ManipulateGaussian: 3D Gaussian Manipulation from Observed Vibrations

Understanding and manipulating the vibration modes of 3D objects is critical for advancements in computer graphics, robotics, and physics-based simulation. A fundamental challenge in precise simulation is constructing interpretable modal representations that facilitate material property estimation, structural analysis, and physically grounded animation. Inspired by recent progress in learning-based dynamic reconstruction, this work extends vibration recovery from 2D dynamic observations into 3D particle simulation. We present a framework for learning and visualizing physically plausible vibration behaviors directly from video data using off-the-shelf 3D Gaussian Splatting (3DGS) models. By leveraging optical flow, deformation modeling, and differentiable rendering, our system estimates modal bases that accurately describe the object's intrinsic structural dynamics.