Aiming at the problems of insufficient interpretability of latent features， limited disentanglement capability， and poor identity consistency in the existing 3D face generation methods， a 3D face generation method based on Latent Feature Enhancement for Disentanglement （LFED） was proposed. Firstly， the hierarchical clustering technique was used to construct a vector discretization module， so as to promote the potential features to absorb prior knowledge and improve the disentanglement performance. Secondly， a positional attention module was designed to integrate location information of the potential features selectively through element-by-element summation operation， so as to ensure the identity consistency of generated faces. Finally， combining prior knowledge and position information， the maximum normalization technique was used to enhance the interpretability of potential features in face generation process. Experimental results demonstrate that the proposed method achieves an accuracy of 95.67% in the latent feature disentanglement metric — Variability Predictability （VP）. Compared with Swap Disentangled Variational Auto-Encoder （SD-VAE）， Local Eigenprojection Disentangled Variational Auto-Encoder （LED-VAE）， and Spherical Harmonic Local Eigenprojection Disentangled Variational Auto-Encoder （SHLED-VAE）， the improvements are 14.96， 14.33， and 12.46 percentage points， respectively. It can be seen that the proposed method enhances disentanglement performance while maintaining good representation and reconstruction capabilities.