Automatic generation of landscape images based on deep generative modelling

With the rapid development of artificial intelligence technology, deep generative models provide new opportunities for the intelligent transformation of landscape design. Aiming at the deficiencies of existing generative methods in terms of scheme diversity, design feature decoupling and small-sample adaptability, this study proposes a hybrid generative architecture that integrates StyleGAN2 and diffusion model, combining with a migration learning strategy to optimise the model generalisation ability in small-sample scenarios. By introducing a reverse denoising mechanism to enhance detail generation, and using PCA and clustering methods to quantify the feature decoupling effect, the model achieves high-fidelity image generation (FID ≤ 25) and feature independence control (clustering purity ≥ 85%) on the publicly available dataset ADELAIDE Landscape Dataset. Experiments show that the model can effectively capture the spatial texture features of Dai villages and terraced fields in the image generation of typical mountain landscapes in Dehong Prefecture.