Data-driven seismic inversion yields promising results compared to physics-driven methods. The performance of datadriven methods highly rely on the quality and quantity of the training dataset. However, it can be rather challenging to obtain effective training dataset for realistic applications. Here we develop a real-time style transform method for creating numerous physically realistic subsurface velocity models from natural images. Particularly, we train an image style transform network to adapt the geological features from subsurface models into the natural images in real-time. Thus a large number of realistic subsurface velocity models are created which can be used as a training set for various data-driven geophysical inversion problems. Particularly, we demonstrate the effectiveness of the training data generated using our approach by testing both data-driven full-waveform inversion and data-driven traveltime tomography using Marmousi-like subsurface models.