GAN 与 Diffusion 在传统纹样设计中的实验研究.

Traditional patterns come as one of the vital components of China's rich cultural heritage embodying the wisdom and aesthetic memory of China. These patterns have been extensively used in various design fields. Artists and designers can draw nourishment and inspiration from the beautiful graphic decorations the rich implications of forms and the unique pattern designs. However traditional manual design methods can no longer meet the diverse and efficient demands of the modern pattern design. Current research on computer-aided pattern design primarily focuses on traditional methods and generative AI approaches. Traditional methods mainly generate new patterns by simulating image morphological features and quantifying image organizational characteristics. Generative AI methods on the other hand use deep neural networks for transfer learning to simulate the distribution of image data thus creating new pattern images and offering new paths and methods for traditional pattern design. While there is already a certain foundational body of research on the generative design of traditional patterns there are still issues in the field of generative technology application research. These include a lack of research from the perspective of universal generative design of traditional patterns neglect of the cultural and artistic foundations of these patterns insufficient attention to the practical application needs of generated patterns and a lack of comprehensive evaluation of generated patterns. To facilitate deep co-creation between designers and AI this paper explores the potential and application of image generation models in the innovative design of traditional patterns from an artistic design perspective. Four mainstream image generation models were initially selected through preliminary experiments on traditional pattern generation. Among these StyleGAN based on GAN and Stable Diffusion based on Diffusion were chosen for further experimentation. The technical aspects of the datasets training processes and model parameters were analyzed and pattern images were evaluated based on diversity clarity and text-image matching. Additionally a survey was conducted to assess the experimental results on five artistic design elements form color aesthetics innovation and application. Combining technical and artistic analyses the experimental results underwent comprehensive multidimensional evaluation. Finally the experimental results were validated from the perspective of design requirements and the superior performance of the two generative design methods in various aspects was explored. This provides case references for designers in selecting and using generative design methods and offers new research perspectives for traditional pattern design studies. The experimental results indicate that both models meet the basic requirements of artistic design. The StyleGAN model produces pattern images closer to the distribution of real images with higher image quality and diversity making it suitable for generating line patterns individual patterns and continuous patterns and meeting the needs for quick generation emphasizing formal beauty. In contrast the Stable Diffusion model better preserves the essence of traditional patterns balancing artistry and creativity and is more aligned with the artistic design needs of traditional patterns suitable for diversified and precise generation requirements and for cultural content emphasizing inheritance and innovation. This study provides an experimental analysis of the application of image generation models in traditional pattern design offering new research perspectives and methods for traditional pattern artistic creation. The findings will contribute to the deep application of generative AI in the design of ethnic and traditional patterns so as to promote the modern transformation of traditional pattern design. [ABSTRACT FROM AUTHOR]