A hybrid attention generative adversarial network for Chinese landscape painting

In traditional Chinese painting, the genre of landscapes is unique and universally valued. For an untrained person to achieve such results is very difficult, requiring mastery of such things as brushwork, composition, and color. In this paper, we propose HA-GAN to transform sketches into Chinese landscape paintings, a new GAN-based framework that builds upon a hybrid attention generator and a discriminator. To generate more realistic landscape paintings, we have designed a hybrid attention module (HA) containing style attention, spatial attention, and channel attention. The proposed hybrid attention module organically integrates the correlation between channels, the spatial long-distance dependence, and the extraction of image-style features into one module. HA combines the advantages of three attention mechanisms and can extract important features of traditional Chinese landscape paintings from multiple dimensions. This combination approach helps the proposed model to understand the input data more accurately and thus improves the model performance. Moreover, a novel loss function is proposed to guide the training process of the model. The experimental results show the advantages of the proposed method compared to the comparison methods both in terms of quantitative and qualitative evaluation. The proposed method can generate realistic landscape paintings similar to those created by human artists.