Scholarlink

<p>Image-based virtual try-on technology aims to realistically superimpose garments onto target individuals. A key challenge lies in accurately aligning garment images with the person’s shape, traditionally addressed through warping techniques. Existing methods often struggle with complex deformations and preserving intricate patterns. To overcome these limitations, we introduce LID-VTON, a novel virtual try-on model comprising a Local Flow Iteratively Refined (LFIR) warping module and a Dense Hybrid Generator Module (DHGM). The LFIR module incorporates a Context Aware Flow Estimator (CAFE) for precise local flow estimation and an Iteratively Refined Attention Module (IRAM) for refining attention maps, significantly improving deformation accuracy, while the DHGM integrates DenseNet-style connections and SE attention mechanisms to improve feature representation and detail preservation. Experimental results on the VITON-HD dataset demonstrate that LID-VTON outperforms comparable methods, generating more realistic and precise try-on images.</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

LID-VTON: Enhancing virtual try-on precision through local flow iterative refinement and dense hybrid generation

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Yueqi Liu,
  • Yang Xu,
  • Jiao Jiang,
  • Zhirong Cheng,
  • Fengyun Zuo

摘要

Image-based virtual try-on technology aims to realistically superimpose garments onto target individuals. A key challenge lies in accurately aligning garment images with the person’s shape, traditionally addressed through warping techniques. Existing methods often struggle with complex deformations and preserving intricate patterns. To overcome these limitations, we introduce LID-VTON, a novel virtual try-on model comprising a Local Flow Iteratively Refined (LFIR) warping module and a Dense Hybrid Generator Module (DHGM). The LFIR module incorporates a Context Aware Flow Estimator (CAFE) for precise local flow estimation and an Iteratively Refined Attention Module (IRAM) for refining attention maps, significantly improving deformation accuracy, while the DHGM integrates DenseNet-style connections and SE attention mechanisms to improve feature representation and detail preservation. Experimental results on the VITON-HD dataset demonstrate that LID-VTON outperforms comparable methods, generating more realistic and precise try-on images.