Due to computational cost, traditional style transfer models may not be suitable to generate stylized images on a device with limited computational capability, e.g., a client device such as a smartphone, tablet, or wearable device. This disclosure describes techniques to train a computationally inexpensive generative model to generate a stylized image from an input image. The model is trained using training images. The model is trained such that the difference between an output stylized image of the generative model and a reference stylized image is minimized. The trained generative model is suitable for use on a client device to generate stylized images, in real time.