Abstract

Talking head generation based on neural radiance fields (NeRF) has gained prominence, primarily owing to its implicit 3D representation capabilities within neural networks. However, most NeRF-based methods often intertwine audio-tovideo conversion in a joint training process, resulting in challenges such as inadequate lip synchronization, limited learning efficiency, large memory requirement and lack of editability. In response to these issues, this paper introduces a fully decoupled NeRF-based method for generating talking head. This method separates the audio-to-video conversion into two stages through the use of facial landmarks. Notably, the Transformer network is used to establish the cross-modal connection between audio and landmarks effectively and generate landmarks conforming to the distribution of training data. Then, these landmarks are combined with Gaussian relative position coding to refine the sampling points on the rays, thereby constructing a dynamic neural radiation field conditioned on these landmarks for rendering the generated head. This decoupled setup enhances both the fidelity and flexibility of mapping audio to video with two independent small-scale networks. Additionally, it supports the generation of the torso part from the head-only image with deformable convolution, further enhancing the realism of the generated talking head. The experimental results demonstrate that our method excels in producing lifelike talking head, and the lightweight neural network models also exhibit superior speed and learning efficiency with less memory requirement.