Abstract

Artifacts often degrade the visual quality of virtual try-on (VTON) and posetransfer applications, impacting user experience. This study introduces a novelconditional inpainting technique designed to detect and remove suchdistortions, improving image aesthetics. Our work is the first to present anend-to-end framework addressing this specific issue, and we developed aspecialized dataset of artifacts in VTON and pose transfer tasks, complete withmasks highlighting the affected areas. Experimental results show that ourmethod not only effectively removes artifacts but also significantly enhancesthe visual quality of the final images, setting a new benchmark in computervision and image processing.

 

Quick Read (beta)