The proposed framework, LoopGaussian, aims to create authentic 3D cinemagraphs from multi-view images of static scenes. The key innovation lies in reconstructing the 3D structure of the scene using 3D Gaussian Splatting. This approach allows for the generation of cinemagraphs with dynamic scenes while offering users control over the magnitude of the scene dynamics. Unlike existing methods that operate in 2D image space, LoopGaussian leverages 3D reconstruction to provide a more realistic and immersive experience for viewers.

The methodology involves several key components, including the use of multi-view images to reconstruct the 3D scene structure. A novel Eulerian motion field is employed to capture the dynamic elements of the scene, enabling the creation of cinemagraphs with realistic motion effects. The process involves clustering 3D Gaussian points, projecting them into a feature space using an autoencoder based on PointNet architecture, and utilizing an MLP for depicting the Eulerian motion field. Additionally, amplitude control of motion and final video rendering parameters are set to enhance the visual quality of the cinemagraphs.