Scholarlink

<p>In this paper, we propose a reliable method to improve the latent space manipulation in Generative Adversarial Networks (GANs) by adopting transformer blocks and the Conjugate Gradient (CG) method. Through extensive experiments on facial attribute editing, our method demonstrates remarkable efficacy and superiority, outperforming existing state-of-the-art techniques in manipulation disentanglement and image quality. Notably, our approach offers models with reduced complexity. In a novel exploration, we extend our method to manipulate images in the maritime domain, a pioneering endeavor that underscores the versatility of our approach. To corroborate the efficiency of our method, we employed both StyleGAN2 and StyleGAN3, showcasing its adaptability across different GAN architectures. We prove that the presented approach guarantees a significant stride in the realm of latent space manipulation, offering not only advanced capabilities in facial attribute editing but also opening avenues for less complex models and pioneering exploitation in the maritime domain.</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

Revolutionizing GANs: a cutting-edge approach to latent space manipulation with transformer blocks and Conjugate Gradient methodology

  • 期刊论文

  • 发表日期：

    2025年12月13日

查看全文
  • Walid Messaoud,
  • Rim Trabelsi,
  • Adnane Cabani,
  • Fatma Abdelkefi

摘要

In this paper, we propose a reliable method to improve the latent space manipulation in Generative Adversarial Networks (GANs) by adopting transformer blocks and the Conjugate Gradient (CG) method. Through extensive experiments on facial attribute editing, our method demonstrates remarkable efficacy and superiority, outperforming existing state-of-the-art techniques in manipulation disentanglement and image quality. Notably, our approach offers models with reduced complexity. In a novel exploration, we extend our method to manipulate images in the maritime domain, a pioneering endeavor that underscores the versatility of our approach. To corroborate the efficiency of our method, we employed both StyleGAN2 and StyleGAN3, showcasing its adaptability across different GAN architectures. We prove that the presented approach guarantees a significant stride in the realm of latent space manipulation, offering not only advanced capabilities in facial attribute editing but also opening avenues for less complex models and pioneering exploitation in the maritime domain.