OmniTryOn: Video Try-On Anything at Once!

A new artificial intelligence framework can transfer multiple wearable items onto a person in a video in a single inference pass, eliminating the need for the external masks that have constrained prior systems, according to a paper submitted to arXiv in June 2026 [1]. The system, called OmniTryOn, is described by its authors as an “external-prior-free generative framework” designed for what they term the “Try-On Anything” task [1][2]. Existing video virtual try-on methods are limited to single-garment transfer and rely on explicit external priors such as garment masks, a dependency the researchers say “inevitably destroys crucial physical dynamics and degrades visual quality” [2]. OmniTryOn instead employs a First Frame Wearable Cache strategy, which supplies diverse wearable objects for the generation process directly through the initial video frame [1][2]. To maintain consistency across frames, the framework uses Spatiotemporally Consistent RoPE, or STC-RoPE, which establishes spatiotemporal anchors to preserve complex human motions and background dynamics [1][2]. The model is optimized through a Gradual Try-On training strategy, which the paper states allows it to “progressively master robust multi-object synthesis” [2]. The researchers also introduced TryAny-Bench, a benchmark that includes a paired video dataset and an evaluation protocol to standardize the new task [1][2]. Extensive experiments on TryAny-Bench showed that OmniTryOn “significantly outperforms existing specialized video virtual try-on models and general video editing baselines,” according to the paper [1][2]. The dataset, code, and models have been made publicly available on GitHub [2]. The work arrives as generative video editing research accelerates, with multiple related preprints appearing on arXiv in the same period exploring adjacent problems in video synthesis and editing [7][8][9].