Abstract: To address the limitations of current virtual try-on methods in adequately preserving local details of target garments, and when diffusion model were used for generation, the Variational AutoEncoder (VAE)’s mapping of input data to low-dimensional space leaded to loss of high-frequency details in models’ hands and faces, a human-centric detail-enhanced virtual try-on method was proposed. Firstly, the clothing-agnostic person map, pose map, and target garment were input into a geometric matching module to generate a coarsely warped garment. Then, a Garment Wrap Refinement (GWR) was constructed to enhance the detailed features of the coarsely warped garment. Subsequently, the warped garment, clothing-agnostic person map, and pose map were concatenated and fed into a UNet alongside textual features, where textual and visual features were fused to progressively generate a clear image through iterative denoising. A Mask Feature Connection (MFC) moduleincorporating a coordinate attention was introduced to precisely localize the model’s position and preserve high-frequency details in regions such as hands and faces, ensuring human-centric results. Finally, the outputs of the (MFC) module and the UNet were fused and decoded to produce the final try-on result. Experimental results demonstrate that the proposed method achieves a 1.41% improvement in the Structural Similarity Index Measure (SSIM) on the DressCode dataset, along with reductions of 7.32%, 31%, and 65% in the Learned Perceptual Image Patch Similarity (LPIPS), Fréchet Inception Distance (FID), and Kernel Inception Distance (KID) metrics, respectively, compared to the LADI-VTON method, these results indicate that the proposed method achieves superior performance in virtual try-on tasks.

Key words: virtual try-on, detail enhancement, coordinate attention, human centric, diffusion model

摘要： 针对当前虚拟试衣方法无法充分保留目标服装的局部细节问题，以及使用扩散模型生成试衣结果时，变分自编码器(VAE)会将输入数据映射到低维空间，从而导致模特手部和脸部高频细节特征丢失的问题，提出一种以人为中心的细节增强虚拟试衣方法。首先，将服装不可知的人体图、人体姿态图和目标服装输入几何匹配模块得到粗扭曲服装结果，再通过构建服装扭曲细化(GWR)模块，增强粗扭曲服装的细节特征；其次，将服装扭曲图、服装不可知的人体图以及人体姿态图等拼接后和文本特征作为UNet的输入，将文本特征与图像特征融合，通过去噪逐步生成清晰的图像；再次，构建掩码特征连接(MFC)模块，引入坐标注意力机制，更准确地定位模特的位置信息，保留模特的手部和脸部细节特征，实现以人为中心；最后，将MFC模块的输出与UNet的输出进行融合解码，得到最终的试衣结果。实验结果表明，与LADI-VTON方法相比，所提方法在Dress Code数据集上的结构相似度指数（SSIM）指标提升了1.41%，在感知相似度（LPIPS）、FID（Fréchet Inception Distance）、KID（Kernel Inception Distance）评价指标上分别降低了7.32%、31%、65%，以上结果验证了所提方法虚拟试衣效果更优。

关键词: 虚拟试衣, 细节增强, 坐标注意力机制, 以人为中心, 扩散模型