Rig3DGS is a novel method for creating controllable 3D human portraits from casual smartphone videos, enabling high-quality reanimation and novel view synthesis. The method uses a set of 3D Gaussians in a canonical space to represent the dynamic scene, including the subject and background. By applying learned deformations guided by a 3D morphable model, Rig3DGS can accurately model facial expressions, head poses, and viewing directions. The key innovation lies in the learnable prior that restricts the deformation to a subspace spanned by the deformations of the closest vertices on the FLAME mesh, ensuring efficient training and effective control over facial expressions and head poses. Extensive experiments demonstrate the effectiveness of Rig3DGS, showing significant improvements in rendering quality compared to prior methods, with 50 times faster training and inference speeds. The method is evaluated on various datasets and compared against baselines such as RigNeRF, INSTA, and PointAvatar, consistently outperforming them in terms of quality and speed.Rig3DGS is a novel method for creating controllable 3D human portraits from casual smartphone videos, enabling high-quality reanimation and novel view synthesis. The method uses a set of 3D Gaussians in a canonical space to represent the dynamic scene, including the subject and background. By applying learned deformations guided by a 3D morphable model, Rig3DGS can accurately model facial expressions, head poses, and viewing directions. The key innovation lies in the learnable prior that restricts the deformation to a subspace spanned by the deformations of the closest vertices on the FLAME mesh, ensuring efficient training and effective control over facial expressions and head poses. Extensive experiments demonstrate the effectiveness of Rig3DGS, showing significant improvements in rendering quality compared to prior methods, with 50 times faster training and inference speeds. The method is evaluated on various datasets and compared against baselines such as RigNeRF, INSTA, and PointAvatar, consistently outperforming them in terms of quality and speed.