Objective To improve the efficiency of digitally transforming traditional patterns of the Ruyuan Yao people into modern footwear design, this study establishes a systematic analytical framework and innovative design methods to explore pathways for converting intangible cultural heritage elements into contemporary product design, thus promoting the activation and innovative use of traditional cultural resources.

Methods A multi-stage mixed-methods approach was used to systematically integrate user needs, pattern digitization, design generation and multi-stakeholder evaluation. In the user-needs analysis stage, BERTopic modeling (Bidirectional Encoder Representations-based Topic Modeling, BERTopic) was applied to e-commerce review texts of loafers to extract user opinions and identify two core dimensions—functionality and aesthetics—as well as their respective priorities. In the pattern-processing stage, a Transformer-based edge detector (EDTER, Edge Detection Transformer) was employed to extract high-fidelity line drawings from traditional patterns and convert them to vector form, which was further supported by Otsu thresholding, morphological processing, and manual refinement to build a standardized and reusable library of Yao patterns. In the design-generation stage, parametric shape grammar was combined with convolutional style-transfer and embedded into a Comfy UI workflow; multimodal constraints were introduced using CLIP, IP-Adapter, and ControlNet, and FLUX models were used for image repainting. Three design strategies—"Fashionable National-Trend" "Modern Minimalist" and "Classic Retro"—were adopted to batch-generate loafer design schemes. In the evaluation stage, the schemes were assessed by a multi-stakeholder panel consisting of 30 industry experts and graduate students in design, along with 25 in-depth interviewees. The Fuzzy Analytic Hierarchy Process (F-AHP) was employed to determine indicator weights with consistency checks, and the Weighted Rank-Sum Ratio (WRSR) method was used to grade and rank the generated schemes.

Results Four main findings were obtained. First, the user evaluation analysis showed that consumers' functional concerns mainly involve wearing comfort, product quality and service experience, while aesthetic concerns focus on design quality, usage scenarios, and cultural dissemination. Accordingly, comfort and quality should be prioritized as baseline improvements, while style and cultural expression serve as premium uplift paths. Second, pattern digitization performed efficiently in a Linux environment with an NVIDIA RTX3060 GPU: the EDTER algorithm processed 1024×1024 images in about 0.5 seconds on average, extracting intact primary strokes and clear edges; the vectorized outputs can be stably adapted to key shoe components such as toe caps, central vamp panels, and welt strips. Third, the design schemes generated based on the three design strategies and multiple craft combinations improved breathability, fit, and visual recognizability, while producing visual features with brand communication potential. Fourth, the evaluation instrument showed good validity and reliability (KMO=0.797; Bartlett's test, P=0.000). WRSR grading identified schemes I4 and I12 as the top tier; both characterized by coordinated pattern layout and reinforced craft at key regions, exhibiting strong cultural expression and brand recognition. Schemes I2, I7, I15, I19 and I29 fell into the second tier, while I22 was rated unacceptable, indicating that some generated schemes still require optimization.

Conclusion The study demonstrates that, after satisfying the basic needs of comfort, durability and service reliability, the application of high-fidelity pattern digitization combined with multimodal-constrained generation methods can significantly enhance the depth and consistency of cultural expression in loafers, effectively balancing design efficiency with controllability of outputs. The proposed paradigm “user-needs driven → pattern digitization → scheme generation → multi-stakeholder evaluation” shows strong systematicity and cross-category transfer potential. However, the current research is limited by the training-data coverage and cross-domain transfer capability of the models. Future work should construct a semantically annotated multimodal cultural-pattern dataset, explore lightweight generation and human−AI co-creation mechanisms, and extend the validation to material selection and manufacturing processes to advance large-scale, commercialized and brandable applications of intangible cultural heritage elements in fashion consumer products.