MU-GAN: Facial Attribute Editing Based on Multi-Attention Mechanism

Facial attribute editing has mainly two objectives: 1) translating image from a source domain to target one, and 2) only changing the facial regions related preserving attribute-excluding details. In this work, we propose multi-attention U-Net-based generative adversarial network (MU-GAN). First, replace classic convolutional encoder-decoder with symmetric U-Net-like structure in generator, then apply an additive attention mechanism build attention-based U-Net connections for adaptively transferring encoder representations complement decoder detail enhance ability. Second, self-attention (SA) is incorporated into layers modeling long-range multi-level dependencies across regions. Experimental results indicate that our method capable of balancing ability details preservation ability, can decouple correlation among attributes. It outperforms state-of-the-art methods terms manipulation accuracy quality. Our code available at https://github.com/SuSir1996/MU-GAN.