Chinese figure painting is one of the most culturally and aesthetically valuable visual art forms in the history of Chinese art. Factors such as the high creation base and long creation cycle of traditional creative forms have largely restricted the innovative development of traditional Chinese portrait painting. In this paper, we use the DualStyleGAN deep learning framework to train the style migration of Chinese traditional portraits, and obtain the best data for generating models through qualitative analysis. Firstly, this paper collects and extracts images of Chinese portrait paintings to construct a dataset of Chinese traditional portrait paintings. Second, the self-constructed Chinese traditional portrait dataset is used as the target style dataset for style migration training. Finally, the optimal performance values are derived by comparing and analysing the effects of loss function and weights on the generation effect under this framework. The results show that DualStyleGAN can effectively migrate the styles of traditional Chinese portraits, and also adjust the degree of stylisation through the weighting values. This study combines Chinese traditional portrait elements with deep learning techniques to expand the application range of portrait style migration and generate Chinese traditional portrait style images with the best visual effect.
Read full abstract