Scholarlink

<p>Virtual try-on (VTON) technology aims to generate realistic images of a target person wearing a specified garment, but existing methods still struggle to preserve fine garment details and achieve accurate garment-body alignment, especially with complex clothing styles and diverse poses. To address these issues, this paper proposes HAR-VTON, a novel diffusion-based VTON framework with core innovations: a hybrid attention mechanism and a receptive field block (RFB). The hybrid attention mechanism fuses a parameter-frozen convolutional block attention module (CBAM) with a trainable cross-attention module to enhance local–global feature interaction, while the RFB enables multiscale feature extraction for better modeling of natural garment fitting on the human body. Additionally, we construct the Cultural-VTON dataset, containing traditional garments from 56 ethnic groups, to evaluate generalization under complex clothing conditions. Experiments on the VITON-HD dataset show HAR-VTON achieves competitive performance (LPIPS: 0.095, SSIM: 0.892, FID: 7.35), outperforming baselines like VITON-HD, HR-VTON, IDM-VTON, and OOTDiffusion in detail preservation and alignment. It also performs well on Cultural-VTON, supporting digital inheritance of ethnic garments. The code is available at <a href="https://github.com/linlinx666/HARVTON.git">https://github.com/linlinx666/HARVTON.git</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

Har-vton: a diffusion-based virtual try-on framework with hybrid attention and receptive field modules

  • 期刊论文

  • 发表日期：

    2025年12月14日

查看全文
  • Yulin Xiong,
  • Yuxin Hong,
  • Xuyan Huang,
  • Jianlin Zhu,
  • Zimao Li,
  • Ruhan He,
  • Xiao Zhang,
  • Bo Yang,
  • Meng Shi

摘要

Virtual try-on (VTON) technology aims to generate realistic images of a target person wearing a specified garment, but existing methods still struggle to preserve fine garment details and achieve accurate garment-body alignment, especially with complex clothing styles and diverse poses. To address these issues, this paper proposes HAR-VTON, a novel diffusion-based VTON framework with core innovations: a hybrid attention mechanism and a receptive field block (RFB). The hybrid attention mechanism fuses a parameter-frozen convolutional block attention module (CBAM) with a trainable cross-attention module to enhance local–global feature interaction, while the RFB enables multiscale feature extraction for better modeling of natural garment fitting on the human body. Additionally, we construct the Cultural-VTON dataset, containing traditional garments from 56 ethnic groups, to evaluate generalization under complex clothing conditions. Experiments on the VITON-HD dataset show HAR-VTON achieves competitive performance (LPIPS: 0.095, SSIM: 0.892, FID: 7.35), outperforming baselines like VITON-HD, HR-VTON, IDM-VTON, and OOTDiffusion in detail preservation and alignment. It also performs well on Cultural-VTON, supporting digital inheritance of ethnic garments. The code is available at https://github.com/linlinx666/HARVTON.git