Abstract: To address the challenge of accurately controlling local features in the 3D face generation task, a 3D face generation method with local feature projection of spherical harmonics is proposed. The method first decomposes the 3D face into several attributes, each corresponding to different latent variables. Then, each attribute is mapped onto a unit sphere, and the feature projection is performed on the shape attributes using the spherical harmonic function. Finally, the feature projection loss is defined and applied to the encoder and decoder of 3D-VAE to learn local latent feature representations with high disentanglement and high interpretability by measuring the differences between the projected features and the latent variables. Comparison experiments are conducted with several mainstream methods on the UHM dataset, and the results of the VP metrics show that the local feature disentanglement effect of the proposed method is optimal among all the methods; moreover, the training time of the proposed method is the smallest among all the methods, which is reduced by about 37% and 54% compared to the SD-VAE and SD-VQVAE methods, respectively.