Scholarlink

<p>With the advancement of neural networks, various elaborately designed models have achieved great success in the field of style transfer. Previous studies have focused on aligning the statistical information of content and style images. Although they have addressed the disentanglement of content and style in images, existing disentanglement methods still suffer from poor separation performance and difficulty in balancing multi-scale feature fusion. To further optimize disentanglement performance and improve stylization quality, this paper designs a novel disentanglement module and fusion module. The disentanglement module disentangles the content and style components of an image. On this basis, the fusion module integrates multi-scale content information from the content image and multi-scale style information from the style image. The fused features are fed into a decoder to generate stylized images. The disentanglement capability of the disentanglement module is enhanced by introducing a disentanglement loss. Extensive experiments demonstrate the effectiveness of the proposed method. Our code used in the paper is available at <a href="https://github.com/cjhjbj/DAFNet">https://github.com/cjhjbj/DAFNet</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

Dafnet: disentanglement-and-fusion network for arbitrary image style transfer

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Jiaheng Chen,
  • Jihong Liu,
  • Hongxia Jiang,
  • Wenwen Wang

摘要

With the advancement of neural networks, various elaborately designed models have achieved great success in the field of style transfer. Previous studies have focused on aligning the statistical information of content and style images. Although they have addressed the disentanglement of content and style in images, existing disentanglement methods still suffer from poor separation performance and difficulty in balancing multi-scale feature fusion. To further optimize disentanglement performance and improve stylization quality, this paper designs a novel disentanglement module and fusion module. The disentanglement module disentangles the content and style components of an image. On this basis, the fusion module integrates multi-scale content information from the content image and multi-scale style information from the style image. The fused features are fed into a decoder to generate stylized images. The disentanglement capability of the disentanglement module is enhanced by introducing a disentanglement loss. Extensive experiments demonstrate the effectiveness of the proposed method. Our code used in the paper is available at https://github.com/cjhjbj/DAFNet.