I like the idea of header only libraries, but I'm not too familiar on what the downsides may be.

I actually ran into a downside. ANL uses gradient look-up tables to generate the Perlin noise. These are static tables, and header-only libraries don't really like static data. Hard to avoid multiple-definition errors if the library is included more than once. I'll probably just stay the course for now, at least until (if) I can eliminate the static tables.

Is it not possible to create these tables at runtime?

It is, but then I was looking at more refactoring to the base noise generator classes to eliminate them. They're hefty enough that I didn't want them to be members of the generators, otherwise each generator would own their own copy. I could make them static members, but that just puts me right back to where I am now, needing a .cpp for static tables. Generating the tables at run-time still needs there to be a pointer variable, somewhere, that all the various instances can reference, unless I give each generator their own instance. There are a few other things I could have done like wrapping them in functions and taking the chance that the inline operator wouldn't be taken literally, and that the compiler would instead consolidate all the individual instances. In the end, my C++ kung fu isn't strong enough to come up with an elegant solution that doesn't require a huge amount of refactoring. I could spend the time refactoring it, or I could just keep it as-is and spend that time instead on adding more function types and writing more documentation.