Animating the Still: Physics-Based 3D Cinemagraph from Multi-View Images Using 3D Gaussian Splatting

Lechao Cheng, Jiyang Li, beihuang, Shengeng Tang, Zhangye Wang, Zhihui Yang, Jingxuan He

Keywords: Cinemagraph

Abstract: 3D Cinemagraphs aim to generate visually compelling media by introducing subtle and continuous motion into otherwise static images. Recent efforts have explored this task through 3D reconstruction techniques, but they often fall short in delivering physically plausible and controllable animations. In this paper, we propose a novel physics-driven framework built upon 3D Gaussian Splatting (3DGS) to address these limitations. Given multi-view images and a user-specified force, our approach first reconstructs a 3D scene using 3DGS, then embeds the reconstruction into a physically consistent simulation environment. By modeling external and internal force fields and performing accurate force analysis within the reconstructed 3D space, we synthesize fine-grained and interpretable motion that aligns with physical intuition. Our method allows users to intuitively control motion effects via high-level physical parameters, achieving a delicate balance between realism and artistic flexibility. Extensive experiments under diverse force conditions demonstrate that our approach produces stable, interpretable, and visually appealing results, surpassing prior methods in both robustness and controllability.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6483