I’ve experimented a bit, but I think the LR filters aren’t the right order, or some technical thing that means I can’t one-for-one replace them. It’s been a while since I look at this. I do want to phase correct the cross~ approach (perhaps loads of allpass~s in there or something).
Even better would be doing an LPC to generate an actual filter from the melband values and skip cross~ altogether. Or I can even dream of a @blocking 2-friendly HIRT so I can do @jamesbradbury 's original approach but in realtime without massive lagspikes due to low-priority threading stuff.
Definitely the plan 
Was thinking of using some of the genetic algorithm crawl stuff we’ve been doing to just learn a mapping that, I imagine(/hope), would generalize out if given enough examples.