Hi, author here - that's a great point. When I first saw those results and how inaccurate they were, I thought there was a chance it was returning me an overfitted actual input image from training. Most likely not, but they were so realistic (and I was used to just seeing llamas until this point), that I thought I'd play it safe.

Also, I came across this article which suggests that at some point users were not allowed to share images generating human faces, artificial or not: https://mixed-news.com/en/openais-dall-e-2-may-now-generate-...

They’re being used to create fake profile pictures.

I'm not sure why anyone bothers. StyleGAN2 profile photos are literally all over social media and they're good enough to fool the human reviewers every time I report them.