Gaussian Motion Field for High-Performance Video Compression

Weichuang Li, Yuxuan Liang

Keywords: Video Compression, Representation Learning

Abstract: Neural video representations have advanced video compression technologies, yet many remain decoding-heavy and struggle to model high-frequency motion. To this end, we introduce Gaussian Motion Field (GMF), a 2D Gaussian–Splatting video codec that represents each frame with a compact set of Gaussians updated by a learned motion field. By predicting per-Gaussian deformations for temporal interpolation, GMF reduces temporal redundancy and requires substantially less capacity than traditional methods that rely on keyframe compression and complex motion estimation. In contrast to NeRV-style models with deep convolutional upsampling, GMF integrates shallow MLPs with lightweight Gaussian representations for efficient decoding. This design yields high storage efficiency and extremely fast decoding: over 1,000 FPS on a single GPU, amounting to roughly a 50$\times$ speedup over recent methods such as HiNeRV, while maintaining comparable visual quality.

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

Submission Number: 23024