this post was submitted on 26 Nov 2023
420 points (90.9% liked)
Technology
59575 readers
3040 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I'm genuinely sceptical. How do they ensure the same looking person is generated each time? From any perspective? You can create fake images of a specific person precisely because you have a dataset of ground truth images.
If it is true... Then yeah. Modelling is now a dead job. And weirdly we're back to pre-photo advertising when everything was just drawn.
You can generate consistent faces simply by using random non-existent names. Which in turn you can use to train a custom LORA with Dreambooth (needs about 20 images) or use ROOP (a single one can be enough). And of course you can just mix and match it as you please, mix multiple real faces together into a new one, use dedicated face generators like https://thispersondoesnotexist.com/ and so on.
This barely even takes effort anymore, e.g. quick ROOP FaceSwap with the photo from the article, which doesn't look quite the same due to only a single input image, gets better with more, but that's just shows how easy it is to generate a new face, which will than be consistent with itself.
The hard part is getting an interesting pose, expression and haircut into this, as well as sponsored products and stuff. Generating realistic images with AI is pretty easy, but getting variation into them so they don't end up all looking the same can get tricky.
Edit: Five more minutes of effort and it starts to look a little closer.
I'm guessing they just generate a bunch of pictures, pick the closest and fix the rest in photoshop.
Not like real models aren't already often photoshopped to (near) unrecognizability.
I'm pretty sure that audience doesn't care if it's a little off ..
Seed numbers bruh
It still doesn't generate the same looking person every time it's just the same kind of style.
You definitely can. Ie, generate 100 pictures and pick the ones that are very similar. Use those to train the concept of "ai lady XYZ" and then generate more and train more.
Keep repeating until the concept "ai lady xyz" is unique and self-consistent.
Mmm.... recursive AI models, nearly as tasty as the recursiveness of the filling of a KitKat.