CADEvolve: Creating Realistic CAD via Program Evolution

TL;DR

CADEvolve tackles the data scarcity challenge in CAD automation by introducing an offline evolution-based data synthesis pipeline that progressively builds rich, multi-operation CadQuery programs from simple primitives. A three-tier CADEvolve-3L dataset (G, P, C) and the CADEvolve-M model enable state-of-the-art Image2CAD reconstruction across multiple benchmarks through supervised fine-tuning and reinforcement learning with geometry-aware rewards. The approach combines LLM-guided proposals, staged validation, canonicalization, and diverse augmentations to produce a scalable, executable CAD corpus, addressing limitations of sketch–extrude corpora and restricted operator sets. The results demonstrate improved reconstruction fidelity and generalization, suggesting that principled data synthesis can substantially boost multimodal CAD workflows and other program-synthesis tasks in engineering domains.

Abstract

Computer-Aided Design (CAD) delivers rapid, editable modeling for engineering and manufacturing. Recent AI progress now makes full automation feasible for various CAD tasks. However, progress is bottlenecked by data: public corpora mostly contain sketch-extrude sequences, lack complex operations, multi-operation composition and design intent, and thus hinder effective fine-tuning. Attempts to bypass this with frozen VLMs often yield simple or invalid programs due to limited 3D grounding in current foundation models. We present CADEvolve, an evolution-based pipeline and dataset that starts from simple primitives and, via VLM-guided edits and validations, incrementally grows CAD programs toward industrial-grade complexity. The result is 8k complex parts expressed as executable CadQuery parametric generators. After multi-stage post-processing and augmentation, we obtain a unified dataset of 1.3m scripts paired with rendered geometry and exercising the full CadQuery operation set. A VLM fine-tuned on CADEvolve achieves state-of-the-art results on the Image2CAD task across the DeepCAD, Fusion 360, and MCB benchmarks.

Figures (11)

• Figure 1: CADEvolve overview. (a) Representation of a shape tuple; (b) seed pool of 46 hand-written generators; (c) VLM proposals conditioned on sampled parents; (d) retrieval-augmented code synthesis; (e) staged validation (execution check, geometry validity, visual–text agreement) with targeted repair; (f) selection and growth of the accepted pool.
• Figure 2: Examples of generated parts. A diverse gallery of accepted CADEvolve-G outputs spanning extrude, revolve, loft, sweep, shell, fillet, chamfer, booleans, and local patterns. Colors are arbitrary.
• Figure 3: Representation of a shape from CADEvolve-G. The representation consists of descriptive textual fields, a Python code that maps geometric parameters to a 3D shape S, and the list of parents from which S has evolved via the CADEvolve algorithm.
• Figure 4: From generator to concrete program. Given a (A) parametric generator $\textit{param2cq}$ and a sampled parameter vector $\mathbf{z}$, we (B) bind parameters (param_i = z["param_i"]), then execute the generator once to resolve conditionals and loops, skipping untaken branches and retaining only geometry-affecting CadQuery operations; (C) apply code–level augmentation; and (D) emit a flat, deterministic script with minimal imports and a unified output (result = …). The script reproduces the shape exactly while exposing its construction history.
• Figure 5: Qualitative comparisons across datasets. Columns: DeepCAD, Fusion360, MCB. Rows (top$\to$bottom): the cadrille baseline, our CADEvolve– C big (RL2) prediction, and the target render. This panel best illustrates the advantages of the CADEvolve dataset: targets include parts built via revolve, sweep, loft, face selectors, and complex hole patterns that cannot be well approximated by the sketch- extrude-boolean schemes used in CAD-Recode and many other datasets, where cadrille typically fails but CADEvolve-M closely reconstructs the input shape.