Scholarlink

<p>Modern diffusion-based style transfer relies on cross-attention to fuse content structures with reference styles. However, conventional cross-attention often entangles structural cues with textural patterns from the style image, leading to shape leakage and distortion of the source content. To address this issue, we propose a Content-Orthogonal Projection mechanism. This mechanism explicitly removes the subspace of style values aligned with content representations by projecting style features onto the orthogonal complement of the content subspace. This design ensures that cross-attention conveys only texture-, color-, and brush-related statistics while strongly suppressing structural artifacts. We integrate this method into the reweighted cross-attention framework, enabling zero-shot style transfer. Extensive experiments demonstrate that our method faithfully preserves content geometry while significantly enhancing stylistic fidelity, achieving superior disentanglement between style and structure. This work provides an efficient and principled pathway toward controllable, texture-oriented style transfer in diffusion models.</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

Content-orthogonal projection for disentangled style transfer in diffusion models

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Hang Chen,
  • Jiahao Cui,
  • Xinlei Tang,
  • Jinyun Zeng,
  • Yanzhe Zou,
  • Xuan Li,
  • Lei Zhang

摘要

Modern diffusion-based style transfer relies on cross-attention to fuse content structures with reference styles. However, conventional cross-attention often entangles structural cues with textural patterns from the style image, leading to shape leakage and distortion of the source content. To address this issue, we propose a Content-Orthogonal Projection mechanism. This mechanism explicitly removes the subspace of style values aligned with content representations by projecting style features onto the orthogonal complement of the content subspace. This design ensures that cross-attention conveys only texture-, color-, and brush-related statistics while strongly suppressing structural artifacts. We integrate this method into the reweighted cross-attention framework, enabling zero-shot style transfer. Extensive experiments demonstrate that our method faithfully preserves content geometry while significantly enhancing stylistic fidelity, achieving superior disentanglement between style and structure. This work provides an efficient and principled pathway toward controllable, texture-oriented style transfer in diffusion models.