the 4,096 token context window
Citation
If you use this model, please cite:
@misc{ataei2026zerotocadagenticsynthesisinterpretable,
title={Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data},
author={Mohammadmehdi Ataei and Farzaneh Askari and Kamal Rahimi Malekshan and Pradeep Kumar Jayaraman},
year={2026},
eprint={2604.24479},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2604.24479}
}
License
This model is released under the Apache License 2.0.