Abstract: Audio-driven talking portraits synthesis endeavors to convert arbitrary input audio sequences into realistic talking portrait videos.Recently,several works on synthesizing talking portraits leveraging neural radiance fields(NeRF) have achieved superior visual results.However,such works still generally suffer from poor audio-lip synchronization,torso jitter,and low clarity in the synthesized videos.To address these issues,a method based on regional saliency features and spatial volume features is proposed to achieve high-fidelity synthesis of talking portraits.On one hand,a regional saliency-aware module is developed,dynamically adjusting the volumetric attributes of spatial points in the head region with multimodal input data and optimizing feature storage through hash tables,thus improving the precision and efficiency of facial detail representation.On the other hand,a spatial feature extraction module is designed for independent torso modeling.Unlike conventional methods that estimate color and density directly from torso surface spatial points,this module constructs a torso field using reference images to provide relevant texture and geometric priors,thereby achieving more precise torso rendering and natural movements.Experiments applied to multiple subjects demonstrate that,in self-reconstruction scenarios,the proposed method improves image quality(PSNR,LPIPS,FID,LMD) by 10.15%,12.12%,0.77%,and 1.09% respectively,and enhances lip-sync accuracy(AUE) by 14.20% compared to the current state-of-the-art baseline model.Concurrently,there is a notable increase of 14.20% in lip synchronization accuracy as measured by Sync metrics.Under cross-driving conditions with out-of-domain audio sources,the lip synchronization accuracy is achieved improvements of 4.74%.

Key words: Talking portrait synthesis, 3D reconstruction, Audio-video synchronization, Nerual radiance fields, Attention mechanism