LLM-driven personalized garment synthesis enhanced by neuro-symbolic RAG and physics-aware pattern optimization

Existing personalized garment generation systems suffer from a persistent tripartite gap among user intent sensing, generative design, and physics-aware fitting, preventing the production of garments that are simultaneously aesthetically personalized and physically manufacturable. Current AI-generated garments often fail to align with the user’s unique biometric and aesthetic profiles—such as skin tone compatibility and morphological features—and lack the adaptive geometric constraints necessary to ensure a precise physical fit across heterogeneous physiques. Without a bidirectional flow between intent sensing and manufacturing-ready engineering, these outputs remain stylistically arbitrary and physically unwearable. To bridge these gaps, this paper presents an LLM-driven garment synthesis framework enhanced by neuro-symbolic Retrieval-Augmented Generation (RAG) and physics-aware pattern optimization for the fully automated generation of personalized garments. The framework features an automated Requirement Analyst that translates a user’s image and semantic intent into refined design specifications. By extracting high-dimensional biometric and colorimetric profiles, the system identifies garment silhouettes and palettes that are inherently harmonized with the user’s unique physical characteristics, ensuring professional styling accuracy before generation. These specifications drive a dual-path Generative Designer to synthesize physically-based rendering (PBR) textures and 2D sewing patterns, which are then adaptively deformed by an Optimization-based Pattern Adapter. We develop a novel Finite-Difference Adam (FD-Adam) optimization algorithm to resolve mechanical conflicts and improve garment fit across representative body morphologies. Experimental validation on representative test cases supports the feasibility of the framework. Furthermore, the physics-driven fitting process effectively dissipates high-strain concentrations on diverse physiques, achieving a preference rate of 62% over two competitive baselines in a controlled double-blind user study. By enabling the direct transition from perceptual imagery to production-ready engineering assets, this work is compatible with standard industrial CAD workflows for offline personalized garment design. 1