AI Image Generation

Chris

Potato del Grande
18,237
-312
The copyrighted material the model is trained on is analagous to a person looking at and remembering something.

Requiring it to not be used is analagous to artists not being able to set foot in art galleries in case something inspires them.

Anyway I want the work of autistic men to crush the hopes and dreams of people who shunned autistic men.
 

Palum

what Suineg set it to
23,448
33,708
The copyrighted material the model is trained on is analagous to a person looking at and remembering something.

Requiring it to not be used is analagous to artists not being able to set foot in art galleries in case something inspires them.

Anyway I want the work of autistic men to crush the hopes and dreams of people who shunned autistic men.
Uh no. There's many many decades of jurisprudence around the concept of transformative work or fair use. What good AI tools can do today is more equivalent to counterfeiting than that. The likely outcome is lots more work needing to be done in XAI to properly attribute training sources and to develop tools that will analyze significant inputs and outputs to make sure they are sufficiently divergent.
 
  • 1Like
  • 1Dislike
Reactions: 1 users

Aldarion

Egg Nazi
8,946
24,469
Every time this comes up people just dive right into the technical details or the legal details, entirely missing the relevant points.

All art is derivative.

People dont have ideas; ideas have people.

The notion of "stealing an idea" (or an image, or a song) is fundamentally dishonest, a pretense made by people trying to make a buck off the argument. Anyone whos done even a few moments of honest introspection knows this is all nonsense. This "your picture looks like mine therefore you stole my picture" stuff is always bullshit made up by people who don't understand how human thought works, whether its a machine or a starving artist doing the ""stealing"".

In a vain effort to head off irrelevant arguments, I'll add this edit: the process defines whether its plagiarism or not, not any similarity in the final product.
 
  • 1Like
Reactions: 1 user

Mist

Eeyore Enthusiast
<Gold Donor>
30,414
22,202
All art is derivative.

People dont have ideas; ideas have people.

The notion of "stealing an idea" (or an image, or a song) is fundamentally dishonest, a pretense made by people trying to make a buck off the argument. Anyone whos done even a few moments of honest introspection knows this is all nonsense. This "your picture looks like mine therefore you stole my picture" stuff is always bullshit made up by people who don't understand how human thought works, whether its a machine or a starving artist doing the ""stealing"".

In a vain effort to head off irrelevant arguments, I'll add this edit: the process defines whether its plagiarism or not, not any similarity in the final product.
You're missing the point. This has nothing to do with arguments about human creativity or derivative art. The training process literally compresses and stores copies of all the training data, aka stolen content in the case of copywritten content, using what is effectively computing's funkiest pseudorandom lossy compression algorithm.

"Generative AI" in its current form is like saying "if you used Pied Piper's middle-out compression and a bag of d20s to make slightly shittier copies of every album and swapped the songs around between albums, you could do Napster legally without having to pay any of the people who made the music." A bonkers nonsense argument. Most people in the AI research community aren't even trying to defend it at this point, they know what they built.
The copyrighted material the model is trained on is analagous to a person looking at and remembering something.

Requiring it to not be used is analagous to artists not being able to set foot in art galleries in case something inspires them.

Anyway I want the work of autistic men to crush the hopes and dreams of people who shunned autistic men.
We know for absolute certain that this isn't how the human brain works. Your brain does not store everything it sees. In fact, human memory absolutely, demonstrably sucks. At the most basic level, your brain generalizes a LOT from very little data. A transformer neural network model generalizes a little bit, from a LOT of data. The difference is massive.
 

Captain Suave

Caesar si viveret, ad remum dareris.
4,774
8,067
You're missing the point. This has nothing to do with arguments about human creativity or derivative art. The training process literally compresses and stores copies of all the training data, aka stolen content in the case of copywritten content, using what is effectively computing's funkiest pseudorandom lossy compression algorithm.

You're completely misconstruing the legal side of the argument. Copyright law is about ideas and works. Compressing data (which isn't really what's happening in GPTs, it's more nuanced than that) is not a work and does not infringe. Only the works produced by the model could be infringing. Some of them certainly seem to infringe, but many are different enough from any single element of the training set that they likely do not.

We're in fundamentally new territory for the law, and the forthcoming jurisprudence will decide how these tools are built in the future.

"Generative AI" in its current form is like saying "if you used Pied Piper's middle-out compression and a bag of d20s to make slightly shittier copies of every album and swapped the songs around between albums, you could do Napster legally without having to pay any of the people who made the music." A bonkers nonsense argument. Most people in the AI research community aren't even trying to defend it at this point, they know what they built.

This is a deeply shitty summarization of how neural nets work.
 
Last edited:
  • 1Like
Reactions: 1 user

Mist

Eeyore Enthusiast
<Gold Donor>
30,414
22,202
Compressing data (which isn't really what's happening in GPTs, it's more nuanced than that)
Facts and experiments are demonstrating that it isn't more nuanced than that.

The training phase is effectively using gradient descent to find the perfect compression algorithm for the provided training set.
 

Captain Suave

Caesar si viveret, ad remum dareris.
4,774
8,067
Facts and experiments are demonstrating that it isn't more nuanced than that.

The training phase is effectively using gradient descent to find the perfect compression algorithm for the provided training set.

It's not compression in the sense that we typically use the term in computing. The neural net is tuned to fundamental relationships in latent space between the training set features. Yes, with the right prompts it can reproduce specific items of the training set (with loss, but close). But it's not JUST compressing the input data. The whole point of the neural net is that it learns to interpolate between points in training set in a way that approaches real understanding.

I can't use 7zip to compress pictures of people and pictures of animals and then get it to successfully produce pictures of animals wearing people's clothing, which GPTs can do.

Also, I could concede this entire point and it still wouldn't matter when it comes to copyright law.
 
Last edited:
  • 1Like
Reactions: 1 user

Mist

Eeyore Enthusiast
<Gold Donor>
30,414
22,202

1704166909517.png


lol

This ain't gonna look good in court.
 
  • 4Worf
  • 1Like
Reactions: 4 users

Kharzette

Watcher of Overs
4,931
3,579
I've been trying to make my high elf eq character for a few days. I've got a 4 gig gpu on my server machine and just remote in and fiddle with it 2 or 3 times a day and let it run. I've got the standard midrange cleric gearset for Quarm; rubicite and adamantite epaulets.

Code:
beautiful Nordic girl, large detailed eyes, cute pointy ears, blonde hair, red platemail breastplate, crystalline epaulets, red helmet, smile, perfect teeth, bright eyes, snowy forest

This is getting me the best result after days of experimentation. I can never quite get all of that to come together though.

It didn't know what adamantine was so I started using diamond, then crystalline. Diamond tended to create gems everywhere else but the shoulders.

This and seemingly every model I've ever used is fascinated by collars. I have them in the negative but there they are every single time.

Getting actual red armor out of it is rare. Sometimes just a piece or two. After several thousand iterations I think I have 2 or 3 rubicitey looking pieces.

Occasionally a horse shows up in the images for no apparent reason. The character itself looks more half than whole elf but I'm happy with it, though sometimes you get odd subdivided teeth. The eyes are creepy enough to look elfy.

Here are some highlights:

Decent armor, cool earring that might be a fishbone if you look at it crosseyed? :emoji_laughing: No helmet though.
00299-819479722.png


A decent rendition of the epaulets, but only on one side and no breastplate or helm:
00063-1848764575.png


Same for this one but with earrings that make me look like I could actually afford to buy my spells.
00132-1848764644.png


At some point I finally coaxed it into giving me a few helmets. Everything else about the image is wrong though:
00042-4189273003.png


Very tamriel elf eyes:
00064-1758482042.png


This one is probably the closest to having everything in the prompt, though it looks goofy:
00079-1758482057.png


A fave, though little about it is correct:
00067-1758482045.png
 

Mist

Eeyore Enthusiast
<Gold Donor>
30,414
22,202
So, a bunch of hyperwoke polyamorous SF weirdos stole everyone's art and writing, mixed in some of that special Podesta Pizza Sauce, put it all in a giant GPU-powered blender, threw up a barebones UI to get useful results back out of it, and people here are still gonna defend it because they want free stuff without having to work for it.

This is hypocrisy.
 

Control

Ahn'Qiraj Raider
2,246
5,640
So, a bunch of hyperwoke polyamorous SF weirdos stole everyone's art and writing, mixed in some of that special Podesta Pizza Sauce, put it all in a giant GPU-powered blender, threw up a barebones UI to get useful results back out of it, and people here are still gonna defend it because they want free stuff without having to work for it.

This is hypocrisy.
Artists vote communist. It would be hypocrisy not to steal share art.
1704332088749.png

1704332116436.png
1704332137306.png
 
  • 1Like
  • 1Worf
Reactions: 1 users

Tholan

Blackwing Lair Raider
783
1,486
penny arcade really went from ugly, to ok, to fucking ugly again
 
  • 2Like
Reactions: 1 users