Reconstructing high-fidelity and animatable 3D human avatars from visual data is a core task for immersive applications such as virtual reality (VR) and digital content creation. While traditional approaches often suffer from high computational costs, slow inference, and visual artifacts, recent advances leverage 3D Gaussian Splatting (3DGS) to enable rapid training and real-time rendering (up to 361 FPS). A common framework leverages parametric models to establish a canonical human representation, followed by deformation of 3D Gaussians into target poses using learnable skinning and novel regularization techniques. Key advances include deformation mechanisms for motion generalization, hybrid Gaussian-mesh representations for complex clothing and geometry, efficient compression and acceleration strategies, and specialized modules for handling occlusions and fine details. This article briefly reviews recent progress in 3DGS-based human reconstruction, we organize methods by input type: single-view and multi-view reconstruction. We discuss the strengths and limitations of each category and highlight promising future directions.
Chen et al. (2025) studied this question.