We proposed an image generation method of ophthalmic optical coherence tomography (OCT) based on the variational auto-encoders to alleviate the problem of insufficient images in deep learning and improve the performance of computer-aided diagnosis algorithms in ophthalmology. We created a generation network of OCT images based on the variational auto-encoders. In addition, we constructed three kinds of retinal OCT image datasets of age-related macular degeneration, diabetic macular edema, and normal situation based on the two public retinal OCT image datasets to train the network and obtain the generation models, respectively. The effectiveness of the image generation method was verified by subjective visual observation and objective experiments. Both the subjective visual observation and the objective experiments show that our method can effectively generate three kinds of retinal OCT images.