If you take the MNIST dataset and generate synthetic data that has variations on the original images, e.g. rotations, added noise, different background images, etc. you will get a higher model accuracy than the baseline model in exchange for a longer training time.