Pinan Wu and Yingjuan Dong
CLIP, style transfer, personalised design, visual elements
With the development of deep learning technology and the growing demand for multimedia content creation, the style transfer of visual elements has become an important bridge connecting human– computer interaction and artistic expression. However, traditional style conversion methods are often limited to single-modal informa- tion, which makes it difficult to accurately capture and transmit cross-domain knowledge, resulting in the need for more realism and distortion of the generated results. This study constructs an efficient and reliable visual element style transfer and personalised design scheme by taking advantage of the contrastive language-image pre- training (CLIP) model’s cross-modal matching ability and generali- sation performance advantages. First, the original CLIP architecture is improved, and a sub-network for extracting latent style features of images is added to strengthen the model’s understanding of abstract aesthetic cognition. Experimental results show that when evaluated on multiple public benchmark datasets, compared with unoptimised versions and other mainstream designs, our method can significantly improve the diversity and fidelity of composite images, and the quantitative index Frechet inception distance (FID) () score dropped to the industry-leading level, with an average value below 4.8. Given the characteristics of diversified consumer tastes in the real world, the research further puts forward the design concept based on user feedback loop iterative optimisation. With the help of an advanced recommendation engine and A/B testing strategy, the system can intelligently filter out the style combination that best meets personal preferences according to each visitor’s past behaviour trajectory and then provide tailor-made customised services. Statistical analysis shows that the customer retention rate has increased by nearly 20% after adopting this process.
Important Links:
Go Back
