Abstract: The static voltage stability critical point is of great significance to both the traditional analysis and the data-driven methods studying a power system in the ultimate states. Under the new situation of a power system, it is no longer practical to obtain the critical data by calling the point-by-point method iteratively. This paper proposes a generative model based on deep learning to realize the critical sample generation for the static voltage stability. Firstly, it is noticed that a critical sample is a special power flow sample. Setting the voltages on non-contact nodes as the features of the critical sample, the sample non-convergence of power flows and the non-zero injected power for the contact nodes are solved. Secondly, the Wasserstein generation adversarial network with gradient penalty combined with the transfer learning is established to learn the special constraints and distributions of the critical samples. Finally, the minimum singular value among the samples is selected to show the quality of the samples. The results of case study show that the combination of the WGAN-GP and the transfer learning is better in generating high quality samples than using the WGAN-GP alone.