Scholarlink

<p>Image style transfer, a fundamental task in computer vision, aims to transform the visual style of a content image to match a desired artistic style. Traditional methods often require explicit reference style images, limiting their flexibility. This paper introduces a novel zero-shot, text-driven image style transfer framework that leverages a pre-trained CLIP model to establish semantic alignment between textual descriptions and visual features. Our approach enables direct mapping of stylistic attributes from language to the image domain without the need for style images. Key innovations include a CLIP-guided semantic adapter module, a cross-attention-based style modulation module, and a content-style decoupling module. Extensive experiments demonstrate that our method generates visually coherent results and outperforms existing text-driven and style transfer techniques, achieving a new benchmark in zero-shot artistic creation.</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

Zero-shot text-driven image style transfer: a cross-modal semantic alignment approach

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Lili Duan,
  • Hang Chen,
  • Jiahao Cui,
  • Xinlei Tang,
  • Jun Liu,
  • Lei Zhang

摘要

Image style transfer, a fundamental task in computer vision, aims to transform the visual style of a content image to match a desired artistic style. Traditional methods often require explicit reference style images, limiting their flexibility. This paper introduces a novel zero-shot, text-driven image style transfer framework that leverages a pre-trained CLIP model to establish semantic alignment between textual descriptions and visual features. Our approach enables direct mapping of stylistic attributes from language to the image domain without the need for style images. Key innovations include a CLIP-guided semantic adapter module, a cross-attention-based style modulation module, and a content-style decoupling module. Extensive experiments demonstrate that our method generates visually coherent results and outperforms existing text-driven and style transfer techniques, achieving a new benchmark in zero-shot artistic creation.