Scholarlink

<p>Generative adversarial networks (GANs) are widely used in image generation tasks, forming one of the cornerstones of artificial intelligence generated content (AIGC), but the generated images usually lack texture details. In this paper, we propose a novel framework termed <i>progressively unfreezing perceptual GAN (PUPGAN)</i>, which can generate images with fine texture details. Particularly, we propose an <i>adaptive perceptual discriminator</i> to extract features effectively and measure the discrepancy between the generated and real images. In contrast to the perceptual loss employed in traditional GANs, the proposed adaptive perceptual discriminator can dynamically adjust its perceptual feature space to better suit various downstream image generation tasks. In addition, we propose an effective <i>progressively unfreezing scheme</i> that gradually unfreezes the parameters of the discriminator during the training process, ensuring a balanced and stable training process for the proposed model. Qualitative and quantitative experiments on three image generation tasks, i.e., single image super-resolution, paired image-to-image translation, and unpaired image-to-image translation, demonstrate the effectiveness of PUPGAN in AIGC.</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

Progressively unfreezing perceptual GAN

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Qi Guo,
  • Huiqiang Wang,
  • Rachid Hedjam,
  • Jinxuan Sun,
  • Yang Chen,
  • Junyu Dong,
  • Guoqiang Zhong,
  • Wei Xiang

摘要

Generative adversarial networks (GANs) are widely used in image generation tasks, forming one of the cornerstones of artificial intelligence generated content (AIGC), but the generated images usually lack texture details. In this paper, we propose a novel framework termed progressively unfreezing perceptual GAN (PUPGAN), which can generate images with fine texture details. Particularly, we propose an adaptive perceptual discriminator to extract features effectively and measure the discrepancy between the generated and real images. In contrast to the perceptual loss employed in traditional GANs, the proposed adaptive perceptual discriminator can dynamically adjust its perceptual feature space to better suit various downstream image generation tasks. In addition, we propose an effective progressively unfreezing scheme that gradually unfreezes the parameters of the discriminator during the training process, ensuring a balanced and stable training process for the proposed model. Qualitative and quantitative experiments on three image generation tasks, i.e., single image super-resolution, paired image-to-image translation, and unpaired image-to-image translation, demonstrate the effectiveness of PUPGAN in AIGC.