Traditional oil painting relies on the artist’s subjective experience and skill accumulation, and the creation process is time-consuming and uncertain. In the field of modern digital art, although computer-aided painting technology provides convenient tools, it still has obvious deficiencies in simulating the texture, hierarchy and visual guidance effect of real oil paintings. This study proposes an oil painting picture hierarchical guidance design method based on the probability distribution of visual attention, and realizes intelligent oil painting generation by constructing Markov attention transfer model and hierarchical fusion generating adversarial neural network. In the method, the state distribution function is used to establish the attention transfer probability matrix, and the joint training framework of structural GAN and texture GAN is designed to train the model using a dataset of 700 oil painting images. The experimental results show that compared with the suboptimal algorithm Pix2PixHD, this method improves 7.024 in FID index, 6.38 in CFID index compared with PD-GAN, and the Manhattan distance is reduced to 36.19, which is significantly better than PD-GAN’s 75.33. As verified by the visual effect evaluation of 41 subjects, the generated oil paintings are better in terms of the overall aesthetics, color fullness and compositional momentum, all of which show good hierarchical and visual guidance effects. This method provides a new technical path for digital art creation, and improves the creation efficiency while maintaining artistry.
