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,

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.