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.

[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

[Leslie Moak Murray]

She has six fingers on her left hand and the inner bend of the elbow line on the raised arm is wrong.  A few more "AI hands":

[Jason Toemmes]

Wish I could upvote your comment 10 more times

[Olivier2023242526]

We are adults responsible for our actions. The courts decide right and wrong, not the manufacturers. If I puncture a car's tires with a fork, should the manufacturer be forced to blunt them?

[randyo83115211]

The restrictions are too much, adobe you need to loosen these [cursing removed] restrictions.....all the images i try to work with in the beta keep coming up with error about community guidelines and all my images i try are of fully clothed women and its frustrating. please work on this, better yet quit restricting them in the first place, im a boudour photographer. Please dont be like the "other" and censor what we can edit of our own images.

[J-D-F]

Adobe isn't going to improve interaction or the tool. They will add their me too ai to the next release and then subsequently pretend that fashion, swimsuits, glamour photography, lingerie, sports and fitness don't exist as markets. All of these censored AI tools will be totally eclipsed by tools without guidelines. That is how it has worked for every previous generation of digital tools, Adobe wants to pretend that corn ads aren't 80% of their userbase, or that flat-style clipart and fonts will save them.

Somehow this monstrosity made it through the filters because obviously a cluster of mutant flesh bulbs is more appropriate than the nethers of a statue. 

[J-D-F]

Incidentally, we can't even discuss problems because culture is so puritanical we censor words like gen i tals and pe nis, which are legitimate real words, aren't slang, and describe something directly related to the issue. 

[kevin stohlmeyer]

You doing ok @J-D-F? You seem really upset over a first generation public beta.

[CoryShubert]

Hey everybody.

We appreciate all the feedback and patience as we work with this exciting tool, and our goal will be to release a finished product that meets all the needs of professionals. In this time of learning and growth, be assured we hear you and are taking all the feedback to the team to sift through.

[daniellei4510]

No one actually knows WHAT the bot is generating. Everyone seems to assume it's nudity. Or violence. In fact, it could be generating such visual gibberish that the bot doesn't even know how to interpret it, so it's only resource is to assume it's against the guidelines. And as far as this issue being fixed soon, don't hold your breath. It took the programmers of Midjourney a year to finally get it to create reasonable looking hands. And they still aren't 100% perfect.

Meanwhile, two weeks ago, Photoshop didn't even HAVE this technology and now everyone is getting upset that it doesn't work more often than it does. But no good deed goes unpunished, as they say. 

[RED444222]

Hello, I am using the new beta generative feature and I am finding that it is not adhering to the prompts I am using and I belive it is a censorship issue, I work in making swimwear and lingerie and that sometimes involves plus sized models and it appears like the images of the bikinis I make are for some reason violating the guidelines when there is no pornographic content in the image. Is swimwear considered breaking the guidelines? Is being overweight in violation of the guidelines? Is there a problem with showing too much skin on the system where it wont adhere to the edit even if it is not considered pornographic although a bit revealing?

Overall its a great concept and Ive seen some great work done with the service but very restrictive to the industry I am in which makes me very sad that even in this day and age we cant show anything above the knee

The ai was actually putting new clothes in all of the images I was attempting to update. lol.

[kevin stohlmeyer]

Hi @RED444222 can you supply some examples of images, results and the prompts used to show how it's failing? It would help illustrate the issue.

[J-D-F]

Doing great! You seem upset that I have an opinion regarding a poorly implemented late-to-market feature-clone? Have a good day! ❤️ 😘

[Vito27662922on8d]

Naw, I'm thinking at this point it's to stop people from taking a model's picture, lassoing the body and telling it to Generate a nude body. My problem was the opposite. I was trying to PUT bikinis on to nude models so I can put them on Social Media.