Recently there was a livestream with Mat Hayward and Terry White (https://www.behance.net/videos/eb2dab18-ad9b-4535-a64e-127037055b8e/New-Creative-Cloud-features-for-Adobe-Stock-Contributors-with-Mat-Hayward-and-Terry-White) where it was revealed that customers have a certain number of "credits" they can spend on generative fill features. Namely, they called this "fast lane", which puts your request in a separate queue and you get your output in 10-15 seconds. Otherwise, you go into the other lane, which is slower, and which could take a lot longer to do the generative fill.

When customers used up their credits, they can buy new ones, and use it to put their requests in the "fast lane". Now - I might be biased, but it seems pretty obvious with me that a part of this revenue should be shared with us - the contributors who helped train the underlying model. Yes - we got that one-time payment to use our assets to create the model, but that is separate from continuously reaping the rewards of our work, and not sharing any revenue with us. So, I think a percfentage of the revenue should be shared with us.

I asked Mat about that in the topic about the livestream (https://www.microstockgroup.com/fotolia-com/adobe-stock-livestream-today-with-mat-hayward-and-terry-white/msg594091/#msg594091), but although the questions posed after mine were anwered, mine was ignored. I'm no mind reader, but it's probably because there isn't any revenue to share with us.

Do you think that we, the contributors whose content trained their model, should be fairly compensated for the continuous revenue that Adobe will get from selling these (but not limited to!) fast lane credits? Is a one-time payment enough to keep you quiet? Or will you raise your voices only when AI completely takes over and you lose the entierty of your revenue? Let me know in the comments <3

....
b) When an AI "image" generates an image - it does reference essentially computer models. However - it is totally easy to say for say a "car" - 55 contributors were "tagged" in that computer representation of a "car".
c) When the asset is generated, contributors get micropayments (i.e., say the "ai" image was "worth" $0.10 to the company, and the revenue split was 50-50. So 55 contributors each get $0.05 / 55 ~ 0.0005 cents. May not seem like a "lot" - but with the millions of images generated daily - quickly adds up. (So say each image was like that, 55 contributors for different 'models' - and say 100000 images were generated referencing their input in a month, then 0.0005 * 100000 = $50)...

besides the fact that there is no, way to trace which images were used & worse even if your 'easy to do' way of marking were possible, for most images (maybe close to 0%) there is no way to identify who the artist is for the billions of images used - many have no names assoc'd and those that do lack verification and an address to pay to). how would your revised training know who (& how) to make payments

but you've stacked the odds in your favor - you claim there would be millions of images generated using only 50 originals, when when many thousands of images would be likely ( a conservative estimate), so your estimate of payment dueis off by several orders of magnitude and certainly not millions of images generated daily

Quote from: SuperPhoto on October 22, 2023, 11:24

....
b) When an AI "image" generates an image - it does reference essentially computer models. However - it is totally easy to say for say a "car" - 55 contributors were "tagged" in that computer representation of a "car".
c) When the asset is generated, contributors get micropayments (i.e., say the "ai" image was "worth" $0.10 to the company, and the revenue split was 50-50. So 55 contributors each get $0.05 / 55 ~ 0.0005 cents. May not seem like a "lot" - but with the millions of images generated daily - quickly adds up. (So say each image was like that, 55 contributors for different 'models' - and say 100000 images were generated referencing their input in a month, then 0.0005 * 100000 = $50)...

besides the fact that there is no, way to trace which images were used & worse even if your 'easy to do' way of marking were possible, for most images (maybe close to 0%) there is no way to identify who the artist is for the billions of images used - many have no names assoc'd and those that do lack verification and an address to pay to). how would your revised training know who (& how) to make payments

but you've stacked the odds in your favor - you claim there would be millions of images generated using only 50 originals, when when many thousands of images would be likely ( a conservative estimate), so your estimate of payment dueis off by several orders of magnitude and certainly not millions of images generated daily

You obviously don't have any computer background, or very little. Do you have ANY programming experience whatsoever, let alone large datasets? Yes, it is VERY possible, and VERY doable.

The illustration/example is designed to keep it simple, so you can understand. Obviously programming would be a little more sophisticated than that. But it is VERY VERY easy to do - it is simply a matter of DOING it.

besides the fact that there is no, way to trace which images were used & worse even if your 'easy to do' way of marking were possible, for most images (maybe close to 0%) there is no way to identify who the artist is for the billions of images used - many have no names assoc'd and those that do lack verification and an address to pay to). how would your revised training know who (& how) to make payments

We're talking about Adobe generative fill. Adobe knows very well where to find the artists who's photos were used and how to pay them as they used images from their own database.

Quote from: cascoly on October 22, 2023, 12:42

Quote from: SuperPhoto on October 22, 2023, 11:24

....
b) When an AI "image" generates an image - it does reference essentially computer models. However - it is totally easy to say for say a "car" - 55 contributors were "tagged" in that computer representation of a "car".
c) When the asset is generated, contributors get micropayments (i.e., say the "ai" image was "worth" $0.10 to the company, and the revenue split was 50-50. So 55 contributors each get $0.05 / 55 ~ 0.0005 cents. May not seem like a "lot" - but with the millions of images generated daily - quickly adds up. (So say each image was like that, 55 contributors for different 'models' - and say 100000 images were generated referencing their input in a month, then 0.0005 * 100000 = $50)...

besides the fact that there is no, way to trace which images were used & worse even if your 'easy to do' way of marking were possible, for most images (maybe close to 0%) there is no way to identify who the artist is for the billions of images used - many have no names assoc'd and those that do lack verification and an address to pay to). how would your revised training know who (& how) to make payments

but you've stacked the odds in your favor - you claim there would be millions of images generated using only 50 originals, when when many thousands of images would be likely ( a conservative estimate), so your estimate of payment dueis off by several orders of magnitude and certainly not millions of images generated daily

You obviously don't have any computer background, or very little. Do you have ANY programming experience whatsoever, let alone large datasets? Yes, it is VERY possible, and VERY doable.

The illustration/example is designed to keep it simple, so you can understand. Obviously programming would be a little more sophisticated than that. But it is VERY VERY easy to do - it is simply a matter of DOING it.

you know NOTHING about my computer experience.

you don't address the biggest problem - how do you identify who made the image & how to contact & pay them?  that information isnt available in most cases

and now you've changed the goal posts, saying funds should be distributed to everyone, not based on where their images were used -  you know little/nothing about ML it seems and that is relevant as  the major fallacy in your proposal assumes you can track where an image is used

you know NOTHING about my computer experience.

you don't address the biggest problem - how do you identify who made the image & how to contact & pay them?  that information isnt available in most cases

and now you've changed the goal posts, saying funds should be distributed to everyone, not based on where their images were used -  you know little/nothing about ML it seems and that is relevant as  the major fallacy in your proposal assumes you can track where an image is used

You are correct, I don't know your computer experience, which is why I was asking. Likewise, you know absolutely nothing about my experience, and are quite arrogant and accusatory in your statements, which would lead me to believe you know very little, if anything related to computer science/computer engineering/programming/etc - and especially - actual "AI" versus "ML" algorithms. Fact is, I actually do know what I am talking about. Have you EVER dealt with large datasets, scraped data, creating actual databases from scratch, written image algorithms, ANY of that? Have you actually ever even created your own "AI" algorithm? If you had - you'd probably know what I am talking about is very feasible, and simply a matter of doing it. It does require "work", and it does seem agencies are trying to figure out how to cut out artists while appearing to be noble (in many ways from a purely greedy standpoint). when fact is, they should be held fully accountable for any theft of artist assets, and compensate accordingly - in the same perpetual/recurring revenue model they so desperately desire.

Answering your questions:

a) For compensation - depends on where/how the data/images/etc were scraped. It would make most sense to simply begin with the major agencies whose data was scraped (i.e., DT/SS/AS/P5/etc). Super easy to figure out who to issue payments to. For other agencies, would be a matter of writing more sophisticated algorithms...

And - as a "tongue in cheek" statement - since "everyone" seems to think actual thinking "AI" exists (most people don't actually understand what actual "ai", and believe it is "thinking", versus the sophisticated theft & pattern re-arrangement being called "ai") - but if "ai" actually existed - it would be super simple - just ask the "ai" to figure out who to issue payments to.

b) it is SUPER easy to track/source image usage. If you don't realize this, it seems you've never scraped data before?

Quote from: Her Ugliness on October 23, 2023, 02:22

Quote from: cascoly on October 22, 2023, 12:42

besides the fact that there is no, way to trace which images were used & worse even if your 'easy to do' way of marking were possible, for most images (maybe close to 0%) there is no way to identify who the artist is for the billions of images used - many have no names assoc'd and those that do lack verification and an address to pay to). how would your revised training know who (& how) to make payments

We're talking about Adobe generative fill. Adobe knows very well where to find the artists who's photos were used and how to pay them as they used images from their own database.

yes, it's about AS specifically but i was responding to the comments that were not limited to AS.

again, how does AS know whose images were used for each creation since ML eliminates any way to track that, even if identifiers were attached initially.  each image is translated into thousands of datapoints and millions/billions of operations are performed to generate each new image.

Okay - you aren't quite thinking correctly here.

It requires revising the machine learning algorith to incorporate identifiers and then attribute those identifiers to outputted information.

Keeping things simple.

Lets say you have 3 contributors, named "|A|" and "|B|" and "|C|".
|A| has images of an apple, and a pear
|B| has an image of an apple
|C| has an image of a pear, and an orange

Let's say the "machine" ("AI") version of an apple is "ML-APPLE" and "ML-PEAR".

The "AI" (ML/machine learning algorithm) creates a "representation" of what it believes an "apple" to be by scraping |A| + |B|'s image.
It then does the same for a pear, by scraping |A| + |C|'s image.

In it's internal representation, it would look like:

[ML-APPLE]:{|A|,|B|} (simply meaning what I stated above - the "ai" version of an apple references |A| + |B|'s image
[ML-PEAR]:{|A|,|C|} (simply meaning what I stated above - the "ai" version of a pears references |A| + |C|'s image
[ML-ORANGE]:{|C|} (simply meaning what I stated above - the "ai" version of an orange references |C|'s image

Let's say you then have a customer that generates images. Let's say they pay $1/image (simplicity), & its a 50-50 share between agency (the "ML" image) + the source contributors.

They decide to make a picture of a "pear". Since the "ML-PEAR" references |A|+|C|'s image - |A|+|C| would be compensated for the use of that image generation.
I.e., $1 = $.50 agency, $.50 to contributors. Since two contributors (A+C) made the "pear" image, the revenue for contributors would be issued to them).

Now let's say you made an image of an orange.
Since |C| was the only source document referenced, |C| would get full credit for this image. (I.e., $1 => $0.50 to agency, $.50 to |C|).

That is a super basic illustration of what I am talking about.

Of course, the pseudocode above is an extremely simplistic concept - it is simply designed to illustrate how it would be done on a most basic level, and some of the requirements to revise the agorithm to attribute source images.

Of course, actual code would be much more sophisticated, and one could then decide whether to attach weights to "how much" of the model was used (i.e., was it a "tiny" pear in the image, or a "big" pear in the image, and should they be compensated accordingly?) As well as how "much" of the "pear" was attributed to a specific contributor. (I.e., did |C| say have 50 images of pears, and |A| only 1 image of pear, that was used in the model/representation of what a 'pear' was - such that |C| should get 50x the 'credit' for the pear image?) Of course - that is a little more in depth, and this example was simply used for illustration purposes.

Fact is - it IS super easy to properly attribute source images, AND - it is ALSO super easy to properly CREDIT source images - on a PERPETUAL RECURRING BASIS.

It is simply a matter of taken the time to revise the algorithm to do so.

...
So it is SUPER easy to simply add "contributor-id" (which can be the name/URL/etc or an actual number that contains all that information). And then SUPER easy to associate WHICH contributors file(s) were used in creating a "composite" image (i.e., an "ai" generated image).

SUPER SUPER EASY. Just a matter of doing it, then fairly compensating contributors with the SAME RECURRING PERPETUAL INCOME REVENUE model that the agencies so desperately and greedily want for themselves, and trying to convince contributors that anything else is "fair" (which of course, it's not). Sharing the recurrnig revenue model, with opt-in/opt-out features so at ANY time the contributor can opt-out if they don't like the terms  - and assets going forward do NOT reference the input items - is fair.

i agree, starting from scratch, adding contributor id is the easy part, but given the number of possible contributors to each piece of an ai-gen image, the book-keeping becomes expensive - many thousands of entries for each image.

but the much larger problem you don't address is actually finding out who the author is and how to pay them as scraping only works if the image has the artist's verifiable info (and how to verify they're who they claim to be when opting out).  true, when agencies create datasets thery have info for payment, when they license these datasets who tracks payments? and will agencies actually sell data sets with the private info of their contributors?

your examples are great for explaining your proposal, but they don't scale up - the amount payable to any artist is 2-3 orders of magnitude less