Changing regions produces completely insane results - no idea where exactly the bugs are, but this is important to fix.

As a particularly crazy example, if you search for Berlin - spelled the same in German and English:

274.118 Ergebnisse f�r Berlin in �Bilder�
275,093 results for Berlin in images (US)

No translation of anything required but nearly 1,000 images are just missing.

And for the Tuscany example, if you switch to the Italian region:

366.635 risultati per Tuscany in immagini
220.371 risultati per toscana in immagini

Searching for Tuscany in French in the French region:

161 562 r�sultats pour Toscane dans images
366 635 r�sultats pour Tuscany dans images

Edited Tuesday to add that I haven't been looking at Freepik and didn't realize they've been accepting genAI content for many months. They now have 26,580,000 genAI assets, so they're bigger than Adobe Stock's genAI collection.

One interesting feature they have is a box under each genAI item for "Base model", and although many contributors just say "Not specified", I also saw Midjourney 5.2, Stable Diffusion 1.5, Stable Diffusion 2.0, Dall-e 1, Stable Diffusion 2.1, Leonardo. Midjourney seems to be the most used of the ones named.

Lots of logos in their collection too.

==========================

I have been tracking the size of the genAI collections each Monday.

Adobe Stock's is the largest by farsecond largest after Freepik; Dreamstime is the only other agency with any sizable collection. Shutterstock's nearly 1.4 million genAI images are mostly a cautionary tale of what can go horribly wrong - I don't know why they'd exhibit how bad their on-site AI tools are for all to see.

Adobe Stock's genAI collection is now 20,649,310 - up over 1 million from 16 Oct (19,590,366) which was up nearly a million from 9 Oct (18,641,502). Before that it was more like half a million per week.

Separately from the issue of usability of these images is where all the extra customers to buy these will come from?

At some point I'd have thought those new contributors drawn in by the AI gold rush would realize there wasn't much return and just drift away. Not this week though

Looking at the content overall, though there are some images only AI could do, the vast majority are just replicating well-covered subjects already in the human-created collection (and with tons of "similars" which further dilutes sales)

Adobe Stock's collection is about half the size of Shutterstock's - but SS bulking up their numbers seemed to be to impress investors rather than anything useful for buyers. Adobe Stock reports 369,012,172 results in all. SS has "445,783,263 stock photos, 3D objects, vectors, and illustrations" but the footer of each page says "We have more than 734,000,000 images across Shutterstock's properties as of June 30, 2023"

Washington Post article (paywall) about major newspapers looking for payment for use of their content for AI tools/training

https://www.washingtonpost.com/technology/2023/10/20/artificial-intelligence-battle-online-data/

There is a mention of the estimates of Shutterstock's payout to contributors for data training on the collection:

"For example, the stock photo site Shutterstock has a partnership to provide training data for OpenAI. Late last year, the company also launched a Contributor Fund to compensate artists whose work has been used to train AI models. An analysis by stock photographer Robert Kneschke estimated that the fund paid out more than $4 million in May � but the median payout was just $0.0069 per image. Shutterstock did not respond to request for comment."

The above contains a link to this PetaPixel story:
https://petapixel.com/2023/07/12/shutterstock-may-have-paid-out-over-4-million-from-its-ai-contributor-fund/

Some other interesting quotes (emphasis mine):

"Until recently, tech companies have been loath to pay for that data. At a listening session on generative AI hosted in April by the U.S. Copyright Office, Sy Damle, a lawyer representing the Silicon Valley venture capital firm Andreessen Horowitz, acknowledged that �the only practical way for these tools to exist is if they can be trained on massive amounts of data without having to license that data.� "

I'd suggest that it is not impractical, just more expensive. What theory of business mandates that your raw materials have to be free?

OpenAI continues to push the notion that their original scraping of data they didn't own or license was fair use:

"An OpenAI spokesperson confirmed that the company is in talks with the newspapers and that discussions were not focused on prior training data, which it argues was obtained legally. �None of the company�s practices have violated copyright law,� the spokesperson said. �Any deal would be for future access to content that is otherwise inaccessible or display uses that go beyond fair use.� "

"Nearly $16 billion in venture capital poured into generative AI in the first three quarters of 2023, according to the analytics firm PitchBook � a flood of cash that in part reflects how expensive the technology is to build. Every component is prohibitively pricey or hard to acquire, from hardware to computing power. Until now, the only free and easy part had been the data. "

I left EyeEm in January 2021 but thought I'd check to see if my portfolio was still deleted after I read your post - and it is. It was $0.21 royalties from the Getty collection that made me leave.

My portfolio URL doesn't exist (404 error) and my old account login doesn't work. I did a couple of searches for photos that I have sales records for to be sure they weren't online, and there's nothing.

I don't know when you deleted your account, but if the difference isn't based on exit date, I wonder what it is?

There are also vectors in the collection that are clearly labeled as "generative AI" but not tagged as such and have somehow been accepted anyway. Two examples:

Lemons in cartoon, doodle style. 2d vector illustration in logo, icon style. AI Generative  


agressive shark | Generative AI


If you look at the image numbers, this isn't from ages ago, but is recent. The rules are just ignored and the reviewers don't catch it.

Google now has its text to image generation, along with other AI tools in the "Search Generative Experience" - SGE

https://www.zdnet.com/article/google-finally-adds-ai-text-to-image-generation-but-its-not-where-you-think/

"On Thursday, Google unveiled two new features for search: Text-to-image generation and a written drafts feature in its AI-powered Search Experience (SGE)."

https://www.techtarget.com/searchenterpriseai/news/366555473/Google-Search-Generative-Experience-AI-updated-for-images
https://www.theverge.com/2023/10/12/23913337/google-ai-powered-search-sge-images-written-drafts
https://techcrunch.com/2023/10/12/googles-ai-powered-search-experience-can-now-generate-images-write-drafts/

It's not really news that agencies are accepting content that violates their published rules, but given the recent WAG strike and ongoing SAG-AFTRA strike where AI use by companies is a key issue, I can see how this became a subject of discussion.

https://www.theverge.com/2023/10/9/23909529/disney-marvel-loki-generative-ai-poster-backlash-season-2

Some interesting points of view in the comments. For example how designers under pressure to keep their jobs feel

"It's an integrated tool in Photoshop at this point as well. If you're not using all the tools available to you, you're simply going to be seen as slow to employers. This pixel-measuring is just as ridiculous as when people tried to determine if filters were used in the 2000s. Having consumers breathing down creatives backs for the tools they pick is just an unnecessary pressure in an otherwise high-tension position to begin with."

"As a senior graphic designer who often does highly visible work for a myriad of household names, I can only shrug. Generative AI tools are an entrenched part of my workflow now, albeit at conceptual stage rather than finished artwork, or client internal communications use otherwise � and the line has become blurred even further with the release of Photoshop's Generative Fill tools. The expected pace of design turnaround is faster than ever and I need to stay ahead of the hack account managers armed with Canva and MidJourney taking evening design courses, or its my job on the line."

"I mean, as a designer I understand very well that I have to keep up at this point. Pandora's box is open. It's out there. It won't go away. I'm also using it. But I do see the injustice in how big tech companies profit off of the labor of millions of artists and then selling the computer made remix back to me as a service. I feel if there's still a window where open outrage can sway some policies to better integrate artists and designers into the value chain it is probably now. If not now - it will be never."

Slightly different angle on the same topic - about how key Adobe's finance head is in making AI product decisions...

https://finance.yahoo.com/news/adobe-cfo-helping-steer-company-110806766.html

My experience is that the understanding finance people have of the guts of the business behind the numbers is very shallow

Do not apologize for the "wall of text" - that was very helpful. It spurred a lot of thoughts about a real tangle for agencies in evaluating what they will and won't accept in the future.

It also highlighted that I need to do a little reading - I am starting with this Google overview and will see if I can educate myself on this topic

https://developers.google.com/machine-learning/gan

I don't use Topaz's tools, but all my uploaded images have been massaged by RAW processors and Photoshop tools. I don't consider those changes to be generated, but clearly we're a long way from processing film in a darkroom

When Photoshop was a baby, it was viewed as suspect by some, but the message from those of using these tools was that it was just doing what you could always have done in the darkroom but without the chemicals and mess. Even if you just stuck with Capture One or Lightroom and Photoshop, we're way beyond that now.

Clearly agencies need to scale back on trying to cut costs by using inadequate tools to evaluate contributor's content, but it does raise the question of what the appropriate yardstick should be.

Firefly 2 (beta) is now like a brain that's three times larger!!

"Alexandru Costin, Adobe�s VP for generative AI and Sensei, told me that the new model wasn�t just trained on more recent images from Adobe Stock and other commercially safe sources, but also that it is significantly larger. �Firefly is an ensemble of multiple models and I think we�ve increased their sizes by a factor of three,� he told me. �So it�s like a brain that�s three times larger and that will know how to make these connections and render more beautiful pixels, more beautiful details for the user.� The company also increased the dataset by almost a factor of two, which in turn should give the model a better understanding of what users are asking for."

https://techcrunch.com/2023/10/10/adobe-firefly-can-now-generate-more-realistic-images/

I didn't do much testing, but with what I did with photos, I can't really see any improvement - and it's not ready for prime time use, IMO. I did not try the match feature - where you upload your own photo to match the style. I'd want to look very carefully at what rights I'm giving Adobe over anything I upload.

I had already decided I need to avoid the generative tools in Photoshop 25 if I want to submit work to stock agencies, but the warning popup in Firefly 2 pointed me at terms and conditions that made it clear no contributor can use these tools if they upload to any agency that trains AI on their own collections.

https://www.adobe.com/legal/licenses-terms/adobe-gen-ai-user-guidelines.html

1. No AI/ML Training
 
When using our generative AI features, you agree you will use them only for your creative work product and not to train AI/ML models.

This means you must not, and must not allow third parties to, use any content, data, output or other information received or derived from any generative AI features, including any Firefly outputs, to directly or indirectly create, train, test, or otherwise improve any machine learning algorithms or artificial intelligence systems, including any architectures, models, or weights.


Later on it mentions that they might introduce some beta features that wouldn't be OK for commercial use in any context but is very vague about how you're supposed to know (emphasis mine):

"However, if Adobe designates in the product or elsewhere that a beta version of a generative AI feature cannot be used commercially..."

https://www.engadget.com/adobes-next-gen-firefly-2-offers-vector-graphics-more-control-and-photorealistic-renders-160030349.html
https://www.pcworld.com/article/2100888/photoshops-ai-powers-land-on-chromebooks-as-adobe-debuts-firefly-2.html

This describes the illustrator genAI features (beta):
https://www.theverge.com/2023/10/10/23911114/adobe-max-firefly-generative-ai-model-photoshop-illustrator-express

There's something here about Adobe's approach to safety and content credentials:
https://blog.adobe.com/en/publish/2023/10/10/adobe-firefly-new-features

The stored thumbnail of a reference image used in genAI will supposedly be used to deal with claims that an inappropriate image was used (and not for training Firefly)

https://www.pcmag.com/news/adobe-unveils-firefly-image-2-ai-at-max-conference
https://9to5mac.com/2023/10/10/adobe-firefly-ai-model-updates/

https://www.zdnet.com/article/adobe-unveils-three-new-generative-ai-models-including-the-next-generation-of-its-firefly-text-to-image-generator/

https://www.forbes.com/sites/johanmoreno/2023/10/10/adobes-all-in-on-ai-debuts-new-firefly-ai-models-for-creatives/

https://www.adweek.com/programmatic/adobe-firefly-generative-ai-offerings/

https://www.fastcompany.com/90964791/adobe-ai-is-coming-for-your-branding

"Adobe also made improvements to its image generation engine, which is trained on its own stock image library. While the model allows Adobe to avoid messy copyright issues faced by other text-to-image generators trained more broadly, the downside is images look like stock photos. Under the Firefly Image 2 Model, though, generative images look less stock-like thanks to additional inputs including human feedback and the company�s �own private data set,� Costin says."

https://arstechnica.com/information-technology/2023/10/adobes-ai-image-generators-get-beefy-updates-including-vector-graphics/

"While Adobe claims that the Firefly Vector Model is "the world�s first generative AI model for vector graphics," a quick Google search disproves this claim, showing earlier vector-based AI models that range from commercial offerings to older experimental research. However, it's safe to say that this is the first vector-generating AI model from a company that is currently considered a major player in the image synthesis space."

I don't have any similar experiences, but I wonder if you might understand more about how this mistake occurred by looking at the metadata in your uploaded file.

When I was experimenting with Photoshop 25 and generative fill, I used an online metadata tool to look at and compare what was in the new file versus one created with PS 24.x. As far as I can tell from the limited tests I did, if you stay away from the AI-powered features, your Photoshop edited files should be fine.

https://drive.google.com/file/d/1Z_2dnk-QdSpAU47AQZSXgwwPYRhIsa9A/view?usp=sharing

It's possible it's the Gigapixel tool that's embedding something that triggered SS's automatic screening, so possibly you need to clean the metadata out after that step and before editing in Photoshop?

Let us know what you find out.

Once Firefly was out of beta, it was available for commercial use - also the generative fill in Photoshop 25

However, most stock agencies do not accept AI generated content, and the content credentials in your Firefly/generative fill from Photoshop will show that AI was used.

Dreamstime, Deposit Photos and I think 123rf accept gen AI content, but not Getty/iStock, Shutterstock or Alamy

Thank you for letting us know - and it�s outrageous you�ve been treated this way.

Adobe Stock has lost the plot on so many issues. Shameful.

This is sad for the real models. All the AI people are starting to look the same.

Two different contributors. These creations could be twins if they were real people

 

The "problem" (from my perspective, and I suspect in short order from a buyer's perspective too) is that there are huge volumes of nicely lit images of not-quite-real look-alike humans. Impossibly fit and beautiful seniors, shiny trendy young adults, very spooky children who look adult but smaller.

I have samples from "Find Similar" searches and it's really spooky.

Sometimes, the hotter the fad, the faster it fades, so possibly this won't last and human models will still find work?