Construction and Verification of a Personalized Design Generation Algorithm for Light Luxury Jewelry Based on a Multi‐Constraint Shape Grammar Enhanced Diffusion Model

ABSTRACT This study aims to solve the problems that the demand for personalized customization in the field of luxury jewelry design is difficult to coordinate with strict structural esthetic norms efficiently, and the existing generation algorithms are not precise enough in the control of complex topological structures. It proposes a personalized design generation algorithm based on Multi‐Constraint Shape Grammar (MCSG) enhanced Diffusion Model (DM). The technological innovation of MCSG lies in the integration of explicit rule constraints and implicit deep generative features, establishing the first computable MCSG system. At the level of algorithm architecture, based on Latent Diffusion Model (LDM), a grammar guide module is designed. It maps the rule codes of shape grammar to the denoising network through cross‐attention mechanism, and realizes the physical level constraint on the generated contour closure and internal pattern details. With the Low‐Rank Adaptation (LoRA) technology, fine‐tune the luxury style to ensure that the generated results accurately match the simple texture of modern luxury esthetics and the personalized input of users while retaining the delicate charm of Asian traditional crafts. Ten independent experiments conducted on the OrnAsia dataset show that the Fréchet Inception Distance (FID) of the proposed model is 12.45 ± 0.21. The Kernel Inception Distance (KID) is 3.15 ± 0.09. The Contrastive Language–Image Pre‐training score (CLIP score) is 32.68 ± 0.18. The Inception Score (IS) is 18.95 ± 0.15. In the geometric structure evaluation, the closed‐loop rate reaches 96.80% ± 0.39%, the non‐manifold error rate drops to 1.20% ± 0.11%, and the manufacturability index reaches 0.910 ± 0.007. Compared with the optimal comparison model in each evaluation dimension, the differences of related indexes all reach statistical significance level after correction by Holm–Bonferroni, indicating that the performance improvement of the model remained stable under different random seeds and sampling batches.