Scholarlink

<p>Image style transfer aims to render an image in a desired style while preserving its original content. Existing methods, particularly those reliant on specific style reference images, often limit their applicability and may compromise stylization quality. To address these limitations, we introduce a novel zero-shot text-guided diffusion image style transfer method named FAGStyle. Our approach leverages Sliding Window Crop (SWC) and Feature Augmentation on Geodesic Surface (FAGS) to enhance inter-patch information interaction, facilitating consistent and uniform stylization. Additionally, we integrate a Pre-Shape self-correlation consistency loss to maintain content consistency. Extensive experimental results demonstrate that FAGStyle outperforms existing methods, achieving superior stylization that retains the semantic content of source images across both imagined and common styles. Our work showcases the efficacy of geodesic feature augmentation in zero-shot text-guided diffusion style transfer. The code and datasets can be accessed at <a href="https://github.com/P2i42/FAGStyle">https://github.com/P2i42/FAGStyle</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

Geodesic feature augmentation for zero-shot text-guided diffusion style transfer

  • 期刊论文

  • 发表日期：

    2025年7月23日

查看全文
  • Yuexing Han,
  • Liheng Ruan,
  • Bing Wang

摘要

Image style transfer aims to render an image in a desired style while preserving its original content. Existing methods, particularly those reliant on specific style reference images, often limit their applicability and may compromise stylization quality. To address these limitations, we introduce a novel zero-shot text-guided diffusion image style transfer method named FAGStyle. Our approach leverages Sliding Window Crop (SWC) and Feature Augmentation on Geodesic Surface (FAGS) to enhance inter-patch information interaction, facilitating consistent and uniform stylization. Additionally, we integrate a Pre-Shape self-correlation consistency loss to maintain content consistency. Extensive experimental results demonstrate that FAGStyle outperforms existing methods, achieving superior stylization that retains the semantic content of source images across both imagined and common styles. Our work showcases the efficacy of geodesic feature augmentation in zero-shot text-guided diffusion style transfer. The code and datasets can be accessed at https://github.com/P2i42/FAGStyle.