Source-linked AI summary
CADEvolve: Creating Realistic CAD via Program Evolution
Maksim Elistratov, Marina Barannikov, Gregory Ivanov, Valentin Khrulkov, Anton Konushin, Andrey Kuznetsov, Dmitrii Zhemchuzhnikov
TL;DR
CAD automation is limited by datasets dominated by simple sketch–extrude histories and by weak single-pass VLM grounding. CADEvolve evolves validated CAD programs into a multi-tier training corpus, and its fine-tuned model achieves state-of-the-art Image2CAD performance across DeepCAD, Fusion 360, and MCB. The dataset is synthetic and is not intended to match any single proprietary industrial CAD distribution.
Problem
Public CAD data lacks complex operations, multi-operation composition, and design intent, while existing geometry datasets generally lack CAD sequences.
Method
CADEvolve uses evolution, VLM-guided program generation, validation, sampling, canonicalization, and geometry-based quality controls to construct executable CAD training data.
Results
CADEvolve achieves state-of-the-art Image2CAD performance on DeepCAD, Fusion 360 Gallery, and MCB.
Takeaways & Limitations
The resulting CAD sequence corpus covers the full CAD operation set with executable multi-operation histories and supports strong reconstruction performance across multiple benchmarks.
Takeaways & Limitations
CADEvolve is synthetic, so its shape and operation frequencies may differ from real-world data and it does not claim fidelity to any particular industrial domain.
Abstract
from arXiv · showhide
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.
1. Introduction
CADEvolve addresses the lack of complex, executable CAD training data by evolving validated multi-operation programs and organizing them into a three-tier corpus. Models trained on the resulting data achieve state-of-the-art Image2CAD performance across three benchmarks.
- Data gap: Public CAD corpora predominantly contain sketch–extrude histories, omitting richer operations, multi-operation composition, and design intent.Frozen VLMs also tend to produce simple shapes and struggle to chain heterogeneous operations reliably.
- Pipeline: CADEvolve uses an evolutionary propose–execute–filter pipeline in which VLM-generated CAD programs are validated and selected iteratively.The pipeline is motivated by the limitations of single-pass VLM generation.
- Results: Models trained on the resulting data achieve state-of-the-art Image2CAD performance on DeepCAD, Fusion 360 Gallery, and MCB.The training process includes sampled programs, canonicalization, augmentation, and geometry-derived reward fine-tuning.
- Dataset: The corpus comprises parametric generators, executable sampled programs with geometry, and canonicalized training scripts.These tiers are named CADEvolve-G, CADEvolve-P, and CADEvolve-C, respectively.
- Contribution: CADEvolve-3L is presented as an open CAD sequence dataset covering the full CADQUERY operator set with executable multi-operation histories.Its associated policy supports the full operator set for Image2CAD.
2. Related Work
Prior CAD generation methods use CSG, B-reps, or editable program sequences, while retrieval and evolutionary approaches address limited priors or one-shot generation. CADEvolve instead applies evolution during data synthesis to release executable histories spanning multiple CAD operations.
- CAD representations: CSG methods compose primitives but struggle with the variety and detail of engineered parts, whereas B-rep generators are brittle and harder to edit.Sequence models better match parametric workflows and preserve editability.
- Dataset gap: CADEvolve targets the gap by releasing multi-operation, executable CAD histories.This contrasts with prior sequence corpora dominated by simpler operation sequences.
- Case-based generation: Case-based methods retrieve related designs and refine prompts, but their capabilities remain constrained by the available corpora.Reported support for richer operations does not remove the underlying data limitation.
- Evolutionary methods: Evolutionary coding methods propose, evaluate, and select candidates through validator loops, while EvoCAD applies evolution at inference.CADEvolve relocates evolution to the data stage rather than using it during inference.
- Image2CAD: Image2CAD methods reconstruct parametric programs from rendered views, and this paper uses the task as a controlled dataset-validation setting.Fixed multi-view renderings specify target geometry without requiring per-shape textual descriptions.
- RL for CAD reconstruction: CAD reinforcement learning uses geometry-based rewards either to optimize command sequences or to post-train code generators with verifiable signals.CADEvolve’s approach is evaluated within this broader reward-guided reconstruction landscape.
3. Dataset Generation and Processing
CADEvolve constructs complex parametric CAD generators through evolutionary proposals, validation, and quality-diversity sampling, then converts them into augmented and canonicalized executable programs for training. Processing addresses template collapse, scale variation, and geometric validity while producing a large training corpus.
- Evolutionary Synthesis: CADEvolve represents each shape with descriptive fields, a CADQUERY parameter-to-solid function, and an evolutionary parent lineage.The representation captures both design intent and inheritance history.
- Evolutionary Synthesis: The initial seed pool contains 46 hand-written generators spanning extrude, revolve, loft, sweep, shell, fillet, chamfer, booleans, and local patterns.This seed pool establishes operator breadth and parameterization styles.
- Evolutionary Synthesis: The propose–execute–filter loop samples parents, elicits child metadata and code, validates execution, geometry, and visual–text agreement, then retains passing children.Targeted self-repair is prompted when a validation stage fails, and the process yields 7,945 validated generators.
- Generator Sampling: Quality-diversity search penalizes invalid or poorly scaled solids and candidates too close to archived samples, while CMA-ES finds 15 valid, diverse instances per generator.Accepted samples must be valid, novel, and within the specified geometric bounds.
- Program Parsing: Single-run tracing binds sampled parameters, resolves control flow, retains constructive CADQUERY operations, and emits deterministic flat scripts.The resulting CADEvolve-P scripts preserve exact geometry while exposing construction history.
- Augmentation: Template collapse caused scripts from each generator to share operation sequences, so validated semantically equivalent rewrites were generated before further augmentation.The initial model learned fixed code skeletons and produced unusable results; rewrites yielded 744,780 variants, while ABC and ShapeNet predictions expanded training to approximately 1.74 million scripts.
- Canonicalization: Canonicalization removes non-geometric code, centers and scales solids, and revalidates collision-prone transformations before training.Post-transform validation retained 1,002,002 programs, and length filtering plus deduplication produced 961,300 scripts shorter than 3k characters.
- Canonicalization: The post-canonization Qwen2-VL-2B baseline improved markedly but remained below state-of-the-art, with limited sketch diversity identified as an error source.This motivates subsequent reinforcement-learning fine-tuning described in the evaluation section.
4. CADEvolve-M: Program-Generating Policy
CADEvolve-M evaluates Image2CAD with multi-view renders decoded into CadQuery programs, using supervised fine-tuning followed by geometry-rewarded online reinforcement learning. Canonicalized data, prediction-derived supervision, and RL improve performance across DeepCAD, Fusion360, and MCB, with domain-specific adaptation benefiting MCB.
- 4. CADEvolve-M: Program-Generating Policy: Image2CAD maps fixed multi-view renders directly to executable CadQuery programs using the VLM’s built-in visual encoder.CADEvolve-P uses seven views, while CADEvolve-C/RL uses eight canonical views; no extra image backbone or adapter is introduced.
- 4. CADEvolve-M: Program-Generating Policy: Training proceeds from token-level supervised fine-tuning on view–code pairs to online RL with IoU rewards and invalidity penalties.The RL configurations run for 20 epochs and use Dr. GRPO and Dr. CPPO variants.
- 4.5. Results: Canonicalized CADEvolve-C regimes outperform earlier augmented CADEvolve-P SFT, with substantial gains from SMALL to MIDDLE and further improvement in BIG.The results support prediction-derived ABC/ShapeNet supervision and canonicalization as important additions beyond semantics-preserving code rewrites.
- 4.5. Results: RL from the CADEvolve-C BIG checkpoint improves CD and IoU over cadrille RL across all three datasets, with a slightly higher invalid rate in RL1.The higher invalid rate is reported alongside more frequent use of complex, collision-prone operations.
- 4.5. Results: Adding MCB training shapes in RL2 substantially improves MCB performance while maintaining near-cadrille performance on DeepCAD and Fusion360.The MCB test split remains fixed, and the added training data addresses MCB’s smoother rendering domain.
5. Limitations
CADEvolve’s synthetic corpus is not intended to reproduce any single industrial CAD distribution, and its programs are expressed specifically in CadQuery. These boundaries limit claims about industrial distribution fidelity and straightforward transfer to other CAD systems.
- Synthetic distribution mismatch: CADEvolve’s synthetic shape and operation frequencies may differ from real-world data, so the dataset does not claim fidelity to any particular industrial domain.The authors nevertheless report improved generalization across multiple benchmarks, including stronger MCB performance.
- CadQuery dialect scope: CADEvolve programs use the CadQuery dialect, and faithful conversion to other CAD systems may be non-trivial.Differences in feature-history representations, kernel behaviors, and constraint semantics create the stated portability boundary.
6. Conclusion
CADEvolve is presented as a general method for synthesizing high-quality supervision when open corpora are scarce, instantiated through a CAD dataset covering the full operation set. The resulting dataset achieves state-of-the-art Image2CAD performance and may support broader multimodal CAD tasks.
- CADEvolve synthesizes high-quality supervision when open corpora are scarce, with CAD as its application domain.
- CADEvolve-3L is the first CAD sequence corpus covering the full CAD operation set.
- State-of-the-art Image2CAD performance is achieved using the resulting dataset.
- The same data foundation could further support PC2CAD/Scan2CAD, Text2CAD, and broader multimodal CAD pipelines.
A. Evolutionary synthesis trajectories
The evolutionary pipeline starts from simple CAD primitives and applies incremental LLM-proposed code edits. Accepted refinements form branching trajectories whose shapes become progressively more complex through multi-operation composition.
- The pipeline starts from simple seed primitives and applies incremental code edits that introduce new operations and structural detail.
- Accepted candidates become parents for subsequent iterations, producing branching lineages and occasional recombination across design directions.
- Generated CAD parts become progressively more complex through compositional multi-operation edits over iterations.
B. Novelty–validity dynamics
Search efficiency degrades in later iterations: invalid proposals increase sharply while accepted novel samples decline. Continuing exploration therefore faces a practical validity and resource trade-off.
- ∼85% invalid proposals are reached in later iterations under strict validation rules.
- 40–50% of novel samples are accepted as the search progresses, indicating diminishing returns under fixed API or compute budgets.
- Extending the search would likely require stronger proposal strategies rather than simply running longer.
C. Rotational augmentation
Rotational augmentation targets pose-independent CAD understanding by generating randomly rotated script variants. This addresses reliance on canonical poses and supports generalization to unseen orientations.
- The same CAD part may be stored or observed under arbitrary rotations while retaining its construction logic and parameterization.
- Without rotational augmentation, reliance on dataset-specific canonical poses can reduce generalization.
- Randomly rotated script variants encourage attention to pose-independent geometric and procedural cues.
D. Comparison to existing benchmarks
CADEvolve is characterized against existing CAD benchmarks using operation coverage, procedural depth, and geometric complexity. Its parts show broad operator coverage alongside substantially deeper programs and more detailed geometry than typical benchmarks.
- Comparison framework: Together, operation occurrence, sequence length, and face count capture operator coverage, procedural depth, and geometric complexity.These complementary statistics provide the basis for comparing CADEvolve with existing CAD benchmarks.
- Operation occurrence: Operation distributions broadly resemble real CAD histories, but CADEvolve uses fewer revolve, chamfer, shell, and mirror operations and more hole, transform, and loft operations.The most frequent operators remain sufficiently represented for training and evaluation.
- Sequence length: CADEvolve scripts span a wide range of lengths, including many programs exceeding 25 operations and a long tail of highly procedural models.Sequence length measures the number of CAD operations per script and indicates greater procedural depth than typical benchmarks.
- Face count: CADEvolve parts frequently contain thousands of polygonal faces, reflecting fine-grained geometric detail and higher shape complexity than existing benchmarks.Face count measures the number of polygonal faces in the resulting geometry.