Will this news affect those who make generative  AI images for stock sites?

OpenAI, Google, others pledge to watermark AI content for safety -White House

https://www.reuters.com/technology/openai-google-others-pledge-watermark-ai-content-safety-white-house-2023-07-21/

One month after the OP, the Adobe Stock genAI collection (tagged; there are more that are not) is over 12 million - 12,004,534

We can celebrate by making a yummy fruit smoothie - although I'm really afraid of this mixer...

The mixer is clearly Borg and in the process of assimilating the kitchen.

Resistance is futile!

Quote from: Jo Ann Snover on July 13, 2023, 10:20

One month after the OP, the Adobe Stock genAI collection (tagged; there are more that are not) is over 12 million - 12,004,534

We can celebrate by making a yummy fruit smoothie - although I'm really afraid of this mixer...

The mixer is clearly Borg and in the process of assimilating the kitchen.

Resistance is futile!

or a hand crank?

Well, there are certainly double standards in reviewing.
Good and useful images that get rejected for "quality issues" (yes, looking at you Adobe) while AI seems to get a free pass-through, no matter how much they're off.

And I get it, AI generated content that looks amazing at first sight (and awful with closer inspection) is easy to produce, and agencies must be flooded with that.
I also get that agencies don't want to miss out on the AI train.

But they have to realize that sloppy reviewing floods their database with junk, and in the end, this doesn't do any good for their customers either.
In the end, they pass the reviewing over to their customers.: bad content doesn't sell and gets buried by the search engine algorithm.
That system works fine for the occasional error that reviewers make, but if you flood your database with junk, nobody will bother anymore to wade through it.

We have the "junk" issue with photography too, but not in this volume, not at this pace.

However bad you think the accepted genAI images are, go take a look at a few real world types of images - things you think a buyer might be searching for - and you'll realize it's worse.

These examples are just a selection from page 1 of the default (Relevance) search. I'm ignoring the fantasy images (even the cute mouse with sunglasses and a bow tie) as that's not where the issue is.

Somehow this woman made it on a bike with one pedal - possibly because she has part of one leg missing - and strange mangled spokes. Extra points for her not holding the handlebars, plus bad Photoshop work between the handlebar and brake lever. Cheating to cover the suspicious construction of the seat to frame connection with her dress!



Apparently Byzantine painters were no better with correct number and shapes of hands and feet than AI - who knew?



This is keyworded as Santorini - as are 4,037 other genAI images. Did not count how many of those look like photos - some are labeled as watercolor or other artistic style. The pseudo paintings or pseudo illustrations seem less problematic in that they are clearly not purporting to show real life.



"Image of long rows of green beds with growing cabbage or lettuce in a large farmer's field". I'd bet on lettuce - though I'm not a farm or vegetable expert.



Here's what a row of cabbages actually looks like

"A happy couple stands smiling in the driveway of a large house with solar panels installed". If you look at their front door (left) their smiles must result from having figured how to get out and down the steps without breaking their necks!



Chair and stool legs, along with hanging lights, cabinet handles and other details, seem to confuddle AI - don't sit on chairs in AI-world.


Impossibly hip and stylish high school student can't escape AI's problems with digits (mangled pinky)


It's all really depressing - instead of celebrating the strengths of genAI images, we're getting cheap and wonky knock-offs of real world places and things. I'm hoping the market (buyers; Adobe clearly loves welcoming this stuff into the collection) speaks and ignores things that aren't useful. Or which get them articles in the Guardian

I have a large folder of Firefly experiments. It's still in beta. I haven't produced a single usable image (for the things I was interested in working on), and the workflow seems awkward to me -not sure how I would use it if it was out of beta.

Earlier in the week I tried the generative fill option on an image I'd created with firefly and it was funny - not intentionally. I wanted to replace the floor surface in an interior shot. It couldn't automatically select the floor, or even automatically extend my manual selection. The perspective and scale on the fills wasn't right. A mess.

It's all really depressing - instead of celebrating the strengths of genAI images, we're getting cheap and wonky knock-offs of real world places and things. I'm hoping the market (buyers; Adobe clearly loves welcoming this stuff into the collection) speaks and ignores things that aren't useful. Or which get them articles in the Guardian

I think that's the real question: how will we all, as customers, accept these images that are meant "to illustrate" a product or a service?
The burger advertised by McDonalds doesn't look like the burger you get at their counters.
The car that I bought, and was advertised as a shiny car driving to smooth city traffic against flashy buildings, while reality is more like getting stuck in shabby neighborhoods.
Santorini never looks like it get's advertised in travel brochures. People everywhere! 
And I have never seen a couple smiling on their driveway because they just installed solar panels.

The point is: we, as a public, already accept an idealized version of reality, and we know that the real world looks more bleak.
We might as well getting pushed in accepting the next version of it. Impressions of how it could look like, with shabby and wonky AI imagery.

After all, on every advertisement there's a small notice: real product might differ from advertised product.

Quote from: Jo Ann Snover on July 28, 2023, 09:48

It's all really depressing - instead of celebrating the strengths of genAI images, we're getting cheap and wonky knock-offs of real world places and things. I'm hoping the market (buyers; Adobe clearly loves welcoming this stuff into the collection) speaks and ignores things that aren't useful. Or which get them articles in the Guardian

I think that's the real question: how will we all, as customers, accept these images that are meant "to illustrate" a product or a service?
The burger advertised by McDonalds doesn't look like the burger you get at their counters. ....

After all, on every advertisement there's a small notice: real product might differ from advertised product.

Point taken - I've been licensing stock and looking at ads/marketing materials for decades. "Authentic" in stock isn't real life - and if you look at all the "shiny-happy people" pictures, genAI or old-school, they're better lit, tidier, posed and typically younger/prettier than most people's everyday life.

However, there's a huge difference between cleaning up, polishing and giving a sunset glow to real life and making things up that sort-of look like the real thing.

Here's an example. Below is a genAI image of Sheffield town hall, Yorkshire, UK. This link will, courtesy of google, show you a large set of pictures of the town hall you'd see if you went to Sheffield.



It's not just the autumn trees or the warm color tone. The genAI image has the clock tower in a different place with respect to the other buildings; the windows and floors are different from the real building. Just about every architectural detail in the genAI building is wrong, although it has a similar feel overall.

I have many hundreds of examples from Adobe Stock of this sort of problem - the town hall image isn't an exception. There are images from cities all over the world; of objects that are impossible or defy the laws of physics (but which are presented as real life not fantasy); human or animal anatomy, vegetable/fruit structure that are wrong (not ironic or artistic or fantastic; just wrong); flags or other emblems improperly rendered.

As long as this type of work is clearly labeled as genAI, buyers can choose. IMO Adobe would be well served by removing the obvious "oops" images. Although perhaps the aircraft carrier with wings should stay

Another general topic for contributors to contemplate - genAI images can end up looking very much alike.

I was looking at a set of images in today's new approvals and thinking there was something wrong with the sort order - that I was seeing old images from a couple of weeks ago.

Turns out there are two separate portfolios each with a batch of very, very similar sunset-beach-palm-tree pictures. No copying, just whatever tool they used has apparently only got so many ways to deliver results.

Older image batch - Today's image batch

I'm not trying to argue that only one contributor can have sunset-beach-palm-tree images, but do think it's worth considering what value to the buyer there is in having such nearly-cloned images. I also wouldn't argue with the composition of these images - they're all very pretty. And there are none of the heffalump traps that genAI so often falls into in the subject matter.

If you imagine a number of photographers going to a particular beach to shoot a pretty sunset on different days at different times of the year - or even the same photographer going back multiple times - I can't imagine them producing such homogenous results as these two sets of images. Our planet's weather systems and land-ocean interfaces are highly variable - it would be very hard to get the same result twice if that was what you wanted.

Does this mean that with the current data set training, at some point we'll be getting more repeats (for similar prompts)? I know nothing about these two contributors, so it is possible that this is some sort of collaboration and this is all really one big set of images that was divvied up and uploaded separately.

Jo Ann Snover - Great post ... 

Turns out there are two separate portfolios each with a batch of very, very similar sunset-beach-palm-tree pictures. No copying, just whatever tool they used has apparently only got so many ways to deliver results.

My guess is that these two accounts belong to the same contributor. Probably using multiple accounts to get around upload limits.

But they have to realize that sloppy reviewing floods their database with junk, and in the end, this doesn't do any good for their customers either.
In the end, they pass the reviewing over to their customers.: bad content doesn't sell and gets buried by the search engine algorithm.
That system works fine for the occasional error that reviewers make, but if you flood your database with junk, nobody will bother anymore to wade through it.

I remember some time ago in the past when reviewers at AS were really strict and fussy and it was hard to get some content accepted. In more recent times, they seem to be getting less strict and it's not often when I get rejections these days.

The collection is now over 13.5 million. Just a handful of the recent crazy acceptances...

Even if I were taller, I don't know how I'd get into this chair...


The basics of a bicycle really aren't that complicated


Customer service: Why are we getting all these complaints about newly-arrived goods being soaking wet?


How dare you start commenting on my extra legs!


This clearly isn't cheesecake, but even as pasta I wouldn't put that in my mouth!


This is closer, but also doesn't look like "cheesecake with berries and straberry"


It's all completely mad!