Modern design intelligent training assistance platform based on the Lingnan architectural image generation algorithm is studied in the paper. In order to make the network better train unpaired image samples, this paper introduces a loss function composed of the same mapping loss term and perceptual loss term on the basis of the original loss function. Then, the novel style transfer algorithm is designed. We propose CycleGAN without training on paired datasets, and the network employs two general generators and two discriminators, respectively. Two generators are used to then convert between the two domains, and two discriminators are used to distinguish the true and false images in the two domains. Then, the designed model is implemented with the novel intelligent training assistance platform and consider the factor of the Lingnan issues. Through the experimental simulation, the performance is validated.