I have trained StyleGAN2 from scratch with a dataset of female portraits | Hacker News Reader

I have trained StyleGAN2 from scratch with a dataset of female portraits

github.com

20 pointsEvgeniyZh20 comments

When StyleGAN2 came out I played with the original NVidia training set for about a week - and when I say a week I mean I did nothing else, day and night, for a week - just exploring "seed space" to see what variety of faces would come out, trying to select for certain faces, etc... I had a grand old time with it, but what you discover in the end is that the models can not nearly generate every imaginable human face. For lack of better words I would say there was a certain "averageness" to the results. By the way beautiful faces tended towards Isabella Rossellini so apparently she has the "most average beautiful face" :) So for sure this generator has blown away everything that came before, but there is still plenty of room for improvement IMO. In the end what impressed me most was the hair. It does hair better than any other software that I know of.

Impressive. When I saw the guy is from Luxembourg, and the sultry visages of the images, I felt sure this had something to do with some of the porn behemoths there. Which I Googled and found were MindGeek or LiveJasmin.

I guess it is aimed at people already conversant with the tech lingo, but to the uninitiated it looks like a deep-fake related generator of attractive white women.

I mean, it kind of is, and that's okay.

I think deep fake means using face of some existing person, while the faces generated using this method does not exist.

It seems the training data was almost certainly built from adult actress portraits of which you can easily scrape endless amounts from, like any porn site.

We're going to have to work through some tough decisions as a society with this kind of technology developing... The likelihood in the coming years of child exploitation images being generated using similar techniques as posted here, I'd say is almost a sure bet.

It dovetails with the current news of Apple's client-side content scanning using perceptual hashes. I'm strongly against that, but with regards to opening Pandora's box with the situation above...It'd be interesting to hear peoples thoughts on what might be able to be done about these sorts of tech induced dilemmas. (Ban the tech? License it? Accept more and deeper blanket surveillance? Allow it all!?)

Click on the picture with the faces -when having a closer look it can be seen that every picture goes from "slightly distorted" to "disturbingly disfigured".

Seems the less accurate images occur at smaller training set sizes (at least that's how I read those image captions about kimg).

Seems that was just v1. There's a link to the 2nd iteration of the project, and the images, to my eyes, are very much improved with respect to realism.

https://github.com/l4rz/scaling-up-stylegan2

This one clearly has a tendency to create asymmetric eyes slightly beyond what my brain perceives as average asymmetry.

One day style gan will be used to recreate pornography and with all its weird anomalies the outcome will be truly horrifying.

Goodness knows there’s plenty of porno training images out there.

https://thisvaginadoesnotexist.com/about.html

> I was curious how it would work on the human anatomy

Mmm-hmmmmmmmm.

For AI, a female can only be white?

why are they all naked?

According to the readme, to avoid common defects related to clothing.

Ah yes. Defects relating to clothing. Such as "wearing clothing" and "not being naked".

I'll tell my wife about clothing based defects.

“Female portraits” is some chefs kiss whitewashing. More like “women under the male gaze”.

Reply on news.ycombinator.com