StyleGAN-CLIP Cross-Modal Generative Model Based Automatic Design and Dissemination Effect Optimization for Regional Culture-Oriented Apparel Patterns
作者
Zhen Zhang,Yunlei Chen,Jialang Zhang
标识
DOI:10.1145/3795154.3795246
摘要
The preservation and dissemination of cultural heritage through modern design faces dual challenges: maintaining cultural authenticity while achieving market appeal. This paper proposes a novel AI-driven framework that integrates cross-modal generation with reinforcement learning-based dissemination optimization for culturally-grounded pattern design. We construct a comprehensive database of 3,000 annotated cultural patterns spanning eight Chinese regional categories (Dunhuang, Miao, Tibetan, Mongolian, Jiangnan, Lingnan, Hakka, and Xinjiang). Our approach combines a fine-tuned CLIP encoder for cultural semantic understanding with StyleGAN2 for high-fidelity pattern synthesis, establishing precise text-to-visual mappings that preserve nuanced cultural features. To bridge the gap between cultural authenticity and commercial viability, we introduce a reinforcement learning optimization module that leverages real-world A/B testing metrics (click-through rate, share rate, and dwell time) to iteratively refine generation policies. Extensive experiments demonstrate that our method achieves 91% cultural classification accuracy and 27.9 FID score, substantially outperforming baseline approaches including vanilla StyleGAN (62% accuracy), Stable Diffusion (71%), and StyleCLIP (74%). Real-world deployment on a major e-commerce platform with 50,000 users validates practical effectiveness: our framework improves CTR by 40.6%, share rate by 33.3%, and average dwell time by 36.6% compared to human-designed patterns, while compressing design cycles from months to approximately two hours. Ablation studies confirm that CLIP fine-tuning contributes the largest performance gain (+13% accuracy), while RL optimization enhances both cultural fidelity and market transmission (+4% accuracy, +10 dissemination index points). This work establishes a scalable paradigm for AI-assisted cultural creative design that harmonizes heritage preservation with contemporary market demands, offering practical tools for the cultural creative industry.