Abstract

Achieving realistic animated human avatars requires accurate modeling of pose-dependent clothing deformations. Existing learning-based methods heavily rely on the Linear Blend Skinning (LBS) of minimally-clothed human models like SMPL to model deformation. However, these methods struggle to handle challenging loose clothing, such as long dresses, because the required canonicalization process is ill-defined where clothing is away from the body, leading to disjointed and fragmented results. To overcome this limitation, we propose a novel hybrid framework to model challenging clothed humans. The core idea is to use two separate strategies to model clothing regions close to the body and distant from the body. Specifically, we leverage LBS to model close regions that are heavily affected by body movements. More importantly, we further introduce a novel free-form generation to model more distant clothing regions that are less affected by body movements. This free-form generation module, conditioned on the underlying body, endows enhanced flexibility and expressive capabilities when modeling challenging loose clothing, such as skirts and dresses. Combining the strengths of LBS-based deformation and free-form generation, our hybrid framework effectively captures the intricate geometric details of loose clothing. Experimental results on the benchmark dataset featuring loose clothing demonstrate that our method achieves superior visual fidelity and realism, particularly on the most challenging cases.