Research map / 3D vision / 3D generation
CADDreamer: CAD Object Generation from Single-view Images
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.
Generates watertight CAD B-rep models from a single image by fine-tuning Wonder3D to jointly predict multi-view normal and primitive semantic maps, then graph-cut segmentation, geometric optimization, and topology-preserving intersection.
Problem
Single-view image-to-CAD reconstruction: existing 3D diffusion generators output dense, unstructured meshes rather than compact, watertight B-rep CAD models with clean primitives and sharp edges.
Approach
Encode primitive semantics as colors in a cross-domain multi-view diffusion model (Wonder3D fine-tuned to output normal and primitive maps), segment the reconstructed mesh into primitive patches via graph cut, then refine primitive parameters under parallel/perpendicular/collinear/intersection constraints and build a B-rep by topology-preserving intersection.
Key results
- CADDreamer achieves CD 1.27 and NC 92.6% versus best baseline CRM at CD 3.97 and NC 64.4%.
- Primitive count segmentation accuracy SEG(P) of 97.9% vs 49.3% for CRM.
- Hanging face ratio of 2.4% vs 35.2% for CRM; B-rep CD 1.36 vs 9.74.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CADDreamer instead of CRM | Single-view reconstruction on synthetic images · Synthetic images | Chamfer distance (×100): 3.97 -> 1.27 | Helped |
| Use CADDreamer instead of CRM | Single-view reconstruction on synthetic images · Synthetic images | Normal consistency (%): 64.4 -> 92.6 | Helped |
| Use CADDreamer instead of CRM | Single-view reconstruction on synthetic images · Synthetic images | SEG(V) (%): 40.2 -> 95.7 | Helped |
| Use CADDreamer instead of CRM | Single-view reconstruction on synthetic images · Synthetic images | SEG(P) (%): 49.3 -> 97.9 | Helped |
| Use CADDreamer instead of CRM | B-rep reconstruction on synthetic images · Synthetic images | Hanging faces (%): 35.2 -> 2.4 | Helped |
| Use CADDreamer instead of CRM | B-rep reconstruction on synthetic images · Synthetic images | B-rep Chamfer distance (×100): 9.74 -> 1.36 | Helped |
Limitations
- Performance constrained by input image quantity and resolution; fails to detect very fine geometric features.
- Struggles with challenging viewpoints and complex occlusions where not all primitives are visible.
Experimental setup
Datasets: ABC DeepCAD
Benchmarks: synthetic CAD test set (1,000 samples) real handheld-captured CAD images
Metrics: Chamfer distance normal consistency SEG(V) SEG(P) hanging face ratio B-rep Chamfer distance
Baselines: CRM LRM InstantMesh SyncDreamer Point2CAD
Builds on: Wonder3D NeuS Efficient RANSAC Era3D
Closest papers by meaning
- Draw Step by Step: Reconstructing CAD Construction Sequences from Point Clouds via Multimodal Diffusion. (CVPR 2024 · similarity 0.88)
- CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images (CVPR 2025 · similarity 0.88)
- Revisiting CAD Model Generation by Learning Raster Sketch (AAAI 2025 · similarity 0.87)
- Generating CAD Code with Vision-Language Models for 3D Designs (ICLR 2025 · similarity 0.87)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
- cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 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.