Context-Preserved Spatial Normalization Based Person Image Generation

Pose guided person image generation (PGPIG) stands as a prominent task in the field of image generation, focusing on generating a target person image with a given pose derived from a reference image. However, the existing methods tend to concentrate on keypoints correlation rather than texture correlation. To address these challenges, we introduce a two-stage model to concurrently establish the strong correlation between source domain and target domain, so as to generate person images from coarse to fine. Firstly, we propose a Multi-Scale Attention Fusion Alignment Network which is dedicated to generating coarse person images with the target poses while preserving color and edge information of the source images. Secondly, our Textural Local Correlation Transfer Network gradually transfers the texture features from the source images to the target poses. Finally, in order to capture more intricate details in the generated person images and achieve spatial alignment of features, we propose a Context-Preserved Spatial Normalization method. Experimental results demonstrate that the person images generated by our model exhibit realistic textures and are basically consistent with the poses of the target images. Our approach yields superior results from both quantitative and qualitative perspectives.