Text-to-CAD generation on CADBench-Wild human evaluation
77.46Attr. FidelityGraph-CAD
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Graph-CADBackbone=Llama3-8B, Training Strategy=SAPCL2026.04 | 77.46 | 76.07 | 61.39 | 71.64 | |
| GPT-5Paradigm=Graph-CAD2026.04 | 76.82 | 74.75 | 54.6 | 68.72 | |
| Claude-opus-4-1Paradigm=Graph-CAD2026.04 | 74.62 | 73.56 | 69.07 | 72.42 | |
| Graph-CADBackbone=Llama3-8B, Training Strategy=SFT2026.04 | 70.45 | 72.68 | 59.71 | 67.61 | |
| Claude-opus-4-1Paradigm=end-to-end2026.04 | 69.23 | 72.85 | 60.72 | 67.6 | |
| GPT-5Paradigm=end-to-end2026.04 | 68.94 | 71.07 | 59.02 | 66.34 | |
| BlenderLLM2026.04 | 67.22 | 65.09 | 46.51 | 59.14 |