P: Generated images violate user guidelines

[Hematita]

So as you can see, it's a PG-13 relatively inoffensive image of a woman in a bunny outfit. The top worked fine, and I was able to complete the top ear, which is cool. When I tried to extend the bottom with generative fill, though, I got this warning. They're just a pair of legs wearing stockings, and I wanted to extend it.

It feels like a false flag - though I could be wrong? I find myself thinking it would do the same for women in swimsuits.

Figured I'd share here.

[Blondinko]

@Pete.Green does your team still need more examples of non relevant warnings about guidelines violation? I get dozens of them every day, for nothing, when trying to generate some blurry backgrounds on my macro shots. I will be more than happy to provide examples if it will help to improve the tool. The generative fill is extremely needed and welcomed in macro photography and currently it is really painful and frustrating to use it and to get this warning in 90% of cases.

[Donkeet]

I get a ton of those banners for completely innocuous prompts. Given that it is being beta tested, I'm sure the team is working on fixing these kinks.

[neart45822646]

Wenn man bei Generative Fill den Text-Prompt "remove moire" eingibt - um zu testen, ob PS vielleicht schon in der Lage ist, ein solches Kommando zu verstehen, wird das mit folgender Warnmeldung quittiert: "The generated images were removed because they violate user guidelines". Dies ist offensichtlich ein Fehler und nicht zutreffend. Es wäre mir jedenfalls neu, dass der Moiré-Effekt irgendetwas Anstößiges oder die Guidelines Verletzendes bedeuten könnte – auch wenn man das semantische Umfeld und Fehldeutungen in Betracht zieht.
Wichtig:
Es wäre gut, in solchen und ähnlichen Fällen eine Korrektur- oder Einspruchs-Funktion zu implementieren!

Und da Verletzungen der Guidelines potenziell mit juristischen Konsequenzen für den Nutzer verbunden sind, ist es unbedingt nötig, dem Nutzer mitzuteilen, worin in den Augen der AI der Verstoß besteht. Stichwort Explainable AI.

Anbei ein Screenshot des Bugs.

[PS Beta Version: 24.7 m2223, MacOS 13.4.1, App&System-Sprache ist Englisch]

[bradleyalphabet]

They need to remove the training wheels. Not paying a premium for Adobe to dictate what art is acceptable. I'm trying to clean up a basic head-and-shoulders portrait and it's being flagged, wtaf.

[Donkeet]

Well you're not paying a premium for this because it is still in beta...

[CAVelarde]

The term "balerina en pointe" gives me the message that it violates guidelines. I tried several different canvas sizes and proportions with the same result.

[Bob5E9A]

Donkeet,

You said, "Well you're not paying a premium for this because it is still in beta..."

Oh yea? I have great alternate graphics software but signed up just for their "super" (lol) Firefly and gen fill beta. Like other AI image beta generators they should work.

So every dime I spend, is a premium. And I will pay a huge penalty to cancel and rebuy when it is fixed.

[Graham24508943nobd]

I would hope that testers will get a BIG discount off their next subscription as a thank you, like most other companies do

[sebastianr15198647]

I will try it 

[lagedor222]

I had exactly the same with a flower, trying to remove highlights:

 

[rickiep73948756]

Hello, 

App Version 24.7.0

macOS Monterey 12.7.6

I make Day of the Dead Skeleton images and use generative fill to do so. It seems that everytime I want to fix a skeleton arm or leg or something, even a tree limb, I guess the AI thinks its a certain male appendage. I do not use the beta for that purpose. Perhaps the AI is too sensitive, I don't know, but I am not violating community guidelines. This, perhaps is not a bug, but it is very annoying.

In the attached files, I added the background and wanted to fix the female left arm at the elbow, it took no less than 5 tries to do so because the AI perceived it as a certain male appendage, in my opinion. This happens a lot when I am fixing issues with the arms and legs, sometimes fingers, tree limbs, etc.

Even trying to send this post it said the use of the word describing a certain male appendage was a "bad word." and I attached that screen shot as well. I am just trying to be as polite as possible when describing the issue to you. Please, do not take offense to this, as I am just bringing the issue to your attention. No offense intended.

Thank you

Rickie Pauley

[MontyWales]

When it works it works

[Alan27596961g40h]

Through a number of experimentations, I think I have a grip on how these restrictions are being assessed. In one example, I had a photo of a woman wearing a tight yoga outfit where the top of her head was cropped. So I tried using generative fill to build the top of the head. I kept getting the policy restriction errors. So I started experimenting by cropping parts of the image. First I cropped the lower part of the body. Still getting errors. Cropped to the neck. Still getting errors. I even went as far as cropping so only her face from the nose up was left. Still getting errors. I quit and re-ran the beta version every time in case Adobe remembered the original image before the cropping.

I think how Adobe AI works is that it surfs the Internet for billions of images and saves them in a database somewhere. Google does this, as well as many other search engines. Then the images are processed using algorithms and are then categorized accordingly. I'm pretty sure these categories include the "naughty" category. So when generative fill starts to process your request, it builds the imagery on the server end. If the produced images include parts previously assessed as dangerous, then the errors will keep popping up as long as you use that same image. The results generated by AI have to be 100 percent clean of violations. So in my unfortunate situation, any images found on the Internet that somehow resembles the photo I have that have been tagged as pornography or the like will trigger the violations.

In conclusion, there's nothing wrong with my photo. Adobe's image scanning technology scanned in too many pornography images. What Adobe should do then is flush out the bad imagery from its database. So this way I won't get these violation errors. I'd rather get a lousy result or no result at all than flash a violation error as if I'm some pervert trying to produce something bad.

[Edu Couchez]

So... with this violation of “guidelines”, sure we hadn't had any of the beautiful Monty Python's movies. Really... these days, playing with your AI, I've felt how the old eastern communist block faced the strong soviet's censorship. I know you have to create a safe environment for the human being, but at cost of the creativity? You're developing a creative tool, not a dogmatic ideologic 1984-esque censorship paradigm. At the end, we're sharping our minds to fool the inner AI censorship to get what we want, as it happens in totalitarian societies. Today for example… hours to generate a decent Prussian-war outfit for a steampunk friend’s picture. Needed lies and tricks to convince AI I didn’t want to harm anyone.

So, Adobe users think we are mature enough to face our own responsibilities. If I can google the most recent Ukraine news and videos, why can I create a Prussian cavalry costume over somebody, because I wanted to add a sword in his hand?

And please, don’t take that as a “heater” proclaim… only as a personal philosophic though in loud voice.

[Stapfoto]

Maybe seen as niple? Refused to fill over and over again.
Also filling the space between cabinet doors is refused because "against the rules". Gereative fill doesn't work with most of the pictures I shoot in an office because of this reason. 

[Karim98A7]

For some reason this feels like an AI-generated comment. Been tinkering a lot with them, and I can recognize an AI when I see one.

[defaultaad2ywvie63z]

I wanted to add a beachbar to the photo of a beach and sea, and used only the prompt "beachbar". However I get a message that the beachbar was not generated because it violated the guidelines. This is really odd! On what base does it violate the guidelines? Can you clarify this?

[wizmosis]

I think the prudish community guidelines have a 'thing' about legs! I wanted to recreate something I saw while swimming so I used a photo of the pool underwater, and asked the prompt to fill it with "plump pale legs underwater." Nothing salacious about my concept. It's something I see everyday and find very beautiful (I'm female by the way) but no, "Guideline violation" 

[jdavidbuerk]

@bradleyalphabet That's a great analogy; remove the training wheels - beating a dead horse at this point, but it's an unacceptable level of restriction / censorship (I hate using that word in this context but I think the concept is applicable).  Many of us shoot primarily portraits, most is totally unremarkable, some of which is boudoir, and some may even be spicier - it's nothing new, so it's silly to think that this should be a problem only now.

@lagedor222 You're right, there's zero reason Generative AI should be flagging that, however the Content-Aware Heal tool, or new beta Remove Tool (which works fantastic, btw, barring the occasional (internet / server?) delay) would be better for.

Also, do we know if the Generative AI will be outputting in higher resolution once it's out of beta?  I read on a 3rd party website that it's limited to 1024px maximum resolution currently, which obviously is tiny by today's standards.  While that works fine for many use-cases, that doesn't work for many others.  I've of course tried to keep working areas smaller with this in mind, but the average user when it comes out of beta will not be aware of, or understand, this limitation.  Another important limitation to keep in mind is this resolution limitation also prevents the Generative AI tool from properly matching many input's grain.  I'm shooting on an R5, which is 45MP SOOC, or 180MP if you've used Super Resolution (which is what I do for headshots).  At that resolution, Generative AI won't keep up with the natural grain of the camera even at 45MP for even just a medium sized area.  I've gotten around that limitation by faking it by matching the grain by applying a Camera RAW filter with matching grain (Amount: 25, Size: 15, for the average R5 file) onto the Generative AI's Smart Layer, and that does the trick most of the time, but it's another issue that needs to be looked at by Adobe before this is rolled out for primetime, becuase that's definitely a limitation, and it's something that nobody but powerusers like most of us here in this thread will be knowlegeable enough to work with.

[Rafael_Castano]

no puedo utilizar el relleno regenerativo en ningun caso, me sale siempre este mensaje,

¿Como puedo cambiarlo?

[Graham24508943nobd]

So much for the latest update. I was trying to create a fantasy forest. The words fairy and pixie violate user guidelines. GIVE ME A BREAK !!

[c.pfaffenbichler]

So much for the latest update. I was trying to create a fantasy forest. The words fairy and pixie violate user guidelines. GIVE ME A BREAK !!

»pixies in the woods« for example seems to work here … the results seem moderately useful but no guideline violation-alert. Maybe the plural makes a difference? 

[Vitalii1]

your AI algorithm restrict Generative Fill results  Ionly try to remove tatoo from the arm of my model

[Graham24508943nobd]

Will try, shame as almost every fairy generation is totally disfigured no matter how I spell it, I tried pegasus that got refused so I used a capital P that got through fine but the generations were completely unusable and really ugly

[Graham24508943nobd]

they're a long way off finished by the look of it