Research map / 3D vision / Face and shape

StepCAD: Mesh-to-CAD Code Generation via LLM Policy and Geometry-Guided Search

NeurIPS 2026 · 3D vision › Face and shape · approach: 3D shape generative modeling · 0 citations (OpenAlex)

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 StepCAD, a state-conditioned Qwen2.5-Coder-1.5B CAD policy plus IoU-guided beam search over local program edits, trained on the new 1.5M-program ARCADE dataset, for mesh-to-CadQuery reconstruction.

Problem

Recovering executable CAD construction programs (CadQuery) from 3D meshes is hard due to long-horizon compositional structure and precise continuous parameters; existing methods predict one-shot sketch-extrude programs with limited operation diversity and no correction.

Approach

Autoregressive policy conditioned on point clouds of both target and intermediate geometry predicts CAD actions step by step, followed by a geometry-aware beam search that refines the program via local edits (continuous parameter refinement, sketch modification, complement cut, operation skipping) ranked by IoU against the target.

Key results

• StepCAD reaches 0.73 IoU on ABC Hard vs. 0.31-0.39 for the strongest baselines (87.2% relative gain) and 0.89 IoU on Mechanical.
• Program-space search adds about +0.134 to +0.142 IoU over the policy alone (0.682 to 0.816 on the ablation set).
• The policy alone reaches 0.52 IoU on ABC Hard and maintains about 100% VSR across all benchmarks.

Surprising findings

• CADReasoner slightly outperforms the policy on the easiest benchmarks (DeepCAD, ABC Easy), but degrades substantially on harder subsets.
• Released CADEvolve checkpoint attains only 0.568 IoU on the ablation set, below a single-shot Qwen2.5 trained on ARCADE-1.5M (0.644), though the authors note model and setup differences confound this comparison.

Limitations

• Refinement is slow (~1 minute per sample) and relies on a fixed set of local edits and finite beam, which may miss corrections.
• IoU and CD measure geometric agreement, not original design intent, constraints, or parameter recovery; alternative valid histories complicate comparison.
• Training is on synthetic data, creating a domain gap; policy action cap of 12 limits generation length.

Experimental setup

Datasets: ARCADE-1.5M ABC DeepCAD Fusion 360 MCB Objaverse CADBench

Benchmarks: CADBench (DeepCAD, Fusion 360, ABC Easy/Medium/Hard, Mechanical, Organic)

Metrics: IoU Chamfer Distance Valid Shape Rate

Baselines: CAD-Recode Cadrille CADEvolve CADReasoner Single-shot Qwen2.5 on ARCADE-1.5M

Builds on: DeepCAD CAD-Recode Point-MAE CadQuery

Closest papers by meaning

• cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning (ICLR 2026 · similarity 0.87)
• CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images (CVPR 2025 · similarity 0.87)
• Generating CAD Code with Vision-Language Models for 3D Designs (ICLR 2025 · similarity 0.86)
• GeoCAD: Local Geometry-Controllable CAD Generation with Large Language Models (NeurIPS 2025 · similarity 0.86)
• CAD-Coder: Text-to-CAD Generation with Chain-of-Thought and Geometric Reward (NeurIPS 2025 · similarity 0.86)
• Draw Step by Step: Reconstructing CAD Construction Sequences from Point Clouds via Multimodal Diffusion. (CVPR 2024 · similarity 0.86)