An image model builds a picture out of two things: a random noise sample nobody chose, and the words someone typed. How much of the result each of them accounts for is the question here.
Every station cuts between two pictures in the same place. One thing was changed and the rest held still, so whatever moves on screen is what that one thing was doing.
Every number here is measured in your browser, from the pictures on screen. Where a published figure exists to check it against, that is printed beside it.
Arrow keys move between stations. So do the marks at the bottom.