The WD models already give a value of "certainty" (don't know if that's the best word for describing this, however I'll use it), so it could be used to give weight to the output.
For example:
let's say the WD model returns:
long_hair: 40
medium_hair: 80
Instead of returning long_hair, medium_hair it could return (long_hair:0.4), (medium_hair:0.8) which would go closer to the image.