Abstract: The complex pore structure and strong heterogeneity of shale make it difficult for conventional numerical simulation methods to capture real patterns, and data acquisition is difficult. To solve this problem, a single-image GAN model for shale random reconstruction based on ConSinGAN is proposed. This model introduced and improved the CBAM mechanism to realize adaptive feature refinement and improve the ability to focus on local details. The model added a multi-branch discriminator to keep the reconstruction results consistent with the training sample in terms of the content information and layout information. The model integrated the Hinge loss function of SVM to maximize distance between the real and fake samples to ensure the training stability, and also used a variety of loss functions. The experimental results show that the model can effectively reproduce the geological properties of shale, such as pore space, MPC.