GenCAD-Self-Repairing: Feasibility Enhancement for 3D CAD Generation

Recent breakthroughs in generative AI, notably diffusion models and Transformer architectures, have revolutionized image, text, and 3D model synthesis. In CAD, early efforts like DeepCAD and Fusion 360 provided design histories but lacked scalable, high-quality generation methods. GenCAD integrated Transformer-based sequence modeling with contrastive learning, enabling image-conditioned CAD program generation. However, approximately 10% of outputs failed to produce valid B-rep models due to geometric inconsistencies, limiting industrial applicability. The scarcity of large, diverse CAD datasets further constrained progress. Overall, while these advances marked significant steps, ensuring the geometric feasibility and robustness of generated designs remains a key challenge for real-world deployment.

The main challenge lies in guaranteeing the geometric validity of generated CAD programs. Despite high overall accuracy, a persistent 10% failure rate due to invalid boundary representations hampers usability. Existing models lack mechanisms to dynamically correct infeasible outputs during generation, leading to wasted computational effort and limited practical application. Additionally, the scarcity of large-scale, annotated datasets restricts model training and generalization. Addressing these issues is crucial for transitioning AI-based CAD from research to industrial practice, where reliability and efficiency are paramount.

This work introduces several key innovations: • Incorporating a guided diffusion process that uses classifier and regressor feedback to steer latent vectors toward feasible regions during generation; • Developing a self-repair pipeline that employs a trained regressor to correct invalid latent vectors post-generation, verified via geometry kernel; • Constructing a specialized dataset with validity labels to train classifiers and regressors effectively. These innovations collectively enable the model to generate higher-feasibility CAD designs while maintaining geometric accuracy, representing a significant leap over traditional sequence-based generation methods.

• �� Dataset creation: sampling images, generating command sequences via GenCAD, labeling validity, and storing latent representations; • Classifier training: a multi-layer perceptron (MLP) trained to distinguish valid/invalid latent vectors; • Regressor training: a self-supervised model predicting corrections for invalid vectors, using paired invalid-valid data; • Diffusion guidance: during denoising, gradient feedback from classifier/regressor adjusts latent vectors, guiding them toward feasible regions; • Post-generation validation: CAD commands are checked via geometry kernel; invalid outputs trigger regressor correction, and the process repeats until feasibility is achieved. This pipeline ensures the generated designs are both geometrically valid and faithful to the input images.

The evaluation used a dedicated validation set of 8,515 B-rep images, comparing baseline GenCAD with variants incorporating the proposed guidance and repair mechanisms. Metrics included feasibility rate and MMD for geometric similarity. The results showed a feasibility increase from 93.1% to 97.0%, with over 65% of infeasible designs fixed. Ablation studies assessed the contribution of each component, confirming that guided diffusion and self-repair jointly improve feasibility. The latent space distribution was analyzed via PCA, confirming that the correction process preserves the original design diversity. Computational costs and potential scalability issues were also discussed.

The primary outcome was a feasibility rate of 97.0%, a 3.9% increase over baseline, fixing 65.84% of infeasible designs. MMD increased slightly from 0.180 to 0.200, indicating a minor decline in geometric accuracy but overall distribution consistency. The latent space analysis showed that the corrected samples remained within the original distribution, validating the method’s stability. These results demonstrate the effectiveness of guided diffusion combined with self-repair in industrial CAD generation, balancing feasibility and accuracy.

This approach can be directly applied in manufacturing, architecture, and product design, enabling automatic generation of valid CAD models from images or sketches. It reduces manual effort, accelerates prototyping, and improves design reliability. In the long term, integrating this framework into comprehensive CAD platforms could lead to fully autonomous design pipelines, transforming how products are developed and manufactured, especially in mass customization and rapid prototyping scenarios.

Despite promising results, the method struggles with highly complex geometries, where accuracy drops and computational costs increase. The reliance on geometry kernel validation limits real-time application in high-throughput environments. Moreover, the model’s robustness across diverse datasets and design styles remains to be validated, requiring further training and adaptation. Future work should focus on optimizing guidance efficiency, extending to more complex shapes, and enhancing generalization capabilities.