Face image super-resolution has received increasing attention. However, since the face has a lot of fine textures, it is very difficult to rebuild for large upscaling factors. We propose a new method for face image SR, using residul dense block(RDB) as the basic unit and the Inception architecture is combined in the low layers. We use the relativistic GAN and the improved perceptual loss defined by the features before activation.For the large scaling factors, our GAN is progressive both in architecture and training. The network proposed achieves excellent performance in the reconstruction of low-resolution face images, especially under large scaling factors such as 4x and 8x.