Imagine if you will training it to generate a 3D model's facial features instead produced in 3d, now sure it will take allot longer then the current version to train and require a greater degree of complexity, but the research paper alone indicates that it took a relatively short period of time for the ai's to train up to the level it currently does which by no means is perfect but still is quite impressive considering (https://www.lyrn.ai/2018/12/26/a-style-based-generator-architecture-for-generative-adversarial-networks/), mayhap by having multiple ones trained to generate various body features and perhaps one skilled at putting them all together, in the future this kind of AI trained generation may be something to look at in some future versions of makehuman, mayhap not any time soon, but its exciting to see the potential for new techniques, afterall can you imagine being able to generate 3D characters with this level of detail.
Its major flaw is from reading it due to the sorta organic way it generates a new face, the only way to actually select what type of anything you're likely to get is to limit a given generation sequence's ai's access to the master database of types in this case faces to say a given race, sex or type, by restricting the available library the type of face it is capable of generating is similarly compromised allowing you to in a round about way set what kind of face is generated, however this will also reduce the quality of the generation unless a vast library of every given race/sex and type is made available in the master library to mitigate this, so frankly its unlikely the kind of thing you could run entirely locally, as at the very least it would require some kind of mass-accessible database of large size would be required at least until computing has advanced to the point where the average hard drive could easily accommodate the massive who knows how many terabyte clone of the master DB and the obscene download required in a speedy fashion.
Still i can see some kind of plugin allowing you to select the type of face and generate a request to a master server which will go off and divy one up for you based on some future version of these AI techniques transmitting to you the finished result to your local client happening in a near-future (5-15 year) timescale, I do hope it ends up in makehuman when it does, as frankly I struggle to make a face that is not ugly as sin in the current editor, its not so much the lack of options as in plethora of them that make any such endeavour on my part seemed doomed, it would be real nice if some fancy AI based doo-hikey could do that kinda work for me
, hence my interest, either way ty to the makehuman devs for all the work and yeah curious as to your more trained take on the potential of such a system.