Leveraging 3D Gaussian Splatting for Real-Time Facial Expression Transfer

Over the recent years, real-time deepfakes have increasingly been used for digital disinformation and deception. While most live facial manipulations rely on face swapping or single-image animation, each method has distinct drawbacks for realism and identity preservation. This vision paper explores the potential of 3D Gaussian Splatting (3DGS), an explicit radiance field method, for real-time facial reenactment in monocular video streams. By leveraging fast mesh-based tracking and efficient parameter prediction via Multi-Layer Perceptrons, our approach aims at enabling low-latency synthesis of facial expressions and movements. We discuss the current limitations of 3DGS-based reenactment and outline future directions for robust, high-fidelity facial animation in the context of authenticity, security, and digital trust.