Scholarlink

<p>Virtual try-on aims to generate photorealistic images of a person wearing a target garment, requiring precise garment–body alignment, fine-grained texture preservation, and robustness to pose variations. We propose AFDiT, a structure-aware virtual try-on framework that integrates appearance flow estimation with transformer-based diffusion generation. Specifically, we design an appearance flow encoder–decoder to predict dense garment deformation and alignment, and a parsing-guided mask fusion strategy to refine inpainting regions while preserving uncovered body parts. For high-quality synthesis, we introduce a warped garment-guided stable diffusion pipeline, injecting garment-specific CLIP embeddings and flow-guided warped garment features into a transformer-based diffusion model. Extensive experiments on VITON-HD demonstrate that AFDiT outperforms state-of-the-art methods in both paired and unpaired settings, achieving superior alignment, realism, and generalization to diverse poses and garments. The code is available at: <a href="https://github.com/MTKSHU/AFDiT">https://github.com/MTKSHU/AFDiT</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

AFDiT: flow-guided transformer diffusion for structure-aware virtual try-on

  • 期刊论文

  • 发表日期：

    2025年6月29日

查看全文
  • Huiming Ding,
  • Beining Wu,
  • Mengtian Li,
  • Zhifeng Xie

摘要

Virtual try-on aims to generate photorealistic images of a person wearing a target garment, requiring precise garment–body alignment, fine-grained texture preservation, and robustness to pose variations. We propose AFDiT, a structure-aware virtual try-on framework that integrates appearance flow estimation with transformer-based diffusion generation. Specifically, we design an appearance flow encoder–decoder to predict dense garment deformation and alignment, and a parsing-guided mask fusion strategy to refine inpainting regions while preserving uncovered body parts. For high-quality synthesis, we introduce a warped garment-guided stable diffusion pipeline, injecting garment-specific CLIP embeddings and flow-guided warped garment features into a transformer-based diffusion model. Extensive experiments on VITON-HD demonstrate that AFDiT outperforms state-of-the-art methods in both paired and unpaired settings, achieving superior alignment, realism, and generalization to diverse poses and garments. The code is available at: https://github.com/MTKSHU/AFDiT.