Conversation

Notices

1. Cyber Yuki (yuki2501@hackers.town)'s status on Thursday, 16-Feb-2023 14:23:15 JST Cyber Yuki

   Okay, you want to see something funny?

   A lot of portraits of women generated by Stable Diffusion look like... check this out...

   Famous Hollywood actresses or popular video game characters.

   I can see Angelina Jolie, Cameron Diaz, Ciri from the Witcher, and oh my god is that Lady Gaga? :blobfacepalm:

   Curiously, and I've been searching through several pages in the Lexica webpage (it's a Stable Diffusion search engine), I haven't found any single portrait that looks like... say... Avril Lavigne, or Rose Leslie, or Ally Sheedy, or... any other non-conventionally-beautiful girls (which you CAN find in DDG search, for example). Nope. It's always the popular beauty queens that appear in Stable Diffusion generated images. The same faces over and over and over and over.

   It's almost like they're clones of each other.

   Okay, how did this happen?

   Here's my hypothesis:

   1) Hollywood tends to prefer women that seem conventionally attractive. Variability be damned, only pick the hottest girls.

   2) Artists use the hot Hollywood girls for their studies. Or, you know, anime faces.

   3) The artists' studies get sucked in by the opt-out model of Stable Diffusion

   4) Suddenly everything Stable Diffusion does is recycling the same faces over and over (and over and over...)

   Until iamge generation AIs start to learn actual human anatomy and introduce variability into their models, the whole thing is doomed to be dull and repetitive.

   I really, REALLY hope I'm not the first one to notice.

   #StableDiffusion #AI

   • Cyber Yuki (yuki2501@hackers.town)'s status on Thursday, 16-Feb-2023 15:27:51 JST Cyber Yuki

     in reply to

     If we're really going to go the AI route, we first need to build a database of variable human faces of all races, ethnicities and variations, looking front, sideways, and 3/4s, with annotated metadata for the chin, nose, eye points, etc. so that 3d models can be obtained from them.

     And I'm not talking about a few hundred. I mean at least ten thousand (and I'm think I'm still underestimating the quantity needed).

     Perhaps a hundred thousand sample human faces (and why not, bodies?) would do the trick. Pay the models a good compensation for having their faces (and / or bodies) get in the database, and voilá. You can start generating images now.

     There's no other way. You want real faces? Scan real humans, but scientifically and with enough statistical variation. From ALL countries.

   • Cyber Yuki (yuki2501@hackers.town)'s status on Thursday, 16-Feb-2023 15:27:54 JST Cyber Yuki

     in reply to

     More search results for your enjoyment: "Female portrait."

     God, I wish I was fucking kidding.

     Adrian Cochrane repeated this.
   • Cyber Yuki (yuki2501@hackers.town)'s status on Thursday, 16-Feb-2023 15:27:56 JST Cyber Yuki

     in reply to

     Fun fact: Search for "punk girl" and you'll get tons of CYBERPUNK characters :blobrofl:

     How do you EXCLUDE cyber from there? I even tried "punk woman but not cyberpunk" and it didn't change anything (I'm dying)

   • Cyber Yuki (yuki2501@hackers.town)'s status on Thursday, 16-Feb-2023 15:27:57 JST Cyber Yuki

     in reply to

     Angelina Jolie is fucking everywhere in that webpage and I can't unsee it 😭