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.

[CShubert]

Hi @Jason Toemmes we are sorry this wasn't made clear.  If you click on the beaker icon in the upper right, it will open the info screen. When you click on the Generative Fill section you will have the ability to share your feedback there.

[VeneficusFerox]

I have similar issues.

I create a fully new image from a blank canvas. Then I use the lasso tool to select a part to replace/improve with or without prompt. Then it tells me that the image violates policy and has been removed. As the base image is still there, I can only conclude that the smaller filled part was removed directly after generation. Why would it be able to produce something against its own policy?

[Kevin Stohlmeyer]

Hi @VeneficusFerox as mentioned in here extensively, try entering a prompt to bypass the false error.

[daniellei4510]

"The reason why these tools tend to fail on hands and feet is because these AI companies tend to steer clear of human anatomy (inherently 'evil'). Meanwhile if you learn real art this is the foundation where you start. "

You wouldn't know it from the number of images of amply chested women I see often on AI platforms. No, the real reason is that AI has difficulties with hands and feet for the same reason that real artists often find them difficult to draw or paint. They are very complex parts of our anatomy. At this time, Midjourney usually (I emphasise usually) does a very good job with hands and fingers. With feet, not so much. 

[daniellei4510]

I said it befor, I'll say it again. It's not the user violating the guidelines...it's the bot refusing to accept its own results. If the error message were to read something along the lines of, "the bot has generated results that are in violation of user guidelines and cannot be displayed," I think we'd have fewer instances of complaints over here. While I've received numerous violation notices, generative fill to date has saved a least a dozen of my images that would have taken hours to fix, if I bothered to fix them at all. Be patient, people. 

[Matthew249224202mcs]

The concern for me at least is that it doesn't look like a bug as much as it suggests that "community standards" censorship is a feature.

[shamde]

Thanks Kevin! If only forums could have multiple pinned comments or worked like reddit with upvotes you might not have to repeat yourself so much! Worked fer

[colliemom01]

I'm getting the same error with a photo of my mom looking into a vanity mirror (wedding photos) - there's nothing offensive about it. Happy to provide photo for investigation just don't want to post it publically.

[sincere_Songbird16B7]

While I understand that Adobe dosn't want Generative Fill to be used to generate adult oriented content for obvious reasons, it is not quite clear to me if it's permitted to be used on on photographs containing nudity. Specifically for object removal in the background.

[Graham24508943nobd]

Seems to me like they need a hell of a lot of bug fixing to to do. Most of my generations are pretty crap with the subject NOT keeping within the selection

[garyblakemeister]

Is the policy going forward strict moderation of tools?  

What if I use your tools to create something Adobe deems unsavory or problematic?  Do they reserve the right to delete my PSD file? This was always the fear with the cloud-based model and it's a bummer to actually see lawyers dictating creative policy.  

You should strongly considering taking off the handcuffs and letting creatives make what they want to make.  I don't trust the Adobe Council of Ethics to rule properly on my creative work.

[Matthew249224202mcs]

Additionally, I do a lot of fashion, boudoir, and art nudes... Why do I have to adhere to the lowest common denominator when it comes to using the tools I pay for in an entirely legal manner to support my business?   

I can totally get behind making it more difficult for anything that involves body swapping, or significantly modifying bodies.  Or taking a dressed subject and swapping them into nudity or near enough...

But if you're starting with a subject that's already in lingerie, it seems we've already gone past the "community standards" and shouldn't have to fight with Adobe to let us work.

[Fran30362383lc5u]

I have discovered that anything in the tonal range of skin is deemed inappropriate. I tried to remove a tattoo from a person of colour, hair across the face of a caucasian girl and extend the skirt of a peach coloured tulle skirt (in the colour range of skin tone) and all were refused. I think the puritans running the system need to consider that skin appears in any image with a human it it and skin in itself is not bad!

[MontyWales]

I was just changing some skin blemishes and getting the error. But by putting just the letter S in the text box it worked OK 

[LiveAndLetLive]

Dear god this response is far more worrying to me than any AI would or could ever be. 

Stop buying into every fear campaign that you hear or see online. 
As somebody who has trained and worked with dozens of models (that's Machine Learning speak for the collection of algorithms that people use to "Make AI go brrrrr" I can't help but chuckle when I see people losing their minds over it. 

Listen, all the big execs calling for memorandums and all that other jazz know the truth of what's up as well, the only reason they are trying to get you all scared is because they want to keep the power out of the hands of the people, and in their hands.
Also, money.

(something tells me you might be one of those kinds of people that actually hears that and thinks it's a good thing). 

Smh. 😕 

[Veronika30330093ichg]

Now that you mention it, I am editing things that are sort of skin colored. The other thing I noticed today is that it seems to give me a really hard time when the selection is oblong and "skin" colored. What's really funny is that yesterday I edited a piece of crown molding that was kind of shaped like that. After fighting with it for a while, it filled alright .. looked like a colorful version of the very thing it's trying to prevent me from generating. lol

[Veronika30330093ichg]

I'm not here enough to know if this is being discussed or not, but entering a prompt is hit or miss for me.  It worked a few times tonight, then absolutely refused no matter what I said.  I was editing a building with not a person in site. 

[LiveAndLetLive]

It is, imo, 100% ridiculous. 

The premise only makes sense if the material is being shared, and then it should fall under the same guidelines as any other image being shown to others.
An adult should be able to make anything they like (barring the obvious illegal content as AI can control for it while other more 'traditional' apps cannot obviously) - sharing those images is a different story.

The real reason these companies censor it, is because they don't want the inevitable "Look what Adobe/Photoshop let's you make!!!" videos and posts that will crop up all over the internet, as it can be a PR nightmare, and I understand completely why they wouldn't want that; but if the same restrictions apply on release a lot of people will just turn to other tools leaving PS in the dust, as AI is the future of art in so many ways. 

As it is I'm not too concerned about it myself as there are other tools out there - and I've already made my decision to cancel my subscription (due to what I see as unethical policies in other areas) - but the argument that people shouldn't be allowed to make (legal) things based on someone elses ideas of what constiutes morality is absolutely ridiculous (and shameful in my opinion). 

Again, these are just my ideas (at least I can say that, so many others seem to miss that point while instilling their own ideas of right/wrong on others) and I don't mean to be rude about it, but the people who would police others in the privacy of their own homes aren't doing the good that they think they are.

[CAVelarde]

Photoshop Beta v24.6.0

MacBook Pro 13 in, M1, 2020

Mac OS Ventura 13.4

I just typed "tesseract" on a blank canvas expecting the usual 4-dimensional cube. Instead I got the message that said the image violated guidelines.

[and4657723]

90% of the images I try to generate end up with "The images violated user guidelines" (see attached image).

The error pops up so often that the AI generative fill is pretty much unusable at this point.

[Bob5E9A]

The gen fill engine appears to be analyzing the request and doing a drawing. Then it determines what it drew is x-rated, or not permitted, so the warning pops up. I tried to change the curving lines on a child's cartoon of a big bears paw and I am assuming it drew a curvy nude butt and rejected it's own drawing. The only answer will be that instead of this rejection, we get a pop up screen where we actually describe what the image is and what we want. Then Adobe uses Chat GPT (no kidding)  to analyze what we said and create a prompt. I saw a YouTube video where a guy did simple, easy training for Chat GPT to write brilliant prompts for Mid Journey. The problem with this beta gen fill is it does not write good prompts for itself because it is jumping to conclusions and needs the users help. A flat out rejection like it is giving does not help.

[c.pfaffenbichler]

Please discuss Photoshop Beta features on the Forum dedicated to Photoshop Beta: 

https://community.adobe.com/t5/photoshop-beta/ct-p/ct-photoshop-beta?page=1&sort=latest_replies&lang...

[Graham24508943nobd]

As I've mentioned before, it'll give you a guideline error even if NOTHING is entered in the prompt box. I've tried removing a stone from one of my own pictures taken last week in woods nearby and even THAT got rejected. Just wish they'd get the bugs ironed out soon before I lose my head with it and go back to using the old PS 2023

[slaughters85j]

I apologize in advance for those that know this already:

All Adobe has done has incorporated Inpainting and Outpainting with the option to weight each (prompt/no prompt) like literally hudreds of other applications are doing right now, spun-up largely due to the whole AI craze; and Adobe's marketing team has dubbed this Generative AI.  Look here to see when OpenAI debued outpainting last August, look familiar? https://openai.com/blog/dall-e-introducing-outpainting (click the time lapse)

Without nerding-out too much, all of these prompt-to-text, inpainting, and outpainting features popping up are all made possible by iterative models trained on very large datasets of original imagery, styles, content, etc. (source of the original models everything is based-off stability AI's models.  Unless you build an original model namely stable-diffusion 1.x - 2.1, which is very time and cost intensive, then your stuck relying on the dataset and controls that the original was made with.  *stable-diffusion is the OpenAI Chat GPT Natural Language Processor service equivalent for exclusive image generation.  As such, there are a finite amount of controls applicaiton producers can place on their respective services.  I would place Mid Journey's (perhaps the most widely known and strait forward ease of use service) service as about as restrictive as Adobe's, Mid Journey is more text to image and image-to-image exclusive (weighted by text/weighted by an existing image). *there are plenty of others for both Mac and Widnows but do require a greater learning curve.  The results of these others can be fed as input images into Adobe PS but require additional steps in your pipeline.

For Adobe, there are really three'ish zones, fully restrictive (I think the mutual consensus is likely that this is where the current beta can be placed), the mid gray zone area (where some controls are in-place to restrict a good chunk of what most would consider in-appropriate content), and then the lower zone being full and open, where the only 'controls' are via trusted ethical notes from the model creators.  Adobe needs to implement controls that bring the their Generative AI service down to the gray zone, with the understanding that no service will be able to prevent mis-use completley.  Regardless, there are many other options out there, none of which, including Adobe, have breadcrumbs that lead back to which application or service was used to create resulting images.  If Adobe wants complete control then they will need to build an original dataset that is solely owned and controlled by Adobe - if this is what they've done, then they need to train it further with gray zone content - otherwise user will and can (for free) use other services and Adobe will find themselves trailing behind.

Since launching their Beta, I haven't seen any iterative improvements to the current version.  Not how a Beta is suppose to work, ....Adobe developers?

[Graham24508943nobd]

Yes you're spot on of course, a beta gets the bugs ironed out before the next update and the testers are informed what fixes have been done. Adobe, you REALLY DO need to intereact more with us to get the job done better