Mitigating Occlusions in Virtual Try-On via A Simple-Yet-Effective Mask-Free Framework
This work tackles occlusion issues in Virtual Try-On (VTON). We taxonomize failures into:
1. Inherent Occlusions – “ghost” garments from the reference image that remain in the result.
2. Acquired Occlusions – distorted human anatomy that visually blocks the new outfit.
To remove both, we propose a mask-free VTON framework with two plug-and-play operations:
- Background Pre-Replacement – swaps the background before generation so the model never confuses clothes with body/background, suppressing inherent occlusions.
- Covering-and-Eliminating – enforces human-aware semantics, yielding anatomically plausible shapes and thus fewer acquired occlusions.
The operations are architecture-agnostic: drop them into GANs or diffusion models without re-design.
@article{du2025mitigating,
title={Mitigating Occlusions in Virtual Try-On via A Simple-Yet-Effective Mask-Free Framework},
author={Du, Chenghu and Xiong, Shengwu and Wang, Junyin and Rong, Yi and Xiong, Shili},
journal={Advances in Neural Information Processing Systems},
year={2025}
}