Inversion of large-scale gravity data is generally a challenging problem due to memory requirements and computational cost. In this study, based on VNet, we present an effective strategy for large-scale gravity inverse problems by simultaneously tackling several base-scale gravity data. We first construct a large number of base-scale geological models including gravity sources with different shapes and dimensions, and also their forward model data sets. Then, the idea of semantic segmentation is used to train an inversion network. In the next step, a finite number of base-scale and similar size area of gravity data, clipped from the original large data set with a fixed stride, are fed into the trained network. Finally, the individual recovered models can be combined together to yield the inversion result for the whole subsurface area. The feasibility and effectiveness of the presented inversion algorithm are discussed on a large-scale and complicated synthetic model. The algorithm is, then, verified for the inversion of the gravity data set obtained over the Morro do Engenho complex in central Brazil.