END OF AN ERA, FRACTALFORUMS.COM IS CONTINUED ON FRACTALFORUMS.ORG

For the Perlin noise, it looks like you're using the old version with the S-curve 3x² - 2x³, which is not continuous in the second derivative. That produces the slight creasing at peaks and troughs. Instead, change the blending function to x*x*x*(x*(6*x-15)+10).

When adding octaves of Perlin noise, you also want to add some big offsets to the position to avoid correlations near the origin.

If you're using the standard hash table method for Perlin noise, it doesn't use many random numbers so the generator isn't important.

For a good generator, I used to use Mersenne Twister and had no problems with it, but switched to complementary multiply with carry. Make sure it's a full CMWC, not just MWC. To pass the lottery standard tests, as well as being suitable for high resolution chaos game fractals, it's not enough to have a long period and low correlations. It's also necessary to produce all possible sequences of a given length somewhere in the cycle. George Marsaglia's CMWC generators do this, and are simpler than Mersenne Twister and more flexible for seeding or customizing.

@ David
Very nice results but I'm not familiar with that type of code and it looks like it might be over my head

@ Xeno
I believe the hash table method is where you have the array of 0-255 in random order?
I'm using a different method which I found here - http://www.gutgames.com/post/Perlin-Noise.aspx
I just used common sense / analogy to make it 3D
I don't notice anything that looks like "3x2 - 2x3" though
Maybe this would be a better approach? http://lotsacode.wordpress.com/2010/02/24/perlin-noise-in-c/

@ Syntopia
GPU noise would be ideal for me, but as of now I don't know how to do anything on GPU

The first link there is not "Perlin" noise. It's a poor interpolation between random values. This type is generally called "value" noise, and even with better interpolation, has visible axis alignment. (I wouldn't recommend his method for cellular noise either - sqrts in a nearest neighbour search!)

The second link is the standard (original) Perlin algorithm, with the hashed array. Specifically, it's Perlin gradient noise, because it calculates random gradient values at each lattice point, and the result is less axis aligned than value noise. You can also modify (or add a function) to produce Perlin value noise by just interpolating the values instead of the gradients.

The smoothing function at the bottom is as I mentioned, x * x * (3 - 2 * x); this is the part that you can improve with the fifth order polynomial. The difference is more visible in 3d than 2d.

What you want is linear interpolation and smooth with the fifth degree. This gives the smoothest curve for a single octave. It's noticeable on either a 3d surface or a 1d plot.

Nice !
Am just trying to find time to do something similar, but also to try an alternative method that I think will be a lot faster (in shader fragments but is a lot slower in UF).

kronikel, solution is to extend terrain with virtual space as show on the attached picture;
blue - virtual space (no need to allocate mem for it)
red - resulting terrain
green - seed poles

use for blue height function where result is avg of seed poles + random deviation based on roughness