DRAN: Detailed Region-Adaptive Normalization for Conditional Image Synthesis

In recent years, conditional image synthesis has attracted growing attention due to its controllability in the image generation process. Although recent works have achieved realistic results, most of them fail to handle fine-grained styles with subtle details. To address this problem, a novel normalization module, named DRAN, is proposed. It learns fine-grained style representation, while maintaining the robustness to general styles. Specifically, we first introduce a multi-level structure, Spatiality-Aware Pyramid Pooling, to guide the model to learn coarse-to-fine features. Then, to adaptively fuse different levels of styles, we propose Dynamic Gating, making it possible to choose different styles according to different spatial regions. To evaluate the effectiveness and generalization ability of DRAN, we conduct a set of experiments on makeup transfer and semantic image synthesis. Quantitative and qualitative experiments show that equipped with DRAN, the baseline models are able to achieve significant improvement in complex style transfer and texture details reconstruction.