We have introduced FlexTry-On, a unified and time-efficient framework designed to handle multiple virtual try-on scenarios within a single model. This work, “FlexTry-On: A Time-Efficient Framework for Multi-Scenario Virtual Try-on,” addresses a key limitation of existing image-based virtual try-on (IVTON) systems, which are typically designed for a specific setting such as cloth-to-person (C2P) or person-to-person (P2P) transfer and often require different conditioning or masking strategies when the input format changes. FlexTry-On instead supports C2P, P2P, cross-domain, and in-the-wild try-on through a shared conditioning and generation pipeline. The framework consists of a Text-Aware Segmentation Module (TASM), which converts heterogeneous garment inputs into task-aware garment and editable-region conditions, and a Dual-DiT Tryon Module (DDTM), which separates garment feature extraction from iterative image synthesis so that garment features are computed only once and reused throughout denoising. We further introduce a Style-GRAM loss to preserve garment appearance under noisy or weakly aligned supervision.
Extensive experiments show that FlexTry-On achieves strong quantitative and qualitative performance while requiring only 1.2 seconds per image, including the best paired/unpaired FID among the evaluated methods on VITON-HD and DressCode-Upper, while the full FlexTry-On* model additionally supports P2P and cross-domain C2P/P2P within the same framework. Matched-data experiments further show that FlexTry-On* substantially improves P2P performance over retrained StableVITON, CAT-DM, and ITA-MDT, while qualitative experiments demonstrate its ability to generalize to challenging category-changing and in-the-wild scenarios. Overall, FlexTry-On provides a practical step toward a more general-purpose virtual try-on system that combines scenario flexibility, garment fidelity, and efficient inference within a single unified architecture.
Code, data, and an interactive demo are available at: Coming Very Soon...