Abstract: Neural implicit representation is an emerging shape representation paradigm, but most traditional implicit representation methods such as DeepSDF do not consider the local feature information of the overall shape, which has the problem of insufficient accuracy of topological details. In order to solve the above problems, an implicit 3D reconstruction method driven by latent vectors of the components is proposed, that is, the implicit field of the construction component reduces the signed distance value of the overall shape target point predicted by the model. LGI-RIF (Reconstruction of Implicit Fields with Local and Global Integration), predicted by the implicit field of the component, which can reconstruct the geometry from the observed data. This method trains the neural network in a low-dimensional latent encoding space and jointly optimizes in the decoder framework, designing three modules, EFP, EFCS and R3DS. The EFP module learns the latent vector distribution of the designed variational autoencoder, constructs the automatic decoder to learn the SDF implicit expression of the overall shape in the EFCS module, and reconstructs the overall 3D shape in the R3DS module. The experimental results show that the reconstruction accuracy of LGI-RIF on the ShapeNet and ModelNet 10 datasets has been further improved, and it has a good visual effect in the visualization reconstruction.