Scholarlink

<p>To overcome the limitations of traditional generation methods in effectively integrating cultural symbols with personalized design, this study explores a virtual character generation approach capable of accurately conveying cultural intellectual property (IP) connotations, thereby enhancing consumer identification with cultural IP virtual characters. The study first introduces the algorithmic structures of Generative Adversarial Network (GAN) series and then proposes a virtual character generation method based on the LAnguage-Free traIning for Text-to-image gEneration (LAFITE) model. LAFITE overcomes conventional models’ dependence on large-scale paired text-image data by adopting a language-free training strategy, enabling precise capture and creative combination of cultural IP elements through deep mining of visual semantic relationships and patterns within images. In the first stage, pseudo-text feature synthesis is used to enhance cultural symbol retrieval, mapping input text into multimodal cultural feature vectors. In the second stage, pseudo-text feature refinement implements contrastive latent optimization, dynamically adjusting cultural feature weights to ensure that generated characters conform to cultural norms while expressing personalized traits. Experiments on the AnimeFace dataset show that LAFITE significantly outperforms traditional methods in quantitative metrics, achieving a Fréchet Inception Distance (FID) of 4.21 and an Inception Score (IS) of 7.85. Lower FID values indicate that the distribution of generated images closely matches real data, while higher IS values reflect superior image quality. Regarding text-image alignment, LAFITE achieves a Contrastive Language-Image Pre-Training Score of 0.287, substantially higher than StyleGAN2’s 0.215, demonstrating more precise capture of cultural semantics. The results indicate that LAFITE is superior for virtual character design and exhibits enhanced text-image alignment capabilities. Overall, this approach improves the efficiency of virtual character generation and effectively enhances consumer identification with cultural IP virtual characters.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The construction of cultural IP virtual characters and consumer identification based on generative adversarial networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Huiya Xing,
  • Xingmiao Wang,
  • Xiangyi Li,
  • Yiming Wang

摘要

To overcome the limitations of traditional generation methods in effectively integrating cultural symbols with personalized design, this study explores a virtual character generation approach capable of accurately conveying cultural intellectual property (IP) connotations, thereby enhancing consumer identification with cultural IP virtual characters. The study first introduces the algorithmic structures of Generative Adversarial Network (GAN) series and then proposes a virtual character generation method based on the LAnguage-Free traIning for Text-to-image gEneration (LAFITE) model. LAFITE overcomes conventional models’ dependence on large-scale paired text-image data by adopting a language-free training strategy, enabling precise capture and creative combination of cultural IP elements through deep mining of visual semantic relationships and patterns within images. In the first stage, pseudo-text feature synthesis is used to enhance cultural symbol retrieval, mapping input text into multimodal cultural feature vectors. In the second stage, pseudo-text feature refinement implements contrastive latent optimization, dynamically adjusting cultural feature weights to ensure that generated characters conform to cultural norms while expressing personalized traits. Experiments on the AnimeFace dataset show that LAFITE significantly outperforms traditional methods in quantitative metrics, achieving a Fréchet Inception Distance (FID) of 4.21 and an Inception Score (IS) of 7.85. Lower FID values indicate that the distribution of generated images closely matches real data, while higher IS values reflect superior image quality. Regarding text-image alignment, LAFITE achieves a Contrastive Language-Image Pre-Training Score of 0.287, substantially higher than StyleGAN2’s 0.215, demonstrating more precise capture of cultural semantics. The results indicate that LAFITE is superior for virtual character design and exhibits enhanced text-image alignment capabilities. Overall, this approach improves the efficiency of virtual character generation and effectively enhances consumer identification with cultural IP virtual characters.