The Relative To Pixel node considers the original resolution of the image with no transformation, while it is more useful for it to consider the realized size instead, because that's what later processing nodes will be using.