Ridiculed Stable Diffusion 3 release excels at AI-generated body horror

markgo

Ars Praefectus
3,922
Subscriptor++
censoring adult content that contains nudity also severely hampers an AI model's ability to generate accurate human anatomy

Perhaps human artists—who have trained on nude models for centuries even when the society they lived in was uptight about nudity—could have told them that.

Edit: Ninja’d by someone whose name sounds like it belongs to a great classical artist, Bernardo Verda.
 
Upvote
185 (186 / -1)
Perhaps human artists—who have trained on nude models for centuries even when the society they lived in was uptight about nudity—could have told them that.
Well they'd have to hire artists to know that. Their whole business is viewing artists with contempt and putting them out of work.
 
Upvote
68 (100 / -32)

Timboman

Smack-Fu Master, in training
64
It is still ENDLESSLY hilarious to me that the best image generation models out there right now are based upon Pony Diffusion XL, a model that was EXPLICITLY finetuned to generate furry porn.

Just goes to show that having a very broad, varied, and well tagged dataset is essential for a model to give good results. Even if a large chunk of that dataset is of various Pokemon going at it all hot and heavy, lol.
 
Upvote
155 (158 / -3)

thrillgore

Ars Praefectus
4,116
Subscriptor
200w.gif

My reaction going through some of these images.

Are we sure this isn't another attempt to poison the dataset? Because that's how I'd do it.
 
Upvote
17 (25 / -8)
Why would they not train on nude data? As long as the models have consented I see no issue.
Maybe there should be a ‘safety check’ to see if NSFW is generated when not requested, but that should not be done by removal of parts of of the training data.
Pretty sure nobody consented to their images being used in these training datasets, not the models, not the owners of the copyright on the images...
 
Upvote
35 (47 / -12)
It's really difficult to believe that when you take an AI that has no sense of ethics, morality, consequence or politick, and try to stick your finger on its scales to enforce an ill-defined moral bias held by its human overseers, that you would end up with sub-optimal results! Gosh, I can't even fathom why most attempts to make neural networks conform to values that aren't even close to universally-held in their training data isn't working out!
 
Upvote
26 (42 / -16)

Edgar Allan Esquire

Ars Praefectus
3,101
Subscriptor
Without nipples as landmarks, the whole body goes to hell. Why else do you think men still have them?
Fighting anime always remind me of an interview with a manga artist that worked on both Shonen and yaoi titles. They'd asked her what the main difference between styles was. "In BL you draw the men with nipples," was her response. You can't unsee the nippless pecs in most Shonen after that.
 
Upvote
63 (63 / 0)

lewax00

Ars Legatus Legionis
17,402
Nice try Stable Diffusion, but I'm not falling for "a man showing his hands" - that's clearly a Rakshasa taking human form.

(Edit: for context since it's maybe a little buried: "However, their most unnerving and unique feature was their reversed hands." - that's generally their distinguishing feature that can give them away to the perceptive.)
 
Upvote
33 (33 / 0)

markgo

Ars Praefectus
3,922
Subscriptor++
...but doesn't account for not understanding left and right, nor knowing how many fingers are on a hand.
It’s a little startling to note than some of the more common and smaller AI anatomical errors (extra finger or malformed joint) are not that dissimilar to genetic defects.

It’s either a coincidence or a deeper truth on how complex things are constructed and how that process can go wrong.
 
Upvote
30 (32 / -2)