Virtual Human Dance Movement Generation Technology Based on Transformer and VITON Network

In recent years, with the rapid development of computer vision and deep learning technologies, virtual human technology has been widely applied in fields such as entertainment, education, gaming, and metaverse. But the existing methods still have deficiencies in movement fluency, music matching and naturalness of virtual human posture. In order to improve the quality of virtual human dance generation, the study proposes a movement generation technique based on Transformer and virtual try-on network. By constructing a dance video dataset, the encoder-decoder structure is used to establish the mapping relationship between audio and dance movements. Virtual human dance synthesis is realized. The experimental results show that the matching rate of this method increases rapidly with the increase of the dataset, and gradually tends to stabilize after the dataset size exceeds 10, with the matching rate close to 0.93. Its smoothness and matching degree of the movements are 0.9 and 0.88, respectively. In terms of the generating speed, the fastest generating speed is $0.50 \mathrm{~s} /$ frame. The study shows that this technology has good performance and can effectively improve the accuracy and realism of virtual human dance generation, providing support for applications such as virtual idol and intelligent entertainment.