Unable to explain differing accuracy using GaussianNB for image recognition on two different data sets

Question

There are two data sets of images representing the digits 0 to 9. One of the datasets is the classical MNIST data set within keras that has 60000 training images in 28*28. The o