Author Topic: Significant growth of Adobe Stock collection over the last 7 months (Read 2129 times)

cobalt · « **Reply #25 on:** November 25, 2024, 05:20 »

You are asking for all of Adobestock to be a controlled edited collection. Which puts additional burdens on the reviewers or would require content be sent to specialized genre reviewers.

This goes against the principle of micros being open platforms where anyone can join and upload as long as the technical quality is fine.

Agencies deal with the duplicates quite easily with the algos. That works much faster than expecting every image reviewer to also be trained as a needed style editor.

In addition, all agencies create beautiful trendy collections for customers that they promote on their websites and newsletters.

I don't think that the duplicate issues, whether with normal camera or ai, is as much of a problem as people think.

Adobe also has a premium collection for high quality and often very artsy content.

Because spammers focus on duplicating, we all have plenty of opportunities to make money because the genuinely truly new content coming is actually very little.

There is so much content that I am now doing with ai, that could have been very, very easily done with a normal camera over the last 20 years. Strangely many of these subjects are not covered at all, often less than 100 files in a theme affecting thousands of people around the world.

And since the ai spammers do no research, to me it feels like they are just uploading pretty flower decorations that do not compete with my files.

Just my 2 cents.

davedigitalfx · « **Reply #26 on:** November 25, 2024, 09:33 »

zeljkok: "That image by Herr Mustafa is brutal"

woooooowwww .. i didnt see that sh*t ...

... AI is raising the quality of Adobe collection

davedigitalfx · « **Reply #27 on:** November 25, 2024, 09:45 »

Quote from: cascoly on November 23, 2024, 12:50

Quote from: Jo Ann Snover on November 21, 2024, 15:36
...

Over all, the collection grew 32% between the end of April 2024 and today, but the genAI portion of the collection grew 82% versus the human-made portion grew 6.1%.

...

do we have actual numbers in each case? 6% of a very large number can be much greater than 82% of a much smaller number

yes right .. .also if we consider the 80-20 law ... that the 20% gives the 80% of the profit .... and the other 80% gives the 20% of the profit ...

cascoly · « **Reply #28 on:** November 25, 2024, 14:14 »

Quote from: Jo Ann Snover on November 23, 2024, 14:28

Quote from: cascoly on November 23, 2024, 12:50
Quote from: Jo Ann Snover on November 21, 2024, 15:36
...
Over all, the collection grew 32% between the end of April 2024 and today, but the genAI portion of the collection grew 82% versus the human-made portion grew 6.1%.
...
do we have actual numbers in each case? 6% of a very large number can be much greater than 82% of a much smaller number

thanks! those are amzing numbers

Adobe Stock growth Apr 30 - Nov 21 2024; roughly 7 months

total collection: 139,801,939 (576,474,125 up from 436,672,186) - up 32%

genAI: 117,178,564 (181,469,399 up from 64,290,835) - up 82%

human made: 22,623,375 (395,004,726 up from 372,381,351) - up 6.1%

human made asset types

photos growth 9,936,406 (228,291,001 up from 218,354,595) - up 4.5%

videos shrunk 3,202,090 (15,884,530 down from 19,086,620) - down 16.8%

illos growth 1,990,088 (33,691,112 up from 31,701,024) - up 6.3%

genAI asset types

photos 49,408,590 (71,313,378 up from 21,904,788) - up 126%

illos 65,134,599 (106,056,450 up from 40,921,851) - up 59%

videos 846,305 (1,237,747 up from 391,442) - up 116%

th anks! those are amazing numbers

tätarätä · « **Reply #29 on:** November 25, 2024, 16:55 »

Shocking numbers.
AI photo uploads 5 times compared to real photos.
AI illustration uploads 33 times compared to real illustrations.

Andrej.S. · « **Reply #30 on:** November 25, 2024, 17:42 »

Quote from: cobalt on November 25, 2024, 05:20

You are asking for all of Adobestock to be a controlled edited collection. Which puts additional burdens on the reviewers or would require content be sent to specialized genre reviewers.

This goes against the principle of micros being open platforms where anyone can join and upload as long as the technical quality is fine.

Agencies deal with the duplicates quite easily with the algos. That works much faster than expecting every image reviewer to also be trained as a needed style editor.

In addition, all agencies create beautiful trendy collections for customers that they promote on their websites and newsletters.

I don't think that the duplicate issues, whether with normal camera or ai, is as much of a problem as people think.

Adobe also has a premium collection for high quality and often very artsy content.

Because spammers focus on duplicating, we all have plenty of opportunities to make money because the genuinely truly new content coming is actually very little.

There is so much content that I am now doing with ai, that could have been very, very easily done with a normal camera over the last 20 years. Strangely many of these subjects are not covered at all, often less than 100 files in a theme affecting thousands of people around the world.

And since the ai spammers do no research, to me it feels like they are just uploading pretty flower decorations that do not compete with my files.

Just my 2 cents.

This is ambivalent thinking.

You often complain about copycats, but on the other hand, you don't want Adobe to control the flood of images.

The problem with the spammers comes from the fact that there is no control and that they orient themselves to the best sellers as the best guess.

If, on the other hand, they had some help, along the lines of: Look here, we still have gaps here, we don't have enough content on ESG, sustainable teaching in schools, special health topics, new creative seasonal images, etc.,

Then all the Indians, Thais and Eastern Europeans would suddenly do something useful and quickly fill in all the gaps. Then there would be far fewer copycats.

Why keep the old system running when you can optimize it?
Even back then, Shutterstock, for example, rejected content with too little sales potential. And Adobe should reintroduce this review system.

zeljkok · « **Reply #31 on:** November 25, 2024, 18:42 »

My opinion, for what it's worth:

It all comes down to basic thing that is just about to kill entire microstock industry: Flood of content.

At the beginning you had to have real camera and be a photographer --> 3 or 4 digit sales on places like Alamy
Proliferation of low cost gear, smart automatic mode - just point and shoot. Everyone becomes a photographer
Advancement of smartphones, with semi-decent cameras --> exponential increase of content, exponential decrease of overall quality
"Rise of Machines" (too many Terminator movies, sorry). AI age removes final requirement, need to have camera in the first place

I believe Adobe honestly wanted in the beginning to have "lean and mean" library - high quality, without similars, etc etc. But they just could not cope, just like everyone else, with floods. So they resorted to algorithms that simply push 'unwanted' content in the swamp, like most other agencies did as well. End result is near.

It's a shame, because there are lots of Rewards in Photography whatever the form one prefers. I can fully understand people disillusioned with with what stock industry has become pulling their entire ports out.

cobalt · « **Reply #32 on:** November 26, 2024, 01:16 »

Like I said, my thinking has evolved, I no longer worry about duplicates because the algos will handle that.

All agencies have newsletters and blogs with needed content. But I sincerly doubt the spammers would listen.

They want the „shortcut“ and believe copying the topselling files is the easy path to get rich quickly.

Adobe could try to do more, maybe a country specific youtube channel with content suggestions. But filling niches is a slow way to grow the port and income.

Whereas making copies of 1000 best selling files will initially bring faster results.

So…I will stick to the niches and not worry about the flood.

Jo Ann Snover · « **Reply #33 on:** November 27, 2024, 10:18 »

The default order when you do a search on Adobe Stock (at least in the US) is "Relevence". Many days and with many searches, the top half of the first page is almost all AI images. IMO the massive volume of new (and generally low quality) genAI acceptances means that leaving it to Adobe's software to sort it out is wishful thinking. That sort of thing used to work when collections were smaller, inspections actually meant something and new work was a smaller proportion of the total collection size.

In five days - from Nov 21 to yesterday afternoon (Nov 25) Adobe Stock's total collection grew by just over 6 million items (6,075,697).

Nearly 5.5 million (5,499,162) were genAI.

Let those numbers sink in. And think about the possibility that the genAI content will slowly drive away everything else, and then buyers will drift away when they're bored with the sameness of everything left. Forget what's in my best interests as an individual contributor; Adobe Stock is taking a huge gamble that (a) genAI nearly-like-real-life content is OK with buyers and (b) that if it is, in time buyers will be able to find cheaper places to get it (Freepik's genAI collection is now 168.18 million) or make it themselves.

I've been licensing stock for just over 20 years. I remember a big celebration at iStock when they reached 1 million images...

cobalt · « **Reply #34 on:** November 27, 2024, 11:14 »

Customers can deactivate and filter out gen ai content with one click...as long as that is possible I am really not worried.

I do hope Adobe at some point implements the "personal feed" option that I have been pitching to agencies for over 10 years.

It would make the search experience so much better if a customer could develop his own feed in addition to the standard search options.

With the high volume of ai coming in, the value of non ai might actually go up. And non ai is still selling very well.

Robert Kneschke just published a very interesting blog posts that ai companies might be forced to license non ai content properly simply because the internet is now flooded with ai and that content is "poison" to the training.

article is in German, but google will help

https://www.alltageinesfotoproduzenten.de/2024/11/27/zerstoert-der-ki-hype-selbst-seine-zukunft/?fbclid=IwY2xjawG0JItleHRuA2FlbQIxMQABHVrWzzTNkGfgdSDRtjtYpdy-27r2t0HlpsFqr_yMnA0F__DPAIsV83Gkiw_aem_qhskGZVkaIMla26q7ALxdA

eta

With the high volume of ai coming in, that is a good reason to keep uploading good and high quality normal photos and videos. I think next year the pendulum will swing back, the novelty of ai will wear off and many customers will want to buy real images again, especially with people, food, locations, lifestyle.

Uncle Pete · « **Reply #35 on:** November 27, 2024, 14:31 »

LAION owns the metadata and release as CC-BY-4.0
"We do not own the copyright of the images or text."
LAION 5B was completed in 2022 before AI had invaded most sites.
LAION 5B and the previous collections are Open Source

Any future training should exclude AI images. I'm not saying it will, but it should. Part of the problem of scraping for AI is more complicated when it comes to text and undisclosed writings that are created by AI. But here, I think we're concerned with images.

"AI has generated 150 years worth of images in less than 12 months, study shows" https://www.designboom.com/technology/ai-has-generated-150-years-worth-of-photographs-in-less-than-12-months-study-shows-08-21-2023/

Just for interest: LAION-5B contains 2.3 billion English samples, 2.2 billion multilingual samples, and 1.2 billion unknown language samples. Depending on bandwidth, it’s feasible to download the entire LAION-5B dataset in 7 days using 10 nodes.

I hope that answers some questions about the size and access. I think from what I've found, the dataset is free, and no one was charged or paid for the content. Non-profit funding was used for the gathering of the data.

Andrej.S. · « **Reply #36 on:** November 28, 2024, 02:19 »

Quote from: Uncle Pete on November 27, 2024, 14:31

LAION owns the metadata and release as CC-BY-4.0
"We do not own the copyright of the images or text."
LAION 5B was completed in 2022 before AI had invaded most sites.
LAION 5B and the previous collections are Open Source

Any future training should exclude AI images. I'm not saying it will, but it should. Part of the problem of scraping for AI is more complicated when it comes to text and undisclosed writings that are created by AI. But here, I think we're concerned with images.

"AI has generated 150 years worth of images in less than 12 months, study shows" https://www.designboom.com/technology/ai-has-generated-150-years-worth-of-photographs-in-less-than-12-months-study-shows-08-21-2023/

Just for interest: LAION-5B contains 2.3 billion English samples, 2.2 billion multilingual samples, and 1.2 billion unknown language samples. Depending on bandwidth, it’s feasible to download the entire LAION-5B dataset in 7 days using 10 nodes.

I hope that answers some questions about the size and access. I think from what I've found, the dataset is free, and no one was charged or paid for the content. Non-profit funding was used for the gathering of the data.

I don't know if anyone has taken a closer look at the LAION-5B dataset.

I don't remember any high-resolution images in there, just some thumbnails with watermarks.

I mean, if someone had painted the pictures, scanned them and then used them as training material, would anyone have complained?

For me, Robert is the epitome of hypocrisy.
He complains about the LAION-5B dataset, but generates his own AI images with Midjourney & Co.

Clown world.

cobalt · « **Reply #37 on:** November 28, 2024, 05:05 »

The question is simply copyright.

laion. and some ai companies are trying to make money by stealing other peoples copyrighted works.

That should be illegal.

Ai companies should properly license all their training material and many large companies do.

And it seems now they might be forced to do just that because ai images are poisoning their free scraping pool.

Karma is a bitch.

Andrej.S. · « **Reply #38 on:** November 28, 2024, 05:46 »

Quote from: cobalt on November 28, 2024, 05:05

The question is simply copyright.

laion. and some ai companies are trying to make money by stealing other peoples copyrighted works.

That should be illegal.

Ai companies should properly license all their training material and many large companies do.

And it seems now they might be forced to do just that because ai images are poisoning their free scraping pool.

Karma is a bitch.

Ok, when you are talking about copyright it is not that easy in this case.
Is it forbidden to use an image protected by watermarks for the training of AI models and then no longer use it or no publish it as such?

From my point of view, the question is much more: how high is the reproducibility of the material used by the AI model?
In other words, can the image used be reproduced 10, 20, 50 or 99% as such with a specific prompt by the user?

A high percentage of reproducibility is technically known as overfitting of the model.
And this is exactly where the law should show limits.
Every developer should disclose its model calibration quality public, so that this specific issue can be evaluated.
Anything else is nonsense and not worth further discussing.

Uncle Pete · « **Reply #39 on:** November 28, 2024, 14:53 »

Quote from: Andrej.S. on November 28, 2024, 02:19

Quote from: Uncle Pete on November 27, 2024, 14:31
LAION owns the metadata and release as CC-BY-4.0
"We do not own the copyright of the images or text."
LAION 5B was completed in 2022 before AI had invaded most sites.
LAION 5B and the previous collections are Open Source

Any future training should exclude AI images. I'm not saying it will, but it should. Part of the problem of scraping for AI is more complicated when it comes to text and undisclosed writings that are created by AI. But here, I think we're concerned with images.

"AI has generated 150 years worth of images in less than 12 months, study shows" https://www.designboom.com/technology/ai-has-generated-150-years-worth-of-photographs-in-less-than-12-months-study-shows-08-21-2023/

Just for interest: LAION-5B contains 2.3 billion English samples, 2.2 billion multilingual samples, and 1.2 billion unknown language samples. Depending on bandwidth, it’s feasible to download the entire LAION-5B dataset in 7 days using 10 nodes.

I hope that answers some questions about the size and access. I think from what I've found, the dataset is free, and no one was charged or paid for the content. Non-profit funding was used for the gathering of the data.

I don't know if anyone has taken a closer look at the LAION-5B dataset.

I don't remember any high-resolution images in there, just some thumbnails with watermarks.

I mean, if someone had painted the pictures, scanned them and then used them as training material, would anyone have complained?

For me, Robert is the epitome of hypocrisy.
He complains about the LAION-5B dataset, but generates his own AI images with Midjourney & Co.

Clown world.

I can't comment on anyone else, or hypocrisy, I don't know. I'm interested in the questions of law, AI and the output.

The way I understand things and I know there are people who disagree, the images are used for training, not for creation of new images. So the machine it trained to make a banana after looking at thousands of pictures of bananas, and learning the shape, color and bits of information of "what is a banana". Then when a new image is made, none of the original images are directly used, as the machine creates from what it has learned.

Also I see many posts and if I read things right, LAION 5B and the previous projects are licensed Creative Commons and they are Open Source, which I interpret to mean, they are FREE not licensed for a fee. Since those datasets are older and not continuous, completed in 2022, anything new that some AI software wants to add, to make the machine better, has to be paid for, from some other source.

MicrostockGroup Sponsors

Author Topic: Significant growth of Adobe Stock collection over the last 7 months (Read 2129 times)

cobalt

davedigitalfx

davedigitalfx

cascoly

tätarätä

Andrej.S.

zeljkok

cobalt

Jo Ann Snover

cobalt

Uncle Pete

Andrej.S.

cobalt

Andrej.S.

Uncle Pete

Related Topics

Sponsors

Microstock Poll Results

Sponsors