The dataset is compiled from flickr portraits, and the model was trained to generate sampels from that distribution, so there will definitely be bias towards the sort of people whose portraits end up on flickr. Per the paper:

>The images were crawled from Flickr (thus inheriting all the biases of that website) and automatically aligned and cropped. Only images under permissive licenses were collected. Various automatic filters were used to prune the set, and finally Mechanical Turk allowed us to remove the occasional statues, paintings, or photos of photos.

It's still much more diverse than previous datasets (which used U.S. celebrity photos), but would require some additional work to match the actual world population distribution.