3D Gaussian Splatting(3DGS) has emerged as a powerful framework for real-time neural rendering. However, enabling interactive deformation remains challenging; applying transformations solely to Gaussian centroids without updating their covariance-encoded orientations causes significant visual artifacts, such as surface distortion and volume discontinuity. To address this, we propose a quaternion-based covariance rotation synchronization method. Our approach computes a local rotation from user-defined inputs, converts it into a quaternion, and composes it with each primitive’s original rotation to ensure geometric consistency. Integrated into a CUDA-based pipeline, the method preserves visual coherence under large and non-rigid transformations without significant overhead. Experimental results demonstrate that our strategy effectively eliminates structural artifacts, providing a practical solution for interactive editing in Gaussian-based rendering systems.
Kwon et al. (Sat,) studied this question.