Attention guided domain alignment for conditional face image generation

Zonglin Li, Shengping Zhang, Zhaoxin Zhang, Quanling Meng, Qinglin Liu, Huiyu Zhou

Abstract: Highlights•We propose an attention guided domain alignment method for conditional face image generation, which exploits the spatial relationship of the facial parts under the guidance of local attention to directly align two distinct domains.•We design a top-k ranking module with a dedicated index system to retrieve semantically feature blocks to alleviate the mismatching and preserve the texture structure. To make the top-k ranking operation differentiable, we formulate it as a regularized optimal transport problem•Upon the retrieved feature blocks, we propose an attention-guided domain alignment (DAM) module that achieves dense correspondences in a high resolution. We design an adaptive feature fusion (ADFF) module that provides reliable feature guidance in face image generation•Extensive experimental results on the CelebAMask-HQ dataset demonstrate that our method is superior to state-of-the-art methods. Due to the significantly reduced computation complexity, our method can perform alignment in a high resolution while achieving faithful style control.