>>106244741
>>106244757

I'm using a dynamic split between High and Low through Wan MoE KSampler, which uses the same formula as the one used by wan team.
And the result in my tests is always skewed towards less steps in High vs Low.