Towards diverse image-to-image translation via adaptive normalization layer and contrast learning

A nice image-to-image translation framework is able to acquire an explicit and credible mapping relationship between the source domain and target domains while satisfying two requirements. One is simplicity, the other is extensibility over multiple translation tasks. To this end, we design a concise but versatile generative model for image-to-image translation. Our method includes three major ingredients. First, inspired by popular unconditional normalization layers, named Spatially Adaptive Normalization(SPADE). We introduce a novel Semantics-Appearance Spatially Adaptive Normalization (SA-SPADE), taking into account both semantic structure and style appearance. This enables semantic composition and style appearance information to be sufficiently captured and integrated by our normalization layers. Thanks to SA-SPADE, our model extends to multiple image-to-image translation tasks in an unsupervised or supervised way. Second, we carefully designed two symmetrical network branches to provide semantic and appearance information for our normalization layer, namely Semantic Branch (SB) and Appearance Branch(AB) respectively. Third, we propose novel Semantic-aware Contrastive Loss (SCL) and Appearance-aware Contrastive Loss (ACL)based on newly un-/self- supervised contrastive learning. That is, SCL guarantees domain-invariant (e.g., pose, structure) representations between the generated image and the input image, while ACL ensures domain-specific representations (e.g., color, texture) between the generated image and the reference image. As a result, we verify the effectiveness of our method by comparing it with various task-dependent image translation models in both qualitative and quantitative evaluations.