Learning Implicit Features with Flow-Infused Transformations for Realistic Virtual Try-On

Target Person Garment Warped Garment GP-VTON StableVITON D4-VTON Ours Figure 1. Warping-based method (e.g. GP-VTON) is prone to severe visual artifacts and distortions (see the green dashed box) affected by the incorrect prediction of warp module. Learning-based methods (e.g. StableVTON and D 4 -VTON) are difficult to reconstruct complex detailed textures (see the red dashed box). Our FIA-VTON designs a Flow Infused Attention module, utilizing the warping flow as an implicit guidance to reconstruct complex detailed textures while maintaining the consistency of the garment.