The curent version works with a modiefied midpoint-displacement method on a sphere. It can generate different levels of detail on demand.
I think the generation could be twekead to only generate a small portion in a heigh level of detail and keep the rest low level, so a player sees everythig in high detail, while the computer only has to compute a fraction in hd.