Method for Speech-Driven 3D Face Animation under Personal Style Guidance

In the process of generating speech-driven 3D face animations, incorporating the personal style of the target character can significantly enhance the realism and expressiveness of the animation. Existing methods are not sufficiently detailed in reflecting personal style and need better adaptation to new characters' styles. To address this, we propose a method for generating speech-driven 3D face animation guided by personal style. First, we design a personal style extractor based on style attention network. This extractor derives latent facial movement features from facial action sequences. These features are then adjusted and integrated based on the semantic space distribution of audio features, forming style features that match different animation frames. Next, we construct a style-guided feature fusion decoder based on Transformer architecture. Thanks to its multi-head attention layers, this decoder can refer to the context of personal style features while mapping audio features to 3D face animations, allowing the generated 3D face animation to better mimic the target character's style. Experiments on the publicly available VOCASET dataset show that our method accurately reflects the style of existing characters while maintaining precise synchronization between lip movements and the driving audio. Additionally, with a two-stage training strategy, the method adapts quickly to new character styles from short video clips. Experiments on the self-constructed dataset of new characters show that the method achieves low vertex errors and high style similarity, demonstrating strong generalization capability for new character styles.