近年来，面部动画生成技术已成为计算机视觉领域备受关注的方向之一，生成的面部动画在唇音同步质量、情感可控程度等方面均有不同程度的进展。然而，现有研究在生成面部表情时，不可避免地影响了唇部运动。为提高唇音同步的准确率，通过生成对抗网络方法研究多模态驱动情感可控模型。该模型由两个自定义的深度学习网络串联而成，在输入任意无表情目标肖像图片后，通过音频、文本、情感肖像图片3种模态数据的驱动，生成高质量唇音同步、情感可控的面部视频。在该过程中，文本特征对音频起到辅助预测唇部动作的作用，是提升唇音同步质量的关键步骤。结果表明：在MEAD数据集上，相较于现有面部动画生成方法MakeItTalk和情感感知运动模型（EAMM），所提模型的唇部特征坐标距离（L-LD）降低了5.93%和33.52%，面部特征坐标距离（F-LD）降低了7.00%和8.79%。可见，所提模型提高了唇音同步及情感可控的准确率。

In recent years, facial animation generation technology has become one of the directions that has attracted much attention in the field of computer vision, and the generated facial animation has made different degrees of progress in the quality of lip-sync and the degree of emotional controllability. However, existing studies often unavoidably affect the lip movements when generating facial expressions. In order to improve the accuracy of lip-sync, generative adversarial networksmethod was used to investigate multi-modal driven emotion controllable model. The model consisted of two deep learning networks connected in series to generate high-quality lip-sync and emotionally controllable facial animation driven by three modal data, namely audio, text, and emotional portrait images, after inputting any expressionless target portrait image. In this process, the text features played a role in assisting the prediction of lip movements for the audio, and they were the key step in this technical study to improve the quality of lip-sync. The results show that on the MEAD dataset, compared to existing facial animation generation methods MakeItTalk and emotion aware motion model (EAMM), the proposed model reduces the lip-landmark distance (L-LD) by 5.93% and 33.52%, and the face-landmark distance (F-LD) by 7.00% and 8.79%, respectively. It is concluded that the proposed model improves the accuracy of lip synchronization and emotion control.