UniVVT: Unified End-to-End Framework for Video Virtual Try-On

A recent study presents UniVVT, an integrated framework designed for high-fidelity video virtual try-on (VVT). This innovative approach redefines VVT as semantically conditioned video generation, removing the necessity for distinct modules for human parsing, pose estimation, and garment warping during inference. It overcomes the challenges posed by prevalent methods that depend on mask-conditioned video inpainting and multi-stage processes, which can complicate deployment and allow for the propagation of errors in geometric priors into the final videos. UniVVT employs a scene-task perceiver based on a Multimodal Large Language Model to encode the source video, target garment, and task instructions into compact, task-aware latent tokens. A lightweight semantic bridge connects these tokens to the video generation process. The paper is accessible on arXiv under identifier 2608.05745 and is noted as a cross-type abstract. This research holds significance for the digital art and AI technology fields, enhancing the ability to create realistic videos of individuals in various garments, with applications in fashion, e-commerce, and digital content creation.