Nanjing sand, a type of greenish-grey schistous sand, is rich in weathered mica fragments, making it anisotropic and significantly differ from the round-grained quartz sand in terms of composition, grading, and mechanical properties. In this paper, the variational-autoencoder (VAE), a generative deep-learning model, was used to randomly generate realistic three-dimensional (3D) schistous sand particles. A training set was developed based on a micro-computed tomography scanned image sequence of the specimens of Nanjing sand. A total of 30,000 particles were chosen as the training set. Two thousand realistic 3D schistous sand particles were randomly generated by the trained VAE model and statistically compared with all the natural sand particles. The agreement of the statistical distribution and parameters between the generated and natural schistous sand particles confirmed the generative fidelity of the trained VAE model. Also, the validity and controllability of the generation is testified by the addition of Gaussian random noise and conducting linear interpolation to the latent variables. This methodology not only can be used for the clump template of the DEM, but also holds the potential of generating other geological entities and modeling complicated engineering geological processes.