To obtain higher resolution images, a new model for image zooming based on total generalized variation is proposed. Original low-resolution image is regarded as sub-space projection of zoomed image to its own space, and the final zoomed image is obtained by application of second order total generalized variation regularization. By analyzing the properties of the new model, an efficient primal-dual algorithm is introduced. Experimental results suggest that the proposed algorithm can remove block effect and keep more details, so the high quality image is reconstructed; the model is better than the total variation mode and the wavelet model in terms of both peak signal to noise ratio and root mean square error.