Deep learning (DL)-based three-dimensional (3D) gravity inversion has emerged as a promising approach for subsurface density reconstruction in near-surface resource exploration, offering significantly lower computational cost than conventional inversion methods for large datasets. However, existing convolutional neural network (CNN)-based gravity inversion studies are largely limited to binary or constant-density models, thereby limiting their robustness in reconstructing realistic subsurface density models. In this study, we evaluate the performance of three encoder–decoder CNN architectures, UNet, UNet++, and ResUNet, for 3D gravity inversion. We modify both the training approach and the network architecture by using synthetic models with varying geometric complexity and density contrasts and by incorporating a Softplus output activation function in the output layer of all three networks. These modifications enable the networks to simultaneously reconstruct subsurface geometry and density contrasts. Among the three architectures, UNet++ consistently achieved the highest Dice scores, the lowest gravity-data misfits, and the lowest prediction uncertainty for the synthetic models. We further validate the proposed framework by applying it to the San Nicolás massive sulfide deposit, demonstrating its applicability to real gravity data and its potential for near-surface exploration.