Abstract: In order to solve the problems of insufficient training data and unreal texture in the current deep learning methods in 3D face reconstruction, a weak supervised hybrid model is proposed. Using a single two-dimensional face image, the face 3D deformation model (3DMM) coefficients were regressed by densely connected convolution network (DenseNet), and the feature differences of different levels were used as weak supervision signals to improve the generalization ability of the model. On this basis, the image convolution network (GCN) was used to extract the facial detail features of the input image to optimize the reconstruction texture. The experimental results show that this method can reconstruct the fine three-dimensional model of human face without training label, and is better than the existing methods in terms of normalized average error.