Sun Yat-sen's OMFA Handles Virtual Try-On and Try-Off Without Masks

Most virtual try-on systems have a dirty secret: they only work when you feed them a pristine product photo of the garment on a white background. Try to lift a shirt off a person in a casual snapshot and stick it on someone else, and the results fall apart. A new paper from Sun Yat-sen University and X-Era AI Lab proposes OMFA (One Model For All), a unified diffusion framework that handles try-on, try-off, and pose transfer inside a single network, without any segmentation masks at inference time.

The paper argues that the field has been treating three closely related problems as separate pipelines when they are really the same problem viewed from different angles. OMFA models garment-person transformations bidirectionally within a single architecture, removing the need for garment templates or segmentation masks.

Why current try-on models are stuck

The dominant recipe for diffusion-based virtual try-on relies on two crutches: clean garment templates (sometimes called exhibition garments) and segmentation masks to isolate clothing regions. That works in an e-commerce catalog, but not when a user snaps a picture of a friend's jacket and wants to see themselves wearing it.

Workarounds exist but come with pain. Some methods extract garments directly from images using segmentation algorithms, but these techniques often suffer from boundary artifacts, occlusion issues, and shape distortion. Others reformulate the problem as virtual try-off, reconstructing a canonical garment image from a dressed person, though such approaches treat try-on and try-off as independent tasks and overlook their inherent duality. Pose is another sore spot: 2D try-on models generally lock output to the input pose, and 3D approaches struggle because high-quality 3D garment data is scarce.