I'm not sure I understand the part about some of the shorelines and river edges being cliff-like. That's usually because the input data has a step in values across that boundary (e.g. goes from 0 to 10 instead of 0 to 1). There might be a way to help using some texture-to-selection tricks but I'm not sure what it would do to the other parts of the data. Any chance of seeing a sample of the problem areas along with a sample of some not-problem areas?
I haven't used that feature for a long time so I can't recall the exact parameters at the moment, but in general, a while is "as long as it takes". Not helpful, I know, but how long will depend on the settings, your surface size, machine processor speed, and installed system memory. I vaguely recall that I had sort of abandoned that feature because it wasn't doing quite what I had originally intended.