Research map / Point clouds / Point cloud classification
Draw Step by Step: Reconstructing CAD Construction Sequences from Point Clouds via Multimodal Diffusion.
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.
Proposes CAD-Diffuser, a masked-token multimodal diffusion model (BERT-based DistilBERT denoiser) with a volume-based noise schedule that reconstructs CAD construction sequences from point clouds in outline-first order.
Problem
Reconstructing parametric CAD construction sequences from raw point clouds in a way that reflects human top-down design intent, with accurate, diverse and reusable sequences.
Approach
Unifies point cloud tokens and CAD sequence tokens at the token level and runs a discrete diffusion as a random walk between tokens and [MASK] (absorbing for target, reflective for source) within a BERT-style masked modeling scheme, with a volume-based noise schedule that masks detail-carving tokens first so outlines are generated first.
Key results
- On DeepCAD, reaches 88.55% command accuracy and 82.92% parameter accuracy, versus 82.69% and 74.58% for HNC-CAD.
- On Fusion 360 Gallery, achieves 85.56% command accuracy and 80.48% parameter accuracy, with invalid rate 1.65%.
- Generation diversity: 0.45 distinct token 4-gram score versus 0.35 for HNC-CAD; user study shows reconstructions are more likely reused than baselines and approach ground truth.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAD-Diffuser instead of HNC-CAD | CAD sequence reconstruction · DeepCAD test | Command Accuracy: 82.69% -> 88.55% | Helped |
| Use CAD-Diffuser instead of HNC-CAD | CAD sequence reconstruction · DeepCAD test | Parameter Accuracy: 74.58% -> 82.92% | Helped |
| Use CAD-Diffuser instead of ExtrudeNet | CAD sequence reconstruction · DeepCAD test | Median CD: 0.337 -> 0.302 | Helped |
| Use CAD-Diffuser instead of HNC-CAD | CAD sequence reconstruction · Fusion 360 Gallery test | Command Accuracy: 75.46% -> 85.56% | Helped |
| Use CAD-Diffuser instead of HNC-CAD | CAD sequence reconstruction · Fusion 360 Gallery test | Parameter Accuracy: 64.52% -> 80.48% | Helped |
| Use CAD-Diffuser instead of SECAD-Net | CAD sequence reconstruction · Fusion 360 Gallery test | IoU: 0.690 -> 0.632 | Hurt |
| Use CAD-Diffuser instead of Point2cyl | CAD sequence reconstruction · Fusion 360 Gallery test | Median CD: 0.418 -> 0.385 | Helped |
| Use CAD-Diffuser instead of HNC-CAD | Unconditional CAD generation · Unconditional generation | Distinct token N-grams: 0.35 -> 0.45 | Helped |
| Use CAD-Diffuser instead of HNC-CAD | Unconditional CAD generation · Unconditional generation | Embedding Similarity: 0.138 -> 0.082 | Helped |
| Use CAD-Diffuser instead of HNC-CAD | Unconditional CAD generation · Unconditional generation | COV: 87.73 -> 88.35 | Helped |
Surprising findings
- Point2cyl, ExtrudeNet and SECAD-Net have high invalid rates and median CD errors, attributed to sketch approximation heuristics that trigger CAD kernel exceptions.
Limitations
- Median Chamfer Distance and IoU are not best on all datasets; the paper notes CD/IoU are on par with 3D reconstruction methods rather than superior.
- Evaluation of design reusability relies on a small user study with 7 evaluators.
Experimental setup
Datasets: DeepCAD Fusion 360 Gallery
Benchmarks: DeepCAD test set Fusion 360 Gallery test set
Metrics: Command Accuracy Parameter Accuracy Median Chamfer Distance IoU Invalid Rate #∆P N-gram Embedding Similarity Edit Distance COV
Baselines: Point2cyl ExtrudeNet SECAD-Net Lambourne's DeepCAD HNC-CAD SkexGen
Builds on: D3PM DiffusionBERT VQ-Diffusion DeepCAD HNC-CAD
Closest papers by meaning
- CAD-SIGNet: CAD Language Inference from Point Clouds using Layer-wise Sketch Instance Guided Attention (CVPR 2024 · similarity 0.89)
- CADDreamer: CAD Object Generation from Single-view Images (CVPR 2025 · similarity 0.88)
- cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Point2CAD: Reverse Engineering CAD Models from 3D Point Clouds (CVPR 2024 · similarity 0.88)
- CAD-GPT: Synthesising CAD Construction Sequence with Spatial Reasoning-Enhanced Multimodal LLMs (AAAI 2025 · similarity 0.87)
- CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images (CVPR 2025 · similarity 0.87)
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.