Scholarlink

<p>The current approaches of creating realistic, high-quality landscape imagery mostly depend on labor-intensive manual design procedures. In an effort to simplify image production for video games, virtual reality, and motion pictures, a new hybrid model called ProWGAN, combining ProGAN and WGAN approaches, is employed for automated landscape synthesis. Five models (FCGAN, DCGAN, ProGAN, WGAN, and ProWGAN) were trained on a dataset of landscape images and compared using multiple evaluation metrics. Compared to traditional models, ProWGAN produces 128-128 size images with the best FID score (29.67), IS (5.11), and lowest critic loss (0.2), fully capturing landscape features in just 5&#xa0;h of training and 50 epochs. The layered method to producing images and progressive learning of ProGAN with the stability of WGAN’s Wasserstein distance showed superior ability to generate realistic landscape images. The results demonstrate how ProWGAN can revolutionize landscape image production by reducing manual work, lowering production time and effort and how a 2d image can be converted into 3d model via MeshRoom.</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

ProWGAN a hybrid generative adversarial network for automated landscape generation in media and video games

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年9月29日

查看全文
  • Kavya Sree Kammari,
  • Yasasree Lasya Annambhotla,
  • Manju Khanna

摘要

The current approaches of creating realistic, high-quality landscape imagery mostly depend on labor-intensive manual design procedures. In an effort to simplify image production for video games, virtual reality, and motion pictures, a new hybrid model called ProWGAN, combining ProGAN and WGAN approaches, is employed for automated landscape synthesis. Five models (FCGAN, DCGAN, ProGAN, WGAN, and ProWGAN) were trained on a dataset of landscape images and compared using multiple evaluation metrics. Compared to traditional models, ProWGAN produces 128-128 size images with the best FID score (29.67), IS (5.11), and lowest critic loss (0.2), fully capturing landscape features in just 5 h of training and 50 epochs. The layered method to producing images and progressive learning of ProGAN with the stability of WGAN’s Wasserstein distance showed superior ability to generate realistic landscape images. The results demonstrate how ProWGAN can revolutionize landscape image production by reducing manual work, lowering production time and effort and how a 2d image can be converted into 3d model via MeshRoom.