Efficient and Robust Video Virtual Try-On Via Enhanced Multi-Garment Alignment

Video virtual try-on aims to generate realistic sequences where garments maintain their identity and adapt accurately to a person's pose and body shape in source video. This task can be regarded as video inpainting, whereas previous methods focus primarily on the specific try-on region while simply “copying” the remaining parts of the person. However, this approach limits the degrees of freedom and heavily relies on precise human parsing. In complex in-the-wild scenarios, dynamic blurring and limb occlusions can introduce errors and discontinuities in the inpainting regions, adversely affecting the video try-on results. Our solution, VidClothEditor, adopts a relaxed editing approach that allows for full-body inpainting and treats non-edited regions as a reconstruction task. It utilizes multiple garment alignment with a proposed region guidance to enhance the naturalness of video try-on results. Additionally, we employ garment-augmented video consistency learning, which significantly reduces the inference time and increases the practical potential for video editing. Comprehensive experiments on the VITON-HD and TikTok datasets confirm VidClothEditor's ability to generate high-quality images and smooth videos. The project website is at video-tryon.github.io.