[methodOverdrive]

I didn't read the paper (so I might be full of shit) but I think the idea is that the semantic sliders represent vectors in a vector space that is learned by a machine learning model, based on a sample set of examples. So the idea would be to take a bunch of examples (made by a 3D artist, taken from existing work, scanned in from real objects, etc) and first rate them in each category (which is subjective, and does still take a fair amount of human work)... but then you use an algorithm to train a model that can generate new examples in that vector space. Then, the sliders just modify the values of the different components of a vector representing a new, generated example. I doubt that the behavior of the semantic sliders is hard-coded - instead, there's a general algorithm for coming up with new sets of semantic sliders. So - ideally, for problems this model works well for, you would ideally be able to dramatically reduce the number of 3D models you need to characterize a whole space of parametrized variants.

EDIT:

Just actually went and read the paper. It's not machine learning, it's crowd-sourced. So it really needs a lot of people working on it... so I think that your concerns are totally well placed).

(Maybe we'll see the machine-learning version of this in the near future!)