I think "generated" may be too strong a word for what this system seems to be doing. It feels more like it has found a bunch of averages among a large enough set of similar pictures of some fairly homogenous celebrities. When enough of those averages are put on a gradient along a particular axis you get transformations that seem magically smooth. In the example you gave I think it didn't have enough source photos along that axis. I'll further guess that this works best for celebrities that are generally considered attractive because attractive faces are already averages.
The point is that the image has been transformed into a latent space encoding and interpolating those latent-space variables creates those mind-blowing effects (imagine just moving from one value to another via a simple convex combination). How are those latent variables constructed and what do they represent (or if they could be described easily) is completely up to non-linear optimization process running on a huge number of dimensions.
I don't understand this as the image being transformed, "space" and "dimensions" seem to originate not with the visual features of the images but with the attributes that the CelebA image set is annotated with. The coverage of the set is not uniform with respect to gender and skin color, I'm not 100% sure how to interpret the attribute values but a some quick math shows that images with "Pale_Skin" outnumber their opposites by about 20 to 1.
This is one of the biggest challenges with AI in my opinion. The models can generate the transformations but they have no concept of correctness when applied to vague generative tasks like creating a face based on a set of existing photos.
Basically, there's just one level of cognition. In this case, the AI would only achieve that expected fidelity if the system is layered with more and more models that aim for correctness and accuracy (does this look like a woman, does this look like a mouth, does this look like a nose, etc). The problem with this approach is that it becomes incredibly hard to determine what's needed to be 100% successful at a complex task.
This is the reason why I think we are still far far away from a fully cognitive AI and is the same reason why you only see AI used for very narrow use cases.
Self-driving cars seem to be the first real attempt to have a broad AI system applied to a super-complex and unpredictable field, but I always see conflicting information regarding the progress and challenges in this area.
I think that could feed the output of the GAN into yet another network that assesses the quality of the generated image and automatically tweak the parameters a bit until it doesn't look like an alien.
In fact, that network is probably already part of the original GAN training phase.
Ultimately, this algorithm tries to find an single directional vector in a 512 dimensional space that approximates what it means to change skin color.
Expecting that this works all the time (or expecting that all points in this 512-dim space result in a beautiful person) is probably a bit too much to ask. :-)
The training set probably consisted of black and white celebs, without representation of latino or south asian celebs to fill in the gradient in between.
https://imgur.com/a/vc5dz0L
AI magic.