Quote from: GeekyBugle on March 22, 2023, 06:38:09 PM
I'm going to say something unpopular (nothing new on my end):

Unless you can prove a particular AI is plagiarizing artists IDGAFF if it's "ethical" or not.

I'm not sure what you mean by "plagiarizing".

What I see artists complain about is the AI being trained by using their art. I don't see it much different than a human copying other people's art to learn how it works. But then, I'm also not much of an artist. I see this as different from creating art and marketing it as by some artist when an AI did it. Something I haven't actually seen people do -- even though I'm sure someone has.

It's easy to determine what was using in training since the big AI projects all have their training datasets online. And they did pretty much scrape the internet for any available art for training.

Where I think the copyright office got it wrong (they said AI art can't be copyrighted, only configurations of AI art) is that getting the prompts right to create the image you want is a lot of original creative work. It isn't "just random". You should see some of the huge positive and negative prompt lists people use to generate particular pictures. Yes, a beginner can brute force it, but a lot of the experts don't.

Quote from: Tod13 on March 23, 2023, 04:01:04 PM
It's easy to determine what was using in training since the big AI projects all have their training datasets online. And they did pretty much scrape the internet for any available art for training.

You'd think so, except AI is also trained by AI images. For example, Adobe is making a big deal that all their Firefly training is on "copyright respected" material they own via Adobe Stock. Except artists have been using Midjourney and SD to create Adobe Stock art:



AI training AI is also about to get a lot more common. Stanford just released Alpaca which used GPT-3 to train their model up to being about as good as that model. GPT-3 was like $4 million to train, training up Llama using GPT-3 took $600 and a few hours.

Quote from: Dracones on March 23, 2023, 04:31:20 PM

Quote from: Tod13 on March 23, 2023, 04:01:04 PM
It's easy to determine what was using in training since the big AI projects all have their training datasets online. And they did pretty much scrape the internet for any available art for training.

You'd think so, except AI is also trained by AI images. For example, Adobe is making a big deal that all their Firefly training is on "copyright respected" material they own via Adobe Stock. Except artists have been using Midjourney and SD to create Adobe Stock art:



AI training AI is also about to get a lot more common. Stanford just released Alpaca which used GPT-3 to train their model up to being about as good as that model. GPT-3 was like $4 million to train, training up Llama using GPT-3 took $600 and a few hours.

Cool. I hadn't seen that. I'd be more impressed if they actually posted the database, since "openly licensed content" is kind of vague. *shrugs* I figure anything you post on the internet will get used for whatever they finder decides to do with it anyway.

The Alpaca web site has been taken down for, among other reasons, toxicity (I guess it's not properly censored to spew only leftard talking points. https://www.theregister.com/2023/03/21/stanford_ai_alpaca_taken_offline/

But do not despair, the whole thing can be run on your PC Or so this article claims) https://medium.com/geekculture/writing-a-medium-article-using-ai-stanford-alpaca-running-on-your-local-pc-e025416a032a

In case you want to download and install it has the instructions in an easy to follow format.

Here's the repo: https://github.com/tatsu-lab/stanford_alpaca

This one claims to have a "cleaned" Data set: https://github.com/gururise/AlpacaDataCleaned

I already downloaded it and will try to run it locally.

I painstakingly drew all 45 students for my latest RPG module, one at a time, one per day, and while AI had nothing to do with it, I look back on my efforts and am really proud of what i did.  It's no DaVinci, but the sweat that went into doing it myself makes it all look glorious in my eyes.

Quote from: Cathode Ray on March 23, 2023, 09:09:40 PM
I painstakingly drew all 45 students for my latest RPG module, one at a time, one per day, and while AI had nothing to do with it, I look back on my efforts and am really proud of what i did.  It's no DaVinci, but the sweat that went into doing it myself makes it all look glorious in my eyes.

I would pay for this as I need NPC portraits for my projects.

Quote from: GeekyBugle on March 22, 2023, 06:38:09 PM
I'm going to say something unpopular (nothing new on my end):

Unless you can prove a particular AI is plagiarizing artists IDGAFF if it's "ethical" or not.

Furthermore I am fully onboard, for someone like me who doesn't have the money to hire artists it's a valid option over using the same public domain art everybody else has already used. It's going to empower people to produce higher quality products.

We have caught a couple now. Some are way easier to spot than others.

Af for art theft. These things are very NOT using only public domain art. They are pulling from DA and everything else they can find.

Quote from: jeff37923 on March 24, 2023, 06:10:19 AM

Quote from: Cathode Ray on March 23, 2023, 09:09:40 PM
I painstakingly drew all 45 students for my latest RPG module, one at a time, one per day, and while AI had nothing to do with it, I look back on my efforts and am really proud of what i did.  It's no DaVinci, but the sweat that went into doing it myself makes it all look glorious in my eyes.

I would pay for this as I need NPC portraits for my projects.

I am flattared!  However, these NPCs are for my Radical High module's character profiles, and all belong to my own RPG. 

Never mind. Read wrong post together. Sorry.

Apologies Omega. All I can say is it is late in the work week, and my brain is fried.

I've been using Stable Diffusion for some time. The argument that it infringes on artworks it was trained on is...wrong. I can experimentally prove it, and you can replicate the experiment I used to do so if you doubt me.

I took a random image post with workflow off the Stable Diffusion reddit, copied the information on the model, prompt, and seed, and told my 3090 to make 40 copies of that exact image. Not only were the images not identical to the source image, but it's questionable if any of them were infringeably similar to each other.

THIS was the original: https://www.reddit.com/r/StableDiffusion/comments/100tp0v/protogenx34_has_absolutely_amazing_detail/

This is my output batch: https://drive.google.com/file/d/1H3EkD1DQ7kToEyHspbLf6a--vOYJ730m/view?usp=share_link

And remember, this was a batch of 40 images with the same model, prompt, and seed. This is as close to generating identical images as Stable Diffusion can manage. My point is that AI generated art is not Minecraft. Identical seeds and prompts do not necessarily produce identical results. Not all models are like Protogen here and generate different images from the same seed, but many are. And, of course, if you're talking recreating an image the AI was trained on, you would expect at least this much inconsistency; generated images must start from a noise seed, but training images created by a human by definition do not have a seed.

Incidents like recreating Iranian Girl, Green Eyes are almost certainly the product of cloud services trying to shave their hardware rental costs by scraping images off a google search and putting it through an Image to Image filter, which is a completely different method of image generation and isn't guaranteed to produce non-infringing results.

This puts the art community in a situation where it's basically impossible to tell if an artwork was made or enhanced via AI and it's impossible to tell what artworks went into training an AI without the creators self-incriminating.

At the moment the US Copyright Office says that AI artworks are not copyrightable, but exactly what that means is up for some debate. A lot of artists use AI to enhance their workflows so where human art ends and AI generated begins is quite a gray area. I think that this isn't really a thought-through position; it's more an attempt to scare businesses away from AI by saying they have no copyright protections. But if it stands in court, that would mean you can use a private AI model and the artworks it generates are copyright-free, meaning you can train another AI with them and the second AI can be used for commercial purposes provided you don't care about defending your copyright. Because the artworks it was trained on are not copyrightable.

Like I said, not a terribly well thought through policy.

Quote from: Omega on March 24, 2023, 07:16:59 AM

Quote from: GeekyBugle on March 22, 2023, 06:38:09 PM
I'm going to say something unpopular (nothing new on my end):

Unless you can prove a particular AI is plagiarizing artists IDGAFF if it's "ethical" or not.

Furthermore I am fully onboard, for someone like me who doesn't have the money to hire artists it's a valid option over using the same public domain art everybody else has already used. It's going to empower people to produce higher quality products.

We have caught a couple now. Some are way easier to spot than others.

Af for art theft. These things are very NOT using only public domain art. They are pulling from DA and everything else they can find.

By that metric, nearly all art is theft. Everyone is copying I'm sorry "taking inspiration from" someone else.

Quote from: Corolinth on March 26, 2023, 10:26:49 AMBy that metric, nearly all art is theft. Everyone is copying I'm sorry "taking inspiration from" someone else.

Correct.

"Art is theft." -- Pablo Picasso

"Nothing is original. Steal from anywhere that resonates with inspiration or fuels your imagination." -- Jim Jarmusch

"Art is too important not to share." -- Romero Britto

"You've got to embrace the future. You can whine about it, but you've got to embrace it." -- Matt Groening

Quote from: zircher on March 26, 2023, 01:50:36 PM"You've got to embrace the future. You can whine about it, but you've got to embrace it." -- Matt Groening

"Also, tune in for season 34 of The Simpsons!" - Matt Groening