AD-Gaussian: an Audio Driven Face Synthesis Method Leveraging Differential Attention and 3D Gaussian Splatting

This paper presents AD-Gaussian, a novel audio-driven face synthesis method. It efficiently integrates multiple input modalities—audio, blinking, and head movements, while employing differential attention to guide specific spatial regions to focus on the critical features that drive their visual changes. By dynamically predicting the deformation of 3D Gaussians, the method enables visually coherent face animations synchronized with speech. Experimental results show that AD-Gaussian outperforms existing methods on key metrics for rendering quality and synchronization accuracy while maintaining real-time performance.