Abstract:

Objective Traditional garment pattern-making relies heavily on expert knowledge and long-term training， while generative artificial intelligence designs often lack structural information for production. This study aims to develop an engineering method method for dresses that integrates user-specific body shape constraints with a diffusion model， enabling direct transformation of three-dimensional body data into manufacturable two-dimensional patterns. The approach is expected to reduce professional barriers and supports rapid customization for small and medium-sized enterprises.

Method A three-stage workflow was established. First， a parametric human model and base garment model were built， before being transferred to nine personalized body shapes via landmark-based registration and mesh deformation. Second， multi-angle views of the three-dimensional garment were converted into Canny and Depth constraints using ControlNet， to guide a Stable Diffusion model for generating design images with clear structural lines. Finally， structural lines were mapped onto the three-dimensional garment， which was segmented and unfolded into two-dimensional patterns using a mass-spring system.

Results Single-factor experiments showed that the Canny and Depth dual-constraint scheme minimized blurriness of the structural lines and produced the most natural lines， with form deviation below one centimeter in fitted areas. Single constraints often caused spatial reduction or structural deviation， while posture control alone lacked garment structure information. Unfolding deformation was quantified across nine key dimensions using reverse analysis based on the short-measure prototype method. Shoulder width exhibited high stability with a standard deviation of 0.02， whereas the side-neck-to-bust-point length and front-armpit-to-waistline length were more sensitive to partition line placement， with standard deviations of 0.28 and 0.27 respectively. Placing partition lines near the bust point， where Gaussian curvature is high， reduced overall deformation energy during unfolding. Deformation was not uniform but concentrated in areas opposite the partition lines. Virtual fitting in CLO 3D software demonstrated that the final garments fitted the parametric models well， with smooth contour lines and appropriate ease. Simulated garment pressure ranged from 1.2 to 2.0 kPa， within the human comfort range. Physical garment validation confirmed feasibility： key measurements including bust and waist circumference showed an average error below 1.5% relative to design specifications. The ready-to-wear garment exhibited good fit without excessive tightness or wrinkles， and the wearer reported sufficient freedom for routine movements.

Conclusion A digital workflow integrating parametric modeling， controllable diffusion models， and physical unfolding was established for reverse garment engineering. It bridges generative artificial intelligence visual concepts and producible structural patterns while maintaining fit accuracy and style integrity. The method reduces reliance on traditional pattern-making expertise and enables efficient personalized customization， particularly for small and medium-sized enterprises. However， the approach faces challenges with highly complex structures such as multi-layered or densely gathered designs. A rigorous mathematical model to quantify the influence of segmentation line placement on deformation rates has not yet been developed. Future work should focus on building a quantitative relationship between segmentation lines and surface deformation energy， expanding the fabric physical parameter database， and introducing curvature maps to guide segmentation line generation near key body points， combined with material-specific compensation strategies. These improvements will enhance the method's general

Key words: garment pattern-making, dress, structural design, diffusion model, reverse engineering