Quote from: Anny1234 on March 31, 2023, 16:23

Because following this logic all works created by people are also "derivative" then because artist creates them after he got inspired by a mix of many kinds of "copyrighted" things he has seen before! Nothing ever is born in the vacuum.

Wrong.

It's a long-standing law principle that copyright does not protect ideas, concepts, systems, or methods of doing something, while it differentiates all this from derivative work.

Obviously you are not a lawyer and you didn't do your homework.

but that's the whole point! AI work is not derivative in an rational sense & no one has yet given an example that shows such

...
Anthropomorphising the app is silly. It isn�t �learning� in the same way as a human. It�s taking and processing images via programming. Saying they don�t use them �directly� is a bizarre concept when we are talking about images being loaded into any type of algorithm or app resulting in �new� images coming out the other end.
...

MOST successful AI doesn't learn like a human but produces vastly superior results - champion GO & chess Ai dont even know the rules of the game!  it's called emergent behavir

Quote from: cascoly on April 01, 2023, 12:34

of course giving it a limited dataset can produce what may be derivative work

Exactly!

Let's take the worst case non-existent scenario and judge it all based on it.
Let's ban all the kitchen knives because very occasionally someone is got killed with them.

The talk was about legitimate companies that allow commercial use because they are sure such cases are too rare (and their goal is to bring them to zero eventually, it's been few months!), not some garage AI.

And even in that overblown unimaginable case with 1 input of data, it may not be creating a derivative, if you have to draw a tree there is no way you can draw it any other way than in a shape of a tree. And if you've seen only one tree ever then well what would you expect. But it doesn't copy-paste nor collage anything - this is not how it works at all.

Quote from: Lowls on March 31, 2023, 06:05

Derivative work is clearly what an AI is producing. in which case compensation isn't a nice thing to do it is a legal requirement. Adobe have not paid for any such license to use the work. They don't own the copywrite we do.
..

not so clear - it's been shown many times that the actual generative AI does not use any images directly but creates an entirely new image from the training set.   

whether anyone has the right to use images scraped from the web is a separate issue, more theoretical, since any payment to authors for the training would be  minuscule a tiny fraction of 100s of m
at the same time writers & journalists arent complaining about scraping for trillions image examined

dont  know why some artists continue to propagate misleading info especially since there's really no upside.
webscraping creates the dataset but the results of those GPT bots do not violate anyone's copyright, again because the AI generates completely  new text w/o using the training data

Well you are conflating two issues there.

firstly the data sets it uses are an unknown quantity so you and others are guessing. Assuming millions is a nonsense because it just doesnt need to. Not when it could use 5.It doesnt work like that. As others have stated and shown examples of that given a limited data set the image will be similar or because it isnt programmed to be innovative it will or rather cab directly duplicate it. I've seen many examples of A.I. generated images and they have created a clearly derived work. And another issue is that asking it to create a photo of a dog in a costume should generate a fairly generic photo. But users don't do that. They'll ask for a specific breed, in a specific costume in a specific environment which limits the data sets it can reference. Similarly ask it to create a dragonfly on some grasses and it will reference only a few images not all.
Your second example of text regarding ChatGBT4 producing unique text is not quite correct. This is new law in the making. plagiarism has a long standing legal framework but even though ChatGBt4 can create text with perimeters (write an essay about World War 2 as if I had written it) producing a worryingly me like essay. But again it has racesd ahead and is being abusesd but not so well publicised is that apps have been created to detect A.I. created text already. Because of the potential for plagiarism. Schools are trialing A.I. detection software as we speak using A.I. to find content produced by a.I.

It is a very new frontier across the spectrum of usage but shortly the law will catch up as it always does. And companies that played fast and loose will find themselves in a very uncomfortable position of compensation and fines. Which is why SS are getting ahead of the curve by at least making a show of being wholesome. Random payments for images used that they do not disclose and payments are extremely variable, which doesnt help with transparency but, they are acknowledging that compensation must be paid.

 Adobe will have been aware of this and I suspect have already stated "we are working on a model of compensation" because the A.I. output isn't so far from its source material as they were led to believe it would be so are playing catch up with compensation. Just a theory of course.

Researchers only extracted 94 direct matches and 109 perceptual near-matches out of 350,000 high-probability-of-memorization images they tested (a set of known duplicates in the 160 million-image dataset used to train Stable Diffusion), resulting in a roughly 0.03 percent memorization rate in this particular scenario.

Also, the researchers note that the "memorization" they've discovered is approximate since the AI model cannot produce identical byte-for-byte copies of the training images. By definition, Stable Diffusion cannot memorize large amounts of data because the size of the 160 million-image training dataset is many orders of magnitude larger than the 2GB Stable Diffusion AI model. That means any memorization that exists in the model is small, rare, and very difficult to accidentally extract.

None of this is written by me

@Lowls Go and write your insulting replies under this article: https://arstechnica.com/information-technology/2023/02/researchers-extract-training-images-from-stable-diffusion-but-its-difficult/amp/

Or maybe straight to a group of AI researchers from Google, DeepMind, UC Berkeley, Princeton, and ETH Zurich who released this paper, they are very interested in your opinion.

(Since I ignore you from now on, so won't see your little insulting replies anyway. )
Thanks to the forum for this feature.

Sorry this has got heated/ personal. There's no need for it.

Back on topic.

Or maybe straight to a group of AI researchers...

Possibly not the best people to trust on the topic as they have literally the most invested in it being legit and above board. Even then the paper produces examples of the AI copying images almost exactly. In fact a lot more exactly than if I just saved the exact same image in a more lossy file format.

Surprisingly nothing about derivatives and copy-pasting...

I mean is there nothing about it? Strip out the buzzwords and leading language it could mean almost anything. Here's a possible translation:

When training an image synthesis model, researchers feed millions of existing images into the model from a dataset, typically obtained from the public web.

When programming an AI app employees copy millions of our illustrations/ photos into the app, typically �taken� from our online portfolios where we posted them for completely unrelated reasons.

The model then compresses knowledge of each image into a series of statistical weights, which form the neural network.

The app then saves the images in a more compressed format and indexes them for easy retrieval when recompiling derivates.

This compressed knowledge is stored in a lower-dimensional representation called "latent space."

These compressed copies of the images are stored in a database.

Sampling from this latent space allows the model to generate new images with similar properties to those in the training data set.

Copying elements from the images in this format allows the app to create the derivative images using the properties of the original images compressed and indexed in its database.

OR:
When programming an AI app employees copy millions of our illustrations/ photos into the app, typically �taken� from our online portfolios where we posted them for completely unrelated reasons. The app then saves the images in a more compressed format and indexes them for easy retrieval when recompiling derivates. These compressed copies of the images are stored in a database. Copying elements from the images in this format allows the app to create the derivative images using the properties of the original images compressed and indexed in its database.

I am sure you can imagine writing a description of saving a jpeg as a tif and then zipping the file and making it sound like magic forming a new image if you used enough tech bro buzz words in a similar way.

What they don't realize is that they are only useful for this transition period, to select and check the images, they will be useless and discarded afterwards. The image generations will be autonomous, or directly driven by the final client. And this in a very short time, see you tomorrow!!! End of story.

Exactly! (though haven't seen anyone on this forum who doesn't realises it).

It is senseless to submit images to stock sites anymore (AI or not, maybe if only Editorial) as they themselves will cease to exist eventually, except Adobe as they have their own unique product. In my humble opinion (note for those who doesn't recognise it).

Some people here have predicted that it may happen within a year or even less.

Quote from: SVH on March 26, 2023, 11:39

Quote from: Her Ugliness on March 25, 2023, 02:29

Notice how even here, whenever Mat announces something new, he will always just reply to questions and comments stricly related to the technical aspect of a feature and ignore every single question and comment about user actual concerns or morality aspects of what Adobe was doing.

I can't recall that Adobe has ever seeked out conversation with contributors about anything, before we were getting presented the end result - Free galleries, currency changes, our images being used to train AI that will make our images worthless - there has never been any kind of "conversation" prior to annoucing the final decision and the "conversation" that took place after that were alwas one-sided with Adobe pretty much ignoring our concerns.

Very true. And if you accuse Mat of corporate speak then he is offended

I do think they listen in via Mat. I actually think that Adobe didn't want to prolong the bonus program but because here, and probably at other places as well, we made a fuzz about it and so they caved in, in the end. Not to lose face and they sure will have weighted the costs versus the potential reputation damage.

But it annoys me too that when you ask something else then a technical question you never hear an answer. Concerns are maybe listened to but certainly not replied to.

And since they are a company (maximum profit seeking) it annoys me that they pretend to be an NGO sometimes only to support artists at all times while in reality they are not. But I guess lot of people fall for that kind of talk. At least here, I see many of them. But you can also see them when you join some of their live Behance meetings. Some actually seem groupies and completely believe in Adobes fantasy tale

And remember there is no them without us but not exactly as how they mean it. It's more that if we don't buy their products or deliver content to them then there is no them. That part is definitely true.

Yup, I'm so glad somebody said that. I see it too and often feel like traitor if I dare to say something against Adobe. Because looks that many people think they really care for us.

You see.

Adobe (via Mat) started this thread but all difficult questions were not answered. Even after some time, answers will not come.

But they are here for you, don't forget

Adobe just take your images, feed a machine with it and generates a lot money.

Did you contributors get even a penny for such a usage?

Not sure I understand what you wrote but yes, we got paid, and yes it was pretty much pennies. I don't know how much we got per image, how many of mine, or what images were used. I may have missed that detail?

No - there's a statement that they'll work out a compensation model and share the details when Firefly exits beta (it's still in beta now)

https://helpx.adobe.com/stock/contributor/help/firefly-faq-for-adobe-stock-contributors.html

Aight, and thanks. I was just curious about it.

At least, Adobe plans some sort of compensation, unlike the rest of the industry which just stealing images to train their machines.

I myself currently stopped uploading stuff to agencies until it's more clear that my rights are respected (and my images will be safe) or i get paid at least an extended licence per image.

I decided this should be its own topic

https://www.microstockgroup.com/fotolia-com/googleadobe-partnership-with-bard-firefly/

This is only available in the beta version of Photoshop (and I have no idea how one gets that):

It's under "Beta Apps" in Creative Cloud Desktop App. Looks that it runs parallel with "normal" PS.