AI says this is me, riding the bus.
Those hands, tho. 😂
Discussion
Gave up trying to get the top buttons closed. Need to throw a shawl on that jpg. Oh well.
Those were the only 2 I could produce wearing some semblance of clothing and not having weird folds all over their skin.
What is with the weird quadruple clavicles?
i tell ya, these things are trained on porn
Probably.
If you ask for something curvy, most of them are very skinny, but with obviously fake boobs, or it goes straight to Jabba the Hut.
Darn near impossible to get it to produce Plump Mom pics, which is the most realistic set.
Goes straight from size 0 to size 25, when most women are in the 8-14 range.
Have you tried SDXL with the Pony Diffusion checkpoint?
😂 Pony, as in My Little Pony?
But yet ones that try hard to exclude porn can’t draw any human that isn’t a deformed sci-fi mutant.
Yeah, why is that? So weird.
i dunno, but i had an experience with a friend in highschool who was quite practised at drawing and his female characters were very masculine while mine were very much feminine
it is very likely that these engines are optimized for something other than accurate representations of common living things, i'm pretty sure the fingers problem exists no matter what animal you get them to do things to
My boyfriend in high school was really into manga. He made me one of the characters in the school comic.
Went on to design video game characters.
yeah my artist buddy in school was all about DC/Marvel artists and yes i was more interested in anime at the time
but before that i was learning to draw humans from a book teaching proper artists, my mother also was very good at it, she was a much better painter than me except when i was on meth and using indian ink and brush, then i was better than her lol
Ha ha. Just realized that I'm probably the basis for some video game characters someplace. LOL
Weird thought.
There are worse things to have inspired.
there is something tactile about the ability to render physical things accurately
i know i have extremely acute proprioception, you can turn the light off in a room i know and i can grab at things within a centimetre of exact no problem at all
my mother also was extremely agile, as a child notorious for skipping across the spine pieces of house roofs confidently, running across the top of solid fence posts
gah dammit i almost catched this one woman who turned out to be a really good 3d designer
Why did they do it? Fear of getting sued.
Why does it suck at drawing humans without porn? 🤷♂️ my guess is that they went overboard with scrubbing the training dataset of any human who isn’t wearing loose clothing out of fear that it would take little fine tuning to turn it into porn.
My second thought is... Longer hair are almost always better
Yeah, it refuses my short hair.
That awkward moment when you realize that there's a reason she always cut every selfie off at the clavicles.
#StellaFullLength
I've seen that before about hands, why can't AI manage them?
i have no idea. nostr:npub1m3xdppkd0njmrqe2ma8a6ys39zvgp5k8u22mev8xsnqp4nh80srqhqa5sf do you know?
I think its that there's less a concept of 'object' in an image, its correlations from one pixel to the next. Because fingers and other objects can be in so many positions and in various contexts its difficult for the model to 'segment' out that object like a torso or facial features which have more stability across images.
That makes sense, thanks
I keeps giving you the super-plus white person face 🤣
Helps preserve anonymity, perhaps? 🤷♂️
It does. 😂
I think the AI gets confused because of the eyes. As far as I know, they categorize by race and then fill in from there, so I tend to get lumped in with Scandinavians and then the rest of my face is just confused.
They go the opposite way on the lips. They usually give me these pouty goose lips. And they take off my freckles. 😭 I sometimes add them back on, cuz I take that personally.
It looks vaguely like me, but less 🌹, more 🍯. In reality, I'm less porn star, more kindergarden teacher.
For comparison
It's funny how subtle, but consistent the difference is.
It is pretty funny. It exposes something about the training data and the internal model weights the ML algorithms are working with.
I bet a research paper could be written on AI image generation around the edge cases by exploring how those models respond to various mixed-race individuals.
Yes, would be interesting to know how that all works.