Scholarlink

<p>Photorealistic audio-driven talking-head synthesis is pivotal for immersive human–computer interaction, yet faces challenges in semantic alignment, personalization, and efficiency. This paper introduces GaPTalk, a novel framework leveraging precision-controlled 3D Gaussians for personalized talking face generation. GaPTalk integrates a contextual audio-to-expression encoding module, a precision-controlled 3D Gaussian rendering module, and a global–local inpainting module for seamless head-body reenactment. Extensive evaluations demonstrate that GaPTalk outperforms state-of-the-art methods in visual quality, lip-sync accuracy, and computational efficiency, achieving the fastest training time (&lt; 1&#xa0;h) and real-time inference. Here, we show that GaPTalk provides a robust and efficient solution for high-quality personalized digital human synthesis. Code and video demos are available at <a href="https://github.com/davisleelx/GaPTalk">https://github.com/davisleelx/GaPTalk</a>.</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

GaPTalk: precision-controlled 3D Gaussian rendering for personalized talking-head synthesis

  • 期刊论文

  • 发表日期：

    2025年10月7日

查看全文
  • Xiaoli Huang,
  • Xiang Li,
  • Shuangyi Tan,
  • Wuhui Chen,
  • Guanbin Li

摘要

Photorealistic audio-driven talking-head synthesis is pivotal for immersive human–computer interaction, yet faces challenges in semantic alignment, personalization, and efficiency. This paper introduces GaPTalk, a novel framework leveraging precision-controlled 3D Gaussians for personalized talking face generation. GaPTalk integrates a contextual audio-to-expression encoding module, a precision-controlled 3D Gaussian rendering module, and a global–local inpainting module for seamless head-body reenactment. Extensive evaluations demonstrate that GaPTalk outperforms state-of-the-art methods in visual quality, lip-sync accuracy, and computational efficiency, achieving the fastest training time (< 1 h) and real-time inference. Here, we show that GaPTalk provides a robust and efficient solution for high-quality personalized digital human synthesis. Code and video demos are available at https://github.com/davisleelx/GaPTalk.