<p>Acquiring the digital microstructures of porous media is vital for studying their transport properties. 3D reconstruction algorithms have provided alternative and cost-efficient options for acquiring such digital microstructures. This work presents a novel 3D multiscale generative adversarial network named U-GAN to reconstruct porous media with equiaxed pore systems. In this adversarial framework, the generator component of U-GAN characterizes long-distance connectivity and large-range correlations of training images via a fusion framework of multiscale features. Specifically, downsampling captures critical feature information step-by-step through three convolutional layers and three normalization layers. The upsampling process introduces skip connections to compensate for information that may be lost during the downsampling process, and deconvolution techniques are employed to achieve the fusion of multiscale features. Concurrently, the multiscale discriminators direct the generator towards a more accurate emulation of both the global architecture and details of the training images; random Gaussian noise is fed into the generator for each scale reconstruction to augment the variety of generated reconstructions. This study validated the effectiveness of the new model through coral sand reconstruction experiments, compared it with existing methods, and also assessed the impact of the training dataset size on algorithm performance. The quality assessment of the reconstructed coral sand was based on visual inspection and morphological descriptors. Moreover, the U-GAN model’s reconstruction diversity, resource cost, transport equivalence, and applicability to other porous media have been further discussed. These results demonstrate that the model can adapt to different training scenarios, producing high-quality reconstructed images that provide sufficient 3D pore structure data for investigating transport properties in porous media with equiaxed pore systems.</p>