Scholarlink

<p>The increasing demand for lifelike character animation in digital media has driven the exploration of artificial intelligence techniques, particularly Generative Adversarial Networks (GANs), to enhance realism. However, existing GAN models often struggle to capture the dynamic and complex motion patterns required for high-quality animation. This study aims to compare the performance of three GAN variants, DCGAN, Pix2Pix, and StyleGAN, in generating realistic character animations using the Human 3.6&#xa0;M dataset. The models were trained under uniform hyperparameters and evaluated using Fréchet Inception Distance (FID), Inception Score (IS), and Mean Squared Error (MSE). Experimental results show that StyleGAN outperformed the others, achieving the lowest FID (29.80 ± 5.08), the highest IS (3.99 ± 0.22), and the lowest MSE (0.019 ± 0.005), indicating superior visual realism, diversity, and motion accuracy. These findings demonstrate that StyleGAN offers a more effective solution for realistic character animation, with practical implications for its integration into film, video game, and virtual environment production workflows.</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

Enhancing character animation realism with generative adversarial networks (GANs): a comparative method study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月29日

查看全文
  • Mars Caroline Wibowo,
  • Daniel Manongga,
  • Hendry Hendry,
  • Teguh Indra Bayu

摘要

The increasing demand for lifelike character animation in digital media has driven the exploration of artificial intelligence techniques, particularly Generative Adversarial Networks (GANs), to enhance realism. However, existing GAN models often struggle to capture the dynamic and complex motion patterns required for high-quality animation. This study aims to compare the performance of three GAN variants, DCGAN, Pix2Pix, and StyleGAN, in generating realistic character animations using the Human 3.6 M dataset. The models were trained under uniform hyperparameters and evaluated using Fréchet Inception Distance (FID), Inception Score (IS), and Mean Squared Error (MSE). Experimental results show that StyleGAN outperformed the others, achieving the lowest FID (29.80 ± 5.08), the highest IS (3.99 ± 0.22), and the lowest MSE (0.019 ± 0.005), indicating superior visual realism, diversity, and motion accuracy. These findings demonstrate that StyleGAN offers a more effective solution for realistic character animation, with practical implications for its integration into film, video game, and virtual environment production workflows.