A method based on deep feature to reconstruct face images was proposed in this paper. The deep features were first embedded into the W⁺ space of StyleGAN2， and the embedded w⁺ vector was optimized by gradient descent and fed into a pre-trained StyleGAN2 model to generate face image with resolution of 1 024×1 024. The experimental results showed that the reconstructed images had high visual similarity with the corresponding real face images. When using the same extraction network， the type-Ⅱevaluation TAR of the reconstructed images from LFW and ColorFeret datasets was 96.04% and 100.00% respectively when the FAR was equal to 0.1%， and the pass rates of liveness detection were 88.67% and 74.67% in two products. The proposed method could reconstruct face images in high resolution and achieve high feature similarity between real and generated face images.