Following the high level steps of what they're doing isn't too hard. The problem is I lack to the knowledge to have the "tools" to understand exactly how to derive the steps on my own from scratch.

The other issue is that as shown in the close ups of both of them, they don't entirely work. There's some step missing to produce even the ideal box filtered result I would expect. It's almost as if the variable line filter width is being negated such that it ends up no longer anti-aliasing the lines at all. But again, I'm not mathy enough to be able to pinpoint exactly why.

I should note there's aliasing in the RGSS method, but that's somewhat expected due to the failings of that particular technique. Thin lines that line up just "right" with the rotation of the super sampling pattern will end up aliasing again, which is what can be seen in the examples for that.