In fact, generated data aren’t “new data” it’s your training set with operations. Each of your pictures will generate “new” transform data in the RAM as zoom, shift, translation, inversion. The goal is not to make synthetic data like VAE or GANs but to use your set to augment the training set. It’s not possible even interesting to save this kind of transformations because they are modified in the RAM. If you want to build synthetic data to increase your dataset and save this pictures you need to generate them with generative models.