Scholarlink

<p>Artificial intelligence in fashion design aims to generate detailed fashion images from rough sketches and textures through deep learning. Traditional approaches concatenate sketch and texture in pixel space, struggling with minimizing semantic discrepancies and achieving accurate texture filling. To address these issues, this study introduces BF-Fashion, a fashion design model based on latent diffusion models (LDMs) with bidirectional feature modulation fusion (bFMFusion) from sketch and texture. The bFMFusion module facilitates bi-directional feature exchanges between sketch and texture, reducing semantic differences and enabling seamless integration across layers of the LDMs generator. Furthermore, a classifier-free guidance strategy enhances the fusion of texture and sketch during fashion generation. Experimental results demonstrate that BF-Fashion can generate high-quality fashion items by leveraging various design elements, achieving state-of-the-art performance in fidelity, similarity, and color difference. The proposed model offers an efficient and fast tool for personalized fashion design, assisting users in their creative process. The source code is available at: <a href="https://github.com/zibingo/BF-Fashion">https://github.com/zibingo/BF-Fashion</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

Bidirectional feature modulation fusion for fashion design using latent diffusion models

  • 期刊论文

  • 发表日期：

    2025年4月13日

查看全文
  • Zibin Lu,
  • Jianhua Guo,
  • Shaopeng Liu,
  • Zhixiang Yin

摘要

Artificial intelligence in fashion design aims to generate detailed fashion images from rough sketches and textures through deep learning. Traditional approaches concatenate sketch and texture in pixel space, struggling with minimizing semantic discrepancies and achieving accurate texture filling. To address these issues, this study introduces BF-Fashion, a fashion design model based on latent diffusion models (LDMs) with bidirectional feature modulation fusion (bFMFusion) from sketch and texture. The bFMFusion module facilitates bi-directional feature exchanges between sketch and texture, reducing semantic differences and enabling seamless integration across layers of the LDMs generator. Furthermore, a classifier-free guidance strategy enhances the fusion of texture and sketch during fashion generation. Experimental results demonstrate that BF-Fashion can generate high-quality fashion items by leveraging various design elements, achieving state-of-the-art performance in fidelity, similarity, and color difference. The proposed model offers an efficient and fast tool for personalized fashion design, assisting users in their creative process. The source code is available at: https://github.com/zibingo/BF-Fashion.