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.

[daniellei4510]

Sporran works fine on adobe.firefly.com. It is not a banned word.

[Graham24508943nobd]

well it doesn't work at all on the beta

[Kevin Stohlmeyer]

@Graham24508943nobd Generative fill is US English only at this time. Sporran is a Scottish word.

[davescm]

'Sporran' on its own resulted in a violation but 'sporran on kilt' was accepted here

Dave

[Graham24508943nobd]

Thanks for that Dave, big help !   I just imported one from an old image I had removed background and used that. Didn't faff about as much as I could have with this of course.  The two lassies had horrible defects so it took ages tweaking, the hands were the worst to correct.

 

[jdavidbuerk]

Reminder to those following this issue; it's been said quite a few times already, but there seems to be some confusion about the "violation."  This generally is not an issue of the words you're using being "censored," but instead it's the AI's output image that is flagging false positives via the overly-restrictive image monitoring that is in place prior to the output being released to a user (likely due to CSAM concerns, but as obviously demonstrated, the filtering is much too strong for typical usage).  Yes, a word filter is surely in place, however that's not what's causing the problem for most cases shown in this thread, where people are inputting innocuous prompts such as "wall," "arm," "background," and so forth.  The reason changing your input terms will frequently help avoid the percieved violation is not becuase of the words input, it's becuase the AI's output has changed and is no longer being internally flagged.  This is the same reason even a blank input prompt on an innocuous image can still be flagged as "inappropriate."

I think most people talking about the Generative AI problems agree that the filtering in place is currently much too stringint for regular, practical use, so this is the real core issue at hand: balancing useful flexibility and expression using the AI with reasonable safety and legality.

My reccomendation to solve the problem of balancing artists' freedom with safety from nefarious use is to employ a digital watermarking function linked to a user's Adobe ID (and presumably the history of their use of Generative AI) on files that have been output using Generative AI, and remove all the server-side filtering except for input keywords of illegal acts.  This way, users are not limited in their use of Generative AI to much more freely work on their images, and Adobe is free from the liability that their service is not being used for illegal activity.  And should someone be using the software to create illegal images, real or fictional or otherwise, the embedded watermark would be able to be looked up by Adobe, their access to Adobe services banned, and their identity shared with the proper authorities.  My understanding is such watermarking technology has existed for quite a long time, and can be employed in a combination of both embedded metadata (which Photoshop of course already does), and inperceptible pixel patterns embedded in exported images (essentially like an imperceptible QR code); printers and scanners for example include similar such countermeasures to trace criminal activity, such as forgery.

[Gregusaurus]

That is one frightening recommendation. For a watermark to be implanted that can be used by the "authorities" to track down the artist who dared to create something that is deemed "against the rules" is so horrible that I am shocked it is being proposed.  Maybe you are being sarcastic? 

[jdavidbuerk]

@Gregusaurus It's not outlandish at all.  Such digital watermarking utilities and services already exist for still images, video, and audio, and are commonly used for tracing and enforcing copyright protections.  "Against the rules?"  Yes, the rule of law - the reason the filtering is in place at all is to prevent laws from being broken; Adobe has to protect itself from liability of their services being used for illegal activity, and I hardly think it should be controversial to pursue people creating illegal imagery.

[Eric30148995ldmy]

Question: Why is there one stringent standard for using generative AI to edit vs say manually editing your photos to add new elements - or even just using content-aware fill?

Photoshop has never, ever had restrictions for usage and has never had tracking implemented for editing. Now that generative AI is a new additional tool for editing, all of a sudden now there needs to be restrictions and monitoring? But those restrictions and monitoring disappear the minute you decide to edit by hand instead? The logic doesn't make sense. It amounts to saying because the editing tools got easier to use there needs to be a restriction.

Having guidelines dictating how you're allowed to use photoshop is unheard of in the history of using photoshop. If the warning is popping up because the software uses it as a crutch to hide the fact that it's incapable of generating an image that's one thing. But if it's popping up because a decision has been made that devs will now have greater control over the type of images that are allowed to be edited via photoshop I think it goes against the very purpose of this software.

[Graham24508943nobd]

🤣🤣🤣

[J-D-F]

It's hilarious that Adobe is triggered by Ai nudity in photo editing software to the point of absurdity for content warnings, but completely unconcerned by copyright infringement from the same tool. 

[jane-e]

@J-D-F  Wrote:

It's hilarious that Adobe is triggered by Ai nudity in photo editing software to the point of absurdity for content warnings, but completely unconcerned by copyright infringement from the same tool. 

The two points that you bring up are addressed in #1 and #3 of the user guidelines:

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

Please read the entire page — it's short.

Also note that while GF is in beta, it is for personal use only and cannot be used commercially. (#4)

[daniellei4510]

Take the time to do a search on stock.adobe.com for nudes. There are plenty of results. If this is in fact a censorship issue, there are other factors involved here that involve very complicated programming issues.  This is a beta. 

Also, with respect to copyright issues--and as I understand it, since a lot is still unknown or up in the air right now--Adobe's eventual plans are to only allow uncopyrighted material and stock contributors' images to be used for AI, for which the latter will eventually be compensated. Stock contributors who don't agree with this can simply stop contributing to Adobe Stock. 

[daniellei4510]

Say what you will, but Generative Fill in Photoshop is breaking new ground and I suspect it will be another six months to a year before it is released in the "working" version of Photoshop. As a retired photographer (fine art nudes) and now an AI designer well aware of the short-comings of AI (no AI result is ever perfect), I have been able to save at least two to three dozen AI images that I would have tossed had it not been for Generative Fill and other AI applications. I've been using Photoshop since it was first released and while I'm fairly good at editing and retouching, there are occasional results beyond my expertise (both with respect to the time it would take to fix the issue, and whether it is worth my time to do so even if I could). 

I've attached an image that I believe, for now, Generative Fill SHOULD be used. Not to create entirely new subjects or add objects that were not in the original image to begin with (use a dedicated AI application for those purposes), but as yet another editing tool within Photoshop.

I've said it elsewhere and I'll say it again: Forget what YouTubers are showing you about how powerful Generative Fill is. They are cherry picking their images and editing the daylights out of failed results to gain clicks. Or just being extremely lucky.

Anyway, here is the image in question. Generative AI fixed the rose in her hat, and extended the top and bottom of the image (Oh! And look! Skin was involved!) I also did some traditional Photoshop editing with regard to skin tones, the brightness of the eyes, mild vignetting, etc. But you get my point. I hope. Be patient. It's a beta.

[Graham24508943nobd]

Imagine my surprise when THIS generated for me a couple of days ago ?  Nose & mouth were badly distorted but I left everything else apart from the background which suits her outfit. I typed in beautiful woman posing and got this.

 

[JR Boulay]

The n*pples are missing!

😉

[Graham24508943nobd]

😆. maybe it's non binary 

[YipeePhot]

I just had a similar problem while repairing an old photo of myself and my younger brother and sister. I was using generative fill to replace a missing part of the photo. The photo is just a photo of the 3 of us sitting "fully clothed" in front of our old house.

[vketteniss]

This had me going crazy for days while trying to outpaint 100 images for a project on a deadline.

TODAY THE PENNY DROPPED! It's the skin colour...! So here's the hack:

Invert image >> skin goes blue >> do your in- or outpainting (without guideline warnings)  >> invert image again >> TADAAA!!!

<Moderator Removed link>

[daniellei4510]

I think it's even more complicated than this. When I have a rare issue with Generative Fill that I am unable to find a work-around for, I head over to firefly.adobe.com. You are correct. If I enter the prompt, "An old man smoking a cigar," I am told I am using a word that violates user guidelines. But if I prompt, "An old man smoking a cigarette," I get four images of four old men smoking a cigarette. So much for the issue being with regard to self-harm.

[Immaculens]

How does this violate a user guideline? I made a selection around a present to remove...

[stephen291894531ul1]

Well, at least you got some good results... I just prompted for a belly button and received an infringement notice. The AI made me suspect it was possessed by the ghost of a 1750 Puritan inquisitor. So I thought, let's just "select the stomach area" and leave the prompt empty. Logically, the AI will complete and add the missing navel. I'm amaze by the propose result almost as much has how these systems can even infringe upon themselves. So out of the three options, it only kept one, without a navel of course but at least, my theory about the Puritan ghost has now been upgraded to:maybe the  reincarnation of a photobombing computer geek freed from when the singularity was on floppy disk.....  🙂

[Graham24508943nobd]

Yesterday on adobe firefly.com I prompted for a cowgirl country singer. out of 16 generations more than half were men. NOT a good reslt by any means. I DO find the qulity online slightly better for humans although too many have that artificial plastic look to them.

[Kevin Stohlmeyer]

What prompt did you enter?

[Graham24508943nobd]

try the period or full stop at the end of your prompt.