| description abstract | Abstract. Recent advancements in generative artificial intelligence have shown strong potential for automating computer-aided design (CAD) generation. Unlike commonly used point cloud or mesh representations in 3D generation tasks, CAD models typically require editable and manufacturable boundary representations (B-reps) that satisfy strict geometric and topological constraints. However, due to the hierarchical structure of B-reps, generative CAD models often use network-friendly intermediate representations, such as CAD command sequences, rather than directly operating on B-reps. This introduces a fundamental limitation: generated representations are not guaranteed to be convertible into valid B-reps by the geometry kernel, as the generative model remains unaware of the kernel’s convertibility constraints during generation. In this work, we propose GenCAD-Self-Repairing, a feasibility-aware CAD generation framework that explicitly incorporates geometry-kernel convertibility into the generative process. Our approach operates entirely in the latent space and steers latent representations toward feasible regions through feasibility-guided diffusion denoising combined with a latent-space self-repair mechanism. Experimental results demonstrate that our method significantly improves CAD generation feasibility by reducing infeasible outputs by 56.9% relative to a state-of-the-art baseline, while largely preserving geometric accuracy. | |
