Intent2CAD: How Semantic-Parametric Supervision Shapes Text-to-CAD Generation
Organizations: Beijing Institute of Technology, Beijing, China · University of New South Wales, Sydney NSW, Australia
Abstract
Text-to-CAD generation aims to make CAD modeling more accessible by generating executable CAD programs from natural-language descriptions. However, practical CAD workflows require more than executable geometry: generated programs should also expose reusable structure that supports later modification. We introduce intent-oriented text-to-CAD generation, where design intent is expressed as recoverable code-level signals: semantic features, reusable parameters, and lightweight parametric relations. Rather than changing the model architecture or decoding procedure, we investigate how supervision representation granularity affects both CAD generation and zero-shot intent-preserving editing. To enable this study, we construct IntentCAD-100K, a 100K paired text--code dataset built by converting low-level CAD construction traces into executable CadQuery programs. We further construct IntentCAD-Edit-1K to evaluate zero-shot intent-preserving editing from the same representation perspective. Intent2CAD uses a two-stage conservative lifting pipeline: semantic lifting rewrites reliable primitive groups into feature-level operations, while parametric lifting exposes lightweight parametric relations. Under controlled fine-tuning, semantic lifting mainly improves generation reliability and feature recovery, reducing IR from 2.39% to 0.22% and achieving 99.96% API Recovery, while parametric lifting makes reusable relation structure explicit and further improves relation-aware zero-shot editing. Datasets and code are available at https://github.com/JinboYang94/Intent2CAD.
Figures & tables
| Setting | IR.% | Mean CD | Median CD | AR.% |
| GPT-5.4 | 41.85 | 6.21 | 0.13 | 71.97 |
| GPT-5.4-mini | 50.20 | 38.72 | 0.20 | 69.17 |
| Qwen3.5-397B-A17B | 41.70 | 12.56 | 0.15 | 75.25 |
| Qwen3.5-plus | 41.74 | 12.49 | 0.14 | 76.43 |
| CAD-Coder [ 8 ] | 1.29 | 145.04 | 102.95 | 6.11 |
| cadrille [ 31 ] | 24.43 | 208.31 | 116.00 | 0.16 |
| Rep. | IR.% | Mean CD | Median CD | AR.% | ERS.% | PRS.% |
|---|---|---|---|---|---|---|
| 2.39 [2.09,2.69] | 1.44 [1.12,2.18] | 0.14 | N/A | N/A | N/A | |
| 0.22 [0.12,0.33] | 0.46 [0.28,0.77] | 0.12 | 99.96 [99.93,99.98] | N/A | N/A | |
| 0.20 [0.12,0.29] | 0.57 [0.31,0.98] | 0.12 | 99.96 [99.93,99.98] | 93.35 [91.38,95.15] | 95.25 [94.35,96.14] |
| Rep. | IR.% | Median CD | ES.% | FE@ES | CE@ES | DE@ES |
|---|---|---|---|---|---|---|
| 36.30 [33.30,39.30] | 0.14 | 41.80 [38.70,44.90] | 48.79 [43.96,53.38] | 39.23 [33.76,44.69] | 34.18 [28.73,40.00] | |
| 7.00 [5.50,8.60] | 0.12 | 88.90 [86.90,90.80] | 95.17 [93.00,97.10] | 91.00 [87.78,93.89] | 77.09 [72.00,81.82] | |
| 4.70 [3.40,6.00] | 0.12 | 92.50 [90.80,94.10] | 92.51 [89.86,94.93] | 95.82 [93.57,97.75] | 88.73 [84.73,92.36] |
| Rep. | Loc. (%) | ES (%) | FE@ES (%) | CE@ES (%) | DE@ES (%) | VLM |
|---|---|---|---|---|---|---|
| 69.50 | 36.50 | 33.75 | 33.33 | 43.33 | 6.58 | |
| 84.00 | 66.00 | 86.25 | 48.33 | 56.67 | 7.86 | |
| 91.50 | 78.50 | 78.75 | 73.33 | 83.33 | 8.94 |
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Set | Avg. semantic features | Avg. circle loops | Avg. fallback loops | Fully fallback samples |
|---|---|---|---|---|
| Train | 1.12 | 0.99 | 0.34 | 5,070 (6.29%) |
| Val. | 1.13 | 0.99 | 0.35 | 622 (6.17%) |
| Test | 1.12 | 1.00 | 0.33 | 615 (6.10%) |
| Geometric constraints | Dimensional constraints | ||||
|---|---|---|---|---|---|
| Type | Count | Overall (%) | Type | Count | Overall (%) |
| PARALLEL | 155,431 | 17.36 | LENGTH | 147,184 | 16.44 |
| HORIZONTAL | 125,251 | 13.99 | DISTANCE | 83,202 | 9.29 |
| PERPENDICULAR | 93,149 | 10.40 | DIAMETER | 77,295 | 8.63 |
| TANGENT | 68,346 | 7.63 | RADIUS | 21,991 | 2.46 |
| VERTICAL | 47,689 | 5.33 | ANGLE | 5,477 | 0.61 |
| Edit type | Count | Percentage |
|---|---|---|
| Feature Edit (FE) | 414 | 41.40% |
| Consistency Edit (CE) | 311 | 31.10% |
| Derived-Relation Edit (DE) | 275 | 27.50% |
| Type | Subtype | Count | Within-type |
|---|---|---|---|
| Feature | Parameter modification | 200 | 48.31% |
| Feature replacement | 214 | 51.69% | |
| Consistency | EQUAL (height+width) | 74 | 23.79% |
| EQUAL (height) | 63 | 20.26% | |
| EQUAL (radius) | 63 | 20.26% | |
| EQUAL (width) | 62 | 19.94% |
| Experiment | Setting / Rep. | SV.% | Mean CT |
|---|---|---|---|
| Diagnostic Baseline Comparison | |||
| GPT-5.4 | 99.97 | 341.08 | |
| GPT-5.4-mini | 99.98 | 258.74 | |
| Qwen3.5-397B-A17B | 99.60 | 2109.67 | |
| Qwen3.5-plus | 99.61 | 2115.25 | |
| CAD-Coder [ 8 ] | 99.66 | 270.51 | |
| Comparison | Metric | Difference | 95% CI |
|---|---|---|---|
| vs. | IR.% | +2.17 | [1.87, 2.45] |
| vs. | Mean CD | +0.98 | [0.38, 1.62] |
| vs. | IR.% | +0.02 | [-0.07, 0.05] |
| vs. | Mean CD | -0.11 | [-0.42, 0.18] |
| vs. | AR.% | -0.004 | [-0.025, 0.016] |
| Comparison | Metric | Difference | 95% CI |
|---|---|---|---|
| vs. | Overall ES.% | +47.10 | [43.80, 50.50] |
| vs. | FE@ES.% | +46.38 | [41.06, 51.45] |
| vs. | CE@ES.% | +51.77 | [45.98, 57.56] |
| vs. | DE@ES.% | +42.91 | [36.36, 49.45] |
| vs. | Overall IR.% | +29.30 | [26.00, 32.60] |
| vs. | Overall ES.% | +3.60 | [1.20, 6.00] |
| Type | IR.% | ES.% | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| IR | ES | IR | ES | |||||||
| Feature | 28.74 | 4.11 | 4.11 | 48.79 | 95.17 | 92.51 | +24.63 | +46.38 | +0.00 | -2.66 |
| Param | 32.50 | 1.50 | 2.50 | 54.50 | 98.00 | 97.50 | +31.00 | +43.50 | -1.00 | -0.50 |
| Replace | 25.23 | 6.54 | 5.61 | 43.46 | 92.52 | 87.85 | +18.69 | +49.06 | +0.93 | -4.67 |
| Consistency | 41.16 | 6.43 | 3.22 | 39.23 | 91.00 | 95.82 | +34.73 | +51.77 | +3.21 | +4.82 |
| Derived-Relation | 42.18 | 12.00 | 7.27 | 34.18 | 77.09 | 88.73 | +30.18 | +42.91 | +4.73 | +11.64 |
| Rep. | IR.% | Mean CD | Median CD | GT tokens |
|---|---|---|---|---|
| 2.39 [2.09, 2.69] | 1.44 [1.12, 2.18] | 0.14 | 496.76 | |
| Prim.+Cmt. | 1.27 [0.87, 1.94] | 0.87 [0.52, 1.22] | 0.12 | 502.55 |
| -Anon. | 14.36 [13.69, 15.06] | 0.38 [0.25, 0.68] | 0.12 | 339.23 |
| 0.22 [0.12, 0.33] | 0.46 [0.28, 0.77] | 0.12 | 339.23 |
Text-to-CAD Generation
Text-to-CAD generation turns natural-language design intent into executable, editable parametric 3D models and assemblies.
See all 54 papers