SCRN: Stepwise Change and Refine Network Based Semantic Distribution for Human Pose Transfer

It is a challenging and meaningful task to achieve person image synthesis by guiding pose. However, two problems have existed in past work: inaccurate generated poses inconsistency with the target texture. To address these issues, we propose Stepwise Change Refine Network (SCRN), two-stage network that aims transfer given images pose while generating more reasonable closer-to-real results. In first stage, coarse are using series of modules same structure called Coarse Blocks. This process gradually changes better shape consistency image. second style features extracted from original distributing semantic information. These used optimize rough obtain final image, resulting appearance Our proposed method preserves both pose’s spatial image’s texture features. Furthermore, introduce new loss function make line human perception. Qualitative quantitative experiments state-of-the-art models demonstrate significant improvements SSIM, FID, PSNR, LPIPS, validating superiority our model.