DA-GAN : Dual-attention generative adversarial networks for real-world exquisite makeup transfer

Existing makeup transfer models do not in general perform satisfactorily when applied to real-world unseen face images. Additionally, most existing models primarily focus on color distribution and struggle to transfer accurate details, particularly for elaborate artistic makeup styles. Recognizing these limitations, we propose the dual-attention generative adversarial networks, which is referred to as DA-GAN. Our work addresses the challenge of wide-ranging and exquisite makeup transfer in real-world settings. DA-GAN mainly incorporates a global attention module, a local attention module, and a makeup rendering module. The global attention module builds a coarse mapping from the reference makeup image to the source image. Based on the coarse mapping, the local attention module is designed to build a more precise mapping and preserve makeup details. The local attention module divides the entire face into a set of local patches. Based on these patches, the makeup rendering module applies an exquisite makeup style to the source image which benefits from a set of local discriminators. Extensive experiments demonstrate that DA-GAN achieves state-of-the-art results in transferring exquisite details for both ordinary and artistic makeup styles in real-world scenarios. © 2024 Elsevier Ltd