if we were to turn headerimg into an image generator based on crowdsourced tagging, well, a lot of it would have to be, actually, real tagging. you can't just tag an image of hillary rodhat clintop with "diplodocus" and "cloaca" without also tagging it with: hillary clinton, woman, politician, etc because without this the network we are training cannot make a functional association between hillary and the more colorful terms. that you need a solid base of ~real~ tagging or it will likely come out as an abstract goulash of pixels -- not entirely boring, just mostly boring.
so i would propose splitting the crowd-sourced tagging into two categories: "normal" and "riced". so long as you mark which category your tags belong to, things can be arranged appropriately, even if it winds up being mostly me sitting there forever and adding the more boring tags. the system simply needs to know which of the two it is