I understand and agree Oshayn.
What I do is take the image like in my example links above into photoshop. Make a grey scale and then use the "magic wand tool" to select the areas; the river first for example. And then copy that into new a new blank layer. I then do the same for each element in the image, each into its own layer
Once I have it all (or all I need or want), I hide the photos and just work on the selections I made. Making each into a displacement map for use in TG. BY saving each layer out by its self (over a black background)
Now when you have all the element in TG you can adjust the displacement individually (including inverting the mask). And no, you won't end up with a terrain that looks exactly like the landscape the photo you used was taken from. But you will have much more detail than a DEM with bad arch seconds, depending on the quality of the image (sun, shadow, clarity) and the quality of the selections you made of the various parts of the terrain from the photo.
It is a highly effective way of creating masks I am finding. I have used this method to make a new braided river, that beats my first one all to hell.
The Mountains get a bit harder, so what I do is create them in TG, and use the river and low level areas from the photo to mask out power fractals to create the surrounding mountains.
Everything you said in your post, Oshyan, is clearly correct. But what I just outlined above is a great work around. In fact in my test over the last few days, I have not found a case where i could not use it, so long as I had a good reverence image to start with. And of course, was willing to take the time to build the masks. It does work very well IMO.