Wow, some cool-looking pics in this thread! Funny that I also happened to be working on terrain .frag when I found this. Although I'm using 2D instead of 3D noise. I find that 2D Perlin noise, when modulated and combined in a certain way, gives great-looking mountains (pictures attached). The black and white picture is colored using one of the second order partial derivatives of the landscape surface. It highlights the ridges nicely.
Modelling these realistically could be difficult and IMHO not worth it.
BUT you can easily mimic that and I did it with higher frequency sin noise applied. See my previous imagery in this thread.
I've been trying to tackle just that problem. Specifically fluvial erosion, like this:
http://www.nordiclandscapes.com/Mountains-Volcanic-landscapes-II/slides/mountain-erosion-vegetation.jpg. Of course there are simulation-based methods that run on a grid, but those would be inconvenient to do in Fragmentarium, and they're not fractal-based.
I found a neat trick for creating pseudo-erosion effects here:
http://www.decarpentier.nl/scape-procedural-extensions, which I implemented in my terrain. But I also have a new method in the works.
My terrain implementation is very similar to his "swiss turbulence" example, with a few differences. He used the classic "ridged multifractal" approach to terrain, where you take the absolute value of each octave of Perlin noise to create sharp ridges, and then modulate the strength of each octave based on the sum of the previous octaves, to basically flatten out low regions of the map creating flat basins between the ridges. What I do differently is, for each octave I generate two different Perlin noise samples, take their absolute values, then multiply them together. This emphasizes individual peaks more, whereas his method tends to create longer ridges. Also, instead of just taking the absolute value, I use a "smooth abs" function that doesn't create such a hard crease, and is parametrized by a smoothness value. The function looks similar to this:
http://amath.colorado.edu/faculty/becker/Dual_vs_Primal_smoothing.pngAnother difference is that I modulate the strength of the noise not only on the height, but also on the cumulative derivative of the previous octaves, which smooths out any non-sloped region, whether high or low.
Another difference is that he forgot to multiply the derivative of each octave by it's frequency (which may have been intentional). I added that in. However, doing so can make the terrain look too messy, so I scale down the contribution of higher octaves by their own multiplier.
The last difference is that I use an additional sample of very low-frequency noise to modulate the height of the mountains, this creates large flat regions in the terrain and makes it look more believable.