Research map / Image generation / Autoregressive and latent generation
Revisiting CAD Model Generation by Learning Raster Sketch
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 RECAD, which generates raster sketch images and extrusion boxes for CAD models using two sequential diffusion networks (a Transformer extrusion denoiser and a 3D UNet sketch denoiser) with a sketch image VAE, replacing sequence-based sketch representations.
Problem
Sequence-based CAD generation methods represent sketches as parametric line/curve sequences, which limits geometric complexity, yields invalid command sequences, and produces unnatural latent interpolations.
Approach
Represent sketch-and-extrude CAD models as pairs of binary raster sketch images and extrusion boxes (eight parameters encoding bounding box, direction, and Boolean operation), and generate them with a two-stage latent diffusion pipeline: extrusion boxes first, then sketch latents conditioned on the boxes.
Key results
- Unconditional generation: RECAD achieves COV 86.67% and MMD 0.768 (x10^-2), outperforming DeepCAD, SkexGen and HNC-CAD on COV and MMD.
- RECAD reaches 99.65% Valid versus 86.24% (DeepCAD), 86.73% (SkexGen) and 76.07% (HNC-CAD).
- HNC-CAD has slightly better JSD (0.558 vs 0.593 for RECAD); RECAD's Novel (94.4%) and Unique (98.3%) are competitive.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RECAD instead of HNC-CAD | Unconditional CAD generation · DeepCAD test set | COV (%): 85.33 -> 86.67 | Helped |
| use RECAD instead of HNC-CAD | Unconditional CAD generation · DeepCAD test set | MMD: 0.772 -> 0.768 | Helped |
| use RECAD instead of HNC-CAD | Unconditional CAD generation · DeepCAD test set | JSD: 0.558 -> 0.593 | Hurt |
| use RECAD instead of HNC-CAD | Unconditional CAD generation · DeepCAD test set | Valid (%): 76.07 -> 99.65 | Helped |
| add sketch retrieval instead of using contour extraction (ablation) | RECAD · DeepCAD test set | Valid (%): 95.88 -> 99.65 | Helped |
| add sketch retrieval instead of using contour extraction (ablation) | RECAD · DeepCAD test set | MMD: 0.764 -> 0.768 | Hurt |
| add sketch retrieval instead of using contour extraction (ablation) | RECAD · DeepCAD test set | JSD: 0.616 -> 0.593 | No clear effect |
Surprising findings
- Model trained mainly on mechanical parts generates reasonable shapes when given extrusion boxes for common objects such as chairs and tables.
Limitations
- Raster representation can impede accurate representation of intricate contours, especially at low resolution (32x32 inputs).
- No built-in mechanism ensures plausibility or realism of generated designs.
- Sketch retrieval limits contour diversity; direct contour extraction is needed for edits.
Experimental setup
Datasets: DeepCAD
Benchmarks: DeepCAD test set
Metrics: COV MMD JSD Valid Novel Unique
Baselines: DeepCAD SkexGen HNC-CAD
Builds on: DDPM Stable Diffusion DeepCAD PNDM
Closest papers by meaning
- ReCAD: Reinforcement Learning Enhanced Parametric CAD Model Generation with Vision-Language Models (AAAI 2026 · similarity 0.91)
- SketchDNN: Joint Continuous-Discrete Diffusion for CAD Sketch Generation (ICML 2025 · similarity 0.90)
- From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.88)
- CADDreamer: CAD Object Generation from Single-view Images (CVPR 2025 · similarity 0.87)
- CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images (CVPR 2025 · similarity 0.86)
- DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model (NeurIPS 2026 · similarity 0.86)
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.