Not all kinds of wavelet filters are symmetrical. The orthogonal wavelets, for example, are not (as opposed to the biorthogonal family, which is preferred in image compression exactly because of its symmetry). You can think of the filters as basis functions, just like spline basis functions. That does not mean that the shapes built from that basis all must be symmetrical.
My explanation probably was a bit too much information cramped into too little space ... I'll try to give at least a few more buzzwords to untangle the various concepts that together can make up a fractal landscape.
1. procedurally generated, deterministic
You don't want to store a planet-sized map, so you want to compute just the currently visible parts on demand. That computation has to yield consistent results every time you revisit that part of the landscape, thus the procedure used has to control randomness in a way that makes it perfectly repeatable.
2. localized
You want to compute small parts of the whole thing, without the need to compute much of the invisible surroundings.
3. efficiency
The process must be reasonably quick, so that you can indeed compute individual tiles of the map without a noticeable pause.
4. some control about the type of fractal
You'll probably want to control some parameters like, are these rolling hills or steep peaks and cliffs. Randomness shouldn't be so wild as to create 20 mile high mountains that would appear rather unrealistic to a human observer. (Okay, maybe your planet is made of aluminum and has no gravity.
)
The wavelets I suggested "automatically" fulfill a few of these criteria. They give you some amount of
control, because you can freely adjust the amplitude of the various frequency bands (corresponding to structure size at various detail levels). That allows one to constrict the overall hight of the mountains, but still get a pretty rough skyline (or vice versa, if you wish). Wavelets are
localized, too, at least if you chose one of the many filter types that have compact support.
Wavelets can be quite
efficient. Google for "Lifting Scheme", and chose a suitable wavelet (say, the Daubechies (2,2)-wavelet, also known as the 5/3 filter pair, or occasionally as the first order spline wavelet). This one can be computed purely in integer arithmetic, with just a dozen or so additions/shifts per pixel, and the computation can be nicely parallelized for SIMD.
The quadtree I suggested because its recursive nature causes it to be
localized. As you descend from the root of the tree to a particular leaf, you are following one specific, unique path associated with that one leaf. If information from that path is used as random seed, you have
determinism. Quadtrees tend to be
efficient, because their depth grows only as the logarithm of the number of leaves (this is true of most any tree data structure).
Regardless of the specific method, wavelets or no wavelets, the four buzzwords are probably qualities that you need to look four in the candidates.