Generative adversarial network (GAN), a deep learning technique, is widely applied in the field of remote sensing because of its ability to extract features from large input data and generate more realistic forecasts of meteorological images. At present, however, the application of GANs in atmospheric motion vector (AMV) retrieval is rare, although AMVs are important data source for numerical weather prediction (NWP), especially in data assimilation. Based on this, a method for retrieving AMVs from geostationary satellite images using pix2pix, a type of GAN, is proposed. The pix2pix model is used to convert remote sensing images into wind vector fields at 850 hPa and 200 hPa. With the best data and model architecture, the AMVs obtained by this method are comparable to the AMVs retrieved using traditional algorithms. This method avoids the drawbacks of traditional algorithms, such as the inability to obtain complete wind fields at a certain level, difficulty of height assignment, and sparse AMVs at lower atmospheric levels. Case analysis shows that this method also performs well for specific weather systems.