VITON-HD: High-Resolution Virtual Try-On via Misalignment-Aware Normalization

The task of image-based virtual try-on aims to transfer a target clothing\nitem onto the corresponding region of a person, which is commonly tackled by\nfitting the item to the desired body part and fusing the warped item with the\nperson. While an increasing number of studies have been conducted, the\nresolution of synthesized images is still limited to low (e.g., 256x192), which\nacts as the critical limitation against satisfying online consumers. We argue\nthat the limitation stems from several challenges: as the resolution increases,\nthe artifacts in the misaligned areas between the warped clothes and the\ndesired clothing regions become noticeable in the final results; the\narchitectures used in existing methods have low performance in generating\nhigh-quality body parts and maintaining the texture sharpness of the clothes.\nTo address the challenges, we propose a novel virtual try-on method called\nVITON-HD that successfully synthesizes 1024x768 virtual try-on images.\nSpecifically, we first prepare the segmentation map to guide our virtual try-on\nsynthesis, and then roughly fit the target clothing item to a given person's\nbody. Next, we propose ALIgnment-Aware Segment (ALIAS) normalization and ALIAS\ngenerator to handle the misaligned areas and preserve the details of 1024x768\ninputs. Through rigorous comparison with existing methods, we demonstrate that\nVITON-HD highly surpasses the baselines in terms of synthesized image quality\nboth qualitatively and quantitatively. Code is available at\nhttps://github.com/shadow2496/VITON-HD.\n