Research map / Large language models / Agents, code and math
CAD-Editor: A Locate-then-Infill Framework with Automated Training Data Synthesis for Text-Based CAD Editing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces CAD-Editor, a text-based CAD editing framework that synthesizes (instruction, original, edited) triplets with design variation models and LVLMs, and uses a locate-then-infill LLM pipeline (Llama-3-8B with LoRA) on sketch-and-extrude sequences.
Problem
Text-based editing of parametric CAD models given natural-language instructions is underexplored, and training requires triplets of original CAD, edit instruction, and edited CAD with accurate correspondence that do not naturally exist.
Approach
Synthesize triplets by pairing a CAD model with its design variation (HNC-CAD auto-completion) and having LVLMs summarize the differences into instructions via stepwise captioning; then decompose editing into a locating stage that outputs a masked CAD sequence (LCS-derived ground truth) and an infilling stage that fills the masks, both using an LLM.
Key results
- CAD-Editor reaches a Valid Ratio of 95.6% versus 84.5% for GPT-4o-IC and 63.2% for GPT-4o-Basic.
- CAD-Editor achieves JSD 0.65, CD 1.18 and D-CLIP 0.11, with H-Eval 43.2 versus 15.6 for GPT-4o-IC.
- Ablation: the locate-then-infill decomposition raises VR from 86.1% (direct) to 96.5% on the 10k subset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAD-Editor instead of GPT-4o-IC | Text-based CAD editing · DeepCAD test | VR: 84.5 -> 95.6 | Helped |
| Use CAD-Editor instead of GPT-4o-IC | Text-based CAD editing · DeepCAD test | JSD: 0.70 -> 0.65 | Helped |
| Use CAD-Editor instead of GPT-4o-IC | Text-based CAD editing · DeepCAD test | CD: 1.55 -> 1.18 | Helped |
| Use CAD-Editor instead of GPT-4o-IC | Text-based CAD editing · DeepCAD test | D-CLIP: - 0.11 -> 0.11 | Helped |
| Use CAD-Editor instead of GPT-4o-IC | Text-based CAD editing · DeepCAD test | H-Eval: 15.6 -> 43.2 | Helped |
| Add stepwise captioning instead of basic captioning (ablation) | CAD-Editor-mini, 10k examples · DeepCAD test | VR: 88.8 -> 90.1 | Helped |
| Add stepwise captioning instead of basic captioning (ablation) | CAD-Editor-mini, 10k examples · DeepCAD test | JSD: 0.78 -> 0.70 | Helped |
| Add stepwise captioning instead of basic captioning (ablation) | CAD-Editor-mini, 10k examples · DeepCAD test | CD: 1.25 -> 1.22 | Helped |
| Add stepwise captioning instead of basic captioning (ablation) | CAD-Editor-mini, 10k examples · DeepCAD test | D-CLIP: - 0.12 -> - 0.07 | Helped |
| Add locate-then-infill instead of direct fine-tuning (ablation) | CAD-Editor · DeepCAD test | VR: 86.1 -> 96.5 | Helped |
| Add locate-then-infill instead of direct fine-tuning (ablation) | CAD-Editor · DeepCAD test | JSD: 0.77 -> 0.67 | Helped |
| Add locate-then-infill instead of direct fine-tuning (ablation) | CAD-Editor · DeepCAD test | CD: 1.26 -> 1.13 | Helped |
Surprising findings
- Adding the human-selected selective dataset slightly lowered VR (96.5 to 95.6) while improving JSD, CD and D-CLIP.
- Baselines like SkexGen, Hnc-CAD and FlexCAD are only compared on generation quality because they cannot perform text-based editing.
Limitations
- Data synthesis relies on costly LVLMs that struggle with multiple images at once.
- Long contexts and extended sequences limit handling of highly complex CAD models.
- Training restricted to models with at most 3 SE pairs due to DeepCAD imbalance.
Experimental setup
Datasets: DeepCAD Fusion 360 Gallery
Benchmarks: DeepCAD test split (2k examples, 5 outputs each) Fusion 360 Gallery-derived set
Metrics: Valid Ratio JSD Chamfer Distance D-CLIP H-Eval human success rate
Baselines: SkexGen Hnc-CAD FlexCAD Text2CAD GPT-4o-Basic GPT-4o-IC
Builds on: HNC-CAD DeepCAD LoRA
Closest papers by meaning
- Text2CAD: Generating Sequential CAD Designs from Beginner-to-Expert Level Text Prompts (NeurIPS 2024 · similarity 0.90)
- FlexCAD: Unified and Versatile Controllable CAD Generation with Fine-tuned Large Language Models (ICLR 2025 · similarity 0.89)
- CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation (CVPR 2025 · similarity 0.88)
- CAD-GPT: Synthesising CAD Construction Sequence with Spatial Reasoning-Enhanced Multimodal LLMs (AAAI 2025 · similarity 0.88)
- cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Clarify Before You Draw: Proactive Agents for Robust Text-to-CAD Generation (ICML 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.