I did some experiments that suggest you can reintroduce more variety in the results given by qwen simply by modifying the sigma values.
First the results (visit this link on imgur for the full image scale):
On the leftmost there is the unmodified simple scheduler, the from left to right, the scale decrease from 0.96 to 0.93. In the top-down direction, various seeds are tested.
This also works with an input image:
Now, how this is done:
Basically, you use your usual SamplerCustomAdvanced node, connected to your BasicScheduler, inbetween, you have this SRL Eval node from srl-nodes, which allows you to run arbitrary code (I'm usually using that node to debug while developing custom nodes). And you replace the variable c by the amount you want to scale down.