High fidelity virtual try-on network via semantic adaptation and distributed componentization

Abstract Image-based virtual try-on systems have significant commercial value in online garment shopping. However, prior methods fail to appropriately handle details, so are defective maintaining the original appearance of organizational items including arms, neck, and in-shop garments. We propose a novel high fidelity network generate realistic results. Specifically, distributed pipeline is used for simultaneous generation items. First, warped using thin plate splines (TPS) give coarse shape reference, then corresponding target semantic map generated, which can adaptively respond distribution different triggered by Second, componentized separately our map-based image adjustment (SMIAN) avoid interference between body parts. Finally, all components integrated overall result SMIAN. A priori dual-modal information incorporated tail layers SMIAN improve convergence rate network. Experiments demonstrate that proposed method retain better details condition than current methods. Our achieves convincing quantitative qualitative results on existing benchmark datasets.