It seems like defining how the semantic sliders behave would be much more work than having an artist do the 3d modelling. And either way, you're going to need that 3d artist in the first place. Am I missing something?
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!)
I also don't think (I haven't read the paper, yet!) you need to do anything to get the semantic sliders other than to label existing data. If you're going to design a single chair, it's probably more time efficient to have an artist make a 3D model to your specifications. On the other hand, if you're going to design something like a customized biomedical part with variations that depend upon the patient, then this could easily be a net win. Not to mention if you'd like to automate that design, you now have a much smaller dimensional parameter space to play with.
As someone who creates 3D models from scanned objects... hah haha hahahahaa. <sob> It's so freaking hard. I wish it weren't so, but it's so frustrating, it is probably the aspect of my job that I enjoy the least. I use a Faro CMM arm with a laser line scanner and polyworks (IIRC that whole package costs about $90k), and you can get decent scans of certain objects without much effort, but it's really hard to get good definition in small details, and some surfaces (shiny ones, transparent ones, ones where light scatters slightly below the surface) and some details (holes, crevices, small protrusions) are really hard to capture. And even still, the resulting model has high complexity... If you want a smooth mesh without holes or other aberrations, that's another layer of work. And if you need to convert it to proper NURBS surfaces, that is yet another layer of work (and one that takes an entirely separate skill set).
I know these technologies are getting better all the time, but that one's still a really hard problem that's waiting for a better solution.
Example: Download a toy plane model from thingiverse. Decide you want it to look "stealthier", this can modify it. The plane model was not one in the original training set.
The process appears to take a bunch of example models, then asks people to score them on various parameters. The algorithm then works out how to use that knowledge to modify shapes in those directions.