I meant that it is limited by the size of the input layer. A 2000x2000 pixel image of a face has to be downscaled, because otherwise you’d need 4 million nodes on the input layer and that’s impractical.