Research map / 3D vision / 3D generation
Mamba-CAD: State Space Model for 3D Computer-Aided Design Generative Modeling
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.
Mamba-CAD pairs a Mamba encoder-decoder (pretrained on CAD sequence reconstruction) with a latent GAN to generate longer parametric CSG CAD sequences, and releases a 77,078-model dataset with longer sequences.
Problem
Generative modeling of parametric CAD sequences is limited to short sequences, while industrial CAD components need longer command sequences to define.
Approach
Use a Mamba-based autoencoder pretrained on CAD sequence reconstruction to learn a latent space, then train a 1-D latent WGAN-style generator to produce latent codes that the frozen Mamba decoder converts into long parametric CAD sequences.
Key results
- Reconstruction on Mamba-CAD dataset: A_p 99.93 and MCD 0.64 vs DeepCAD 87.38 and 2.97, with average length 38.91 vs 32.52.
- Random generation: MMD 1.94 and JSD 4.28 vs HNC-CAD 2.65 and 6.07, with COV 75.61% vs 70.43%.
- Completion with 40% masked inputs: Mamba-CAD A_p 98.91 and IR 8.03% vs DeepCAD 69.77% and 42.36%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Mamba-CAD instead of DeepCAD (++) | CAD sequence reconstruction · CAD sequence reconstruction | A_p: 87.68 -> 99.93 | Helped |
| Use Mamba-CAD instead of DeepCAD (++) | CAD sequence completion · CAD sequence completion | A_p: 71.27 -> 98.91 | Helped |
| Use Mamba-CAD instead of DeepCAD (++) | Fusion360 reconstruction · Fusion360 reconstruction | A_p: 87.72 -> 98.21 | Helped |
| Use Mamba-CAD instead of HNC-CAD | Random CAD generation · Random CAD generation | MMD: 2.65 -> 1.94 | Helped |
| Replace Transformer blocks with Mamba blocks (ablation) | Mamba-CAD · CAD sequence reconstruction | A_p: 83.41 -> 99.93 | Helped |
| Use CNN layers instead of MLPs in Compress Block and Scale Block (ablation) | Mamba-CAD · CAD sequence reconstruction | A_p: 83.95 -> 99.93 | Helped |
| Remove the added bottleneck layer between Encoder and Decoder (ablation) | Mamba-CAD · CAD sequence reconstruction | A_p: 99.57 -> 99.93 | Helped |
Surprising findings
- Replacing Mamba blocks with Transformer blocks (variant A) yields the weakest reconstruction, with average length 26.93, showing the Mamba backbone is the main driver of gains.
Limitations
- Only supports a limited set of CSG commands (sketch and extrude); Bezier curves and BRep operations such as fillet and chamfer are not covered.
- Unconditional generation only; no text, image or point-cloud conditioning.
Experimental setup
Datasets: DeepCAD ABC dataset Fusion360 Gallery reconstruction dataset
Benchmarks: Mamba-CAD dataset Fusion360 reconstruction dataset
Metrics: A_c A_p MCD IR AL L>=60 SR MMD JSD COV
Baselines: DeepCAD DeepCAD (++) TM-CAD SkexGen HNC-CAD
Builds on: Mamba DeepCAD WGAN SkexGen HNC-CAD
Closest papers by meaning
- CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation (CVPR 2025 · similarity 0.88)
- Text2CAD: Generating Sequential CAD Designs from Beginner-to-Expert Level Text Prompts (NeurIPS 2024 · similarity 0.87)
- Generating CAD Code with Vision-Language Models for 3D Designs (ICLR 2025 · similarity 0.86)
- From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.86)
- CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images (CVPR 2025 · similarity 0.86)
- Text-to-CAD Generation Through Infusing Visual Feedback in Large Language Models (ICML 2025 · 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.