Goms, I played a little more with my method last night, and had some further thoughts. I think you can get some decent variation, by maybe using difference scaler nodes or adding more of those chains in the .tgd with different seed and then masking between them. I had streets, pavements and grass areas by doing the last of those. Rather than relying on heights, I used the outputs of the conditionals as masks.
A 'Get Average' type node where you select a sample area would be quite usefull to make up a blur function, but I'm not sure it would help much in your case. What you could do, and it would be a bit of work, is comparing a voronoi noise with a spacial offset, then create a ramp between those two values. But I'm not sure if this would work when applied.
Another problem, is that TG really doesn't like sharp linear angles, you would have noticed this 'cracking' effect along your building edges. Unfortunately, if I had to animate something like this, I think it could become a world of hate with disapearing triangles on frame rendering. I would be tempted in using populations, at least I'd have decent textures.
That said, I was trying to make a form of squared rock a while back, which I think some of the techniques in this thread could be very useful, primarily mixing the x and z only cell scalar noises, coupled with the same seed diff scaler to blend, and conditionals to allow the tallest through. You could get some nice cracks and blocks when you apply this to a fractal terrain. But I guess I should reserect that topic rather than spam this one with idle ramblings from me.
Cheers