It that is true, it should be possible to get large wins in LOC if one either is willing to give up orders of magnitude in speed and memory usage, or is equipped with a sufficiently advanced compiler.
There likely also is some room for improvement if one is willing to change the file format (.psd is not known for its consistency), but that would likely be a rounding error.
The reduction comes from making the code way more abstract and much more difficult to reason about in concrete contexts. So ya, it is probably possible, but no, our heads might not be very capable in writing/maintaining/debugging such code even if the computer could execute it. As an extreme, consider "very small code" contests in the demo scene, where various crazy tricks (self modifying code) are used to drastically reduce code sizes, none of which would ever be used in practice and have little SE value.
That being said, a 100x reduction in code size seems like an exaggeration.
I challenge anyone making such claims to construct an even _moderately_ realistic example of where such reduction would be possible.
Here is a very simple and already very optimistic model that is still way below a 100x reduction: say my program consists nearly exclusively of sorting of various arrays and I write out bubble sort in full each time using for loops etc. rather than call a sort() procedure. Bubble sort takes roughly 20 lines of code to implement, so I would at best get a reduction of close to a 20x, rather than a 100x one.