Hi @devonorourke,
You are correct in lowering the minimum count sum perameters. In this case lowering them to 1 or zero may even be warranted.
The DIECODE percent variance explained is a relative percentage across each axis (in this case only three). For example, if the rank is two there will only be two axis and they will sum to 100% of the variance. This can make the % variance increase dramatically over other methods with many PCs. So I would not compare it between methods (like PCoA) and only compare it between other axis in the biplot.
As far as the gradients go, a clear sign you are dealing with a gradient is if you see a horseshoe affect in other methods. (See this paper for a more thorough explication of that). From what I see on your UniFrac plot that is not a problem.
Also be sure to update to the most recent version of DEICODE (v0.2.3). There was an issue with the loadings centering in v0.2.2.
I hope this helps,
Cameron