Scholarlink

<p>Makeup transfer is a computer vision technique aimed at enhancing facial beauty by transferring makeup styles from reference images to source images. However, existing methods often struggle with varying lighting conditions, leading to imprecise and unnatural results. In this paper, we propose Lighting Transfer Generative Adversarial Networks (LTGAN) to address this challenge. LTGAN incorporates a Lighting Transfer Module (LTM) that transfers the lighting conditions of the source image to the reference image, reducing the impact of irrelevant lighting information on the makeup transfer process. Furthermore, we introduce a Detail Refining Encoder (DREnc) to capture both global and local, as well as multi-level features, by utilizing a CLIP (Contrastive Language-Image Pre-Training) image encoder and further process the extracted features from CLIP by HiLo (High Low) attention module. A hybrid loss function, including local loss based on facial masks, is employed to achieve fine local makeup transfer while maintaining the identity of the source image. Experimental results demonstrate that our proposed LTGAN model, equipped with a ConvNeXt V2 block for feature extraction, achieves more natural and state-of-the-art makeup transfer under extreme lighting conditions compared to existing methods. By considering the neck as part of the makeup area, our model further enhances the overall naturalness of the transfer effect. We made the code publicly available at <a href="https://github.com/yeyu-cpu/LTGAN">https://github.com/yeyu-cpu/LTGAN</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

Enhancing makeup transfer robustness under varied lighting conditions with lighting transfer GAN

  • 期刊论文

  • 发表日期：

    2026年2月27日

查看全文
  • Yifei Song,
  • Yongfang Wang

摘要

Makeup transfer is a computer vision technique aimed at enhancing facial beauty by transferring makeup styles from reference images to source images. However, existing methods often struggle with varying lighting conditions, leading to imprecise and unnatural results. In this paper, we propose Lighting Transfer Generative Adversarial Networks (LTGAN) to address this challenge. LTGAN incorporates a Lighting Transfer Module (LTM) that transfers the lighting conditions of the source image to the reference image, reducing the impact of irrelevant lighting information on the makeup transfer process. Furthermore, we introduce a Detail Refining Encoder (DREnc) to capture both global and local, as well as multi-level features, by utilizing a CLIP (Contrastive Language-Image Pre-Training) image encoder and further process the extracted features from CLIP by HiLo (High Low) attention module. A hybrid loss function, including local loss based on facial masks, is employed to achieve fine local makeup transfer while maintaining the identity of the source image. Experimental results demonstrate that our proposed LTGAN model, equipped with a ConvNeXt V2 block for feature extraction, achieves more natural and state-of-the-art makeup transfer under extreme lighting conditions compared to existing methods. By considering the neck as part of the makeup area, our model further enhances the overall naturalness of the transfer effect. We made the code publicly available at https://github.com/yeyu-cpu/LTGAN.