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.

[Sarah27914909k9u3]

Please, PLEASE loosen up the community guidelines for generative fill. The current system finds all kinds of PG or G-rated items to violate community standards. It's ridiculous and makes the feature much less useful, which is a shame, because it's an awesome feature.

On another note, once the feature leaves beta, I do truly hope all the censorship will depart as well. I pay a lot of money for the creative suite, and I do not do this so I can be thwarted and judged by Adobe.

[Michael Bullo]

What kind of keywords are getting flagged?

[daniellei4510]

Not sure about that. I think a lot of what photographers and AI creators are doing involve women, so it comes across that way. I've had the same thing happen with blank areas, cows, babies, you name it. I don't believe for a minute that this is the issue.  

[daniellei4510]

Do you, as I do, get the impression that people POST to this board and never really READ the comments and responses? Do people even understand what a beata is about? For starters, this is an editing tool, not a means to create new images. Yeah, youtubers are showing all these crazy instances about removing entire groups of people or replacing a person holding one thing and then holding something completely different, but trust me: they pick and choose in advance what works and what dosen't. Everyone needs to calm down. This technology didn't exist a couple of weeks ago and now you want absolute perfection on a freaking beta. If you don't like it, stop using it. Cancel your Photoshop membership. Go back to doing it manually, if you even adept enough to do so.

[Bob5E9A]

After many, many experiments with this beta gen fill and the constant weird fills, I personally choose to believe that I get this message because gen fill tried to created a fill and failed. It does not see the object I am trying to fill, just pixels. Since it does not actually see my object with eyes, it does not understand the scale or connection point, etc. Using a prompt gets an image related to the prompt message but not related to my object. Frankly unless gen fill grows eyes and a human brand, don't know how PS can fix this, except to build in some type of user discussion interface tool which interfaces with logic software, like Chat GPT, to create the perfect prompt. The first step in this beta would be for PS to build a discussion interface and let us know what prompt it created so we can modify it.

[daniellei4510]

Generative Fill DOES work, despite its present drawbacks. The mother of the child in this image I've attached gave me permission to post this. After her baby was born, he required the use of a feeding tube. While she shared photos of her new baby with relatives, she told them to NOT share online any of the baby's photos that included the feeding tube, which in fact it must wear 24/7 for the time being. She asked if I could help, since she knew I was a photo retoucher. To be honest, I'm a pretty good one, but removing the tube using Photoshop would have required a good deal of time, espeically since she wanted a number of images to be retouched. Fortunately, Generative Fill had just been announced and I downloaded the beta to give it a try.

Attached is one example of how incredible this new technology is. I simply selected the bandage on his cheek, along with the orange tubing, and selected Generative Fill, without a text prompt. It worked flawlessly on all four images the mother provided me with and she was finally able to give the go-ahead to relatives and inlaws to post images of her newborn.

Generative Fill works. Not always, but it works. Just don't try to over-extend its capabilities. And don't blame skin as being cause of the issue. It simply isn't true. 

[georges73572917]

AI gave me a naked demon standing in the picture, but a closed mouth is a no-no.lol

[Bob5E9A]

I once worked for one software company. Another company, who was always first to market, tried to recruite me. I commented of them being out first so often. The person said they release a product even though it is not ready, but it makes them look cutting edge. PS was so far behind Mid Journey, plus AI is "big", so they had to get something out. Will they be able to take a baby picture, like above, without a tube and use gen fill to put a tube in place, starting with no tube? Don't see how that could happen.

[Bob5E9A]

OK. Now take the baby picture on the right with no tube and have Gen Fill create a tube and put it in place. Yes, you can use prompts.

[daniellei4510]

I put up an example. Admittably far from a perfect one, but with some photoshopping finess, it could probably pass as real. In any case, my original comparison post was to show what generative fill CAN do, not what it can't, and to falsify the notion that skin has something to do with the false flag warnings. That simply isn't the case. 

[Ray26434972uq1g]

All I wanted to do was have generative fill complete the elbow, and it gave me the dreaded "violated user guideline." It works excellently with landscapes, but the generative fill is very cautious about body parts.

[daniellei4510]

One thing I discovered, quite by accident, is that if you make a very small selection anywhere else in the image far removed from your main selection, this will sometimes help avoid the violation message.

[daniellei4510]

Also, you should make your selection roughly in the shape of what you want created. This gives the bot a better idea of what it has to work with.

[Gordon30403982akk6]

You can also just press the space bar once to add a blank space in the prompt. That sometimes helps with the error message, too. 🙂

[Kevinzacher]

The content restrictons can be limiting.  I am working on an image of a fashion model fully clothed shadow boxing towards the camera.  She is smiling and it is a fun image.  Since when is boxing inaproprite anyway.  

Hope you can correct this limitation.

[Ray26434972uq1g]

This worked for my images, no error

[Sarah27914909k9u3]

Thanks for your reply, Michael. I don't think it's really about the words in the prompt. I say that because sometimes Photoshop redacts a couple of the generated images for "violating community guidelines," but allows another one to get through. It almost seems as if, if the generated result contains too much skin-like color, then it rejects it. For example, say I've generated a beautiful woman but she has six fingers and no thumb. Maybe I select one of the fingers and ask the generative fill to create a thumb here. Sometimes some of those queries get rejected. There's absolutely nothing sexual about practically any of my asks, but my subjects are often both men and women with some AI-generated flaw. And that's really my point. My requests are in no way sexual, but the AI chokes on them anyway. It's very frustrating! Especially because I know this would be an amazing feature — if only I could consistently use it.

[daniellei4510]

This is a case where I would use the Remove tool instead of generative fill.

[johnb60867848]

Sometimes it IS about the words in the prompt.  I find some words will result in the VUG warning every time and within a few seconds - before any image coud be generated but small variations of the keyword work fine. example "High heels" fails every time but "High heel shoes" works. Similary "Toes" fails within seconds but "feet" work - at leat to the point of attempting to generate an image, then it is hit-and-miss after that.

[johnb60867848]

I see an update 24.7 (Beta) has been released.  No time to test it as yet. Hopefully this shows some improvement

[Bob5E9A]

Since the Photoshop people don't want to tell us what is going on, we can only guess. I say when PS doesn't know what to do, it gives us the violation message. It does not have a clue what "high heels" are but knows what "shoes" are. Maybe it is all about how well it is trained, not that we are creating something bad. Upon reflections that is the only answer: when it is confused you get the violation response, a cop out from Photoshop who blames us. Photoshop could make this so easy on themselves by doing what Microsoft did with the Bing image generator. Microsoft secured Dal-E as the engine. They should pay a billion to get an engine that is based on Mid-Journey. Just like the Photoshop people are not wise enough to build in an "auto, live" spell check into this board, they may not have the expertise to do what Mid Journey did, ever.

[MontyWales]

Well MidJourney used Google Safe Search for now. They have the same problems. But You can't really compaire the 2 as MJ doesn't do Outfill or Infill yet  

[Bob5E9A]

Gen fill is not that smart. Maybe some day? Who knows? Sara you said, "Maybe I select one of the fingers and ask the generative fill to create a thumb here. " Again, Gen fill is not that smart. Gen Fill has no clue how to create a thumb and know how to scale it and how to position it and blend it in.  So you get the violation. PS could give you a message saying, "Auto fill is in training and should be able to do that soon.", but the easy way out is to make it look like you did something wrong. Can someone draw a cartoon where the Auto fill programmers are laughing their heads off at a confused user when this message is sent. 🙂

[Graham24508943nobd]

Give the devil it's due, the landfill generative fills are 99% spot on as is the extending frame fill. 

Changing clothes is still very poor  though

[Bob5E9A]

montywales, re: Well MidJourney used Google Safe Search for now. They have the same problems. But, you can't really compaire the two, as MJ doesn't do Outfill or Infill, yet.  

And PS may never be able to do gen fil, as we expect. Far too difficult compared to simply just drawing a picture on a blank sheet, like MJ does. It would require MJ and PS to train itself on our picture inorder to throughly understand it, so it can create an accurate piece (a fill) that totally matches the surrounding pixels.  The only hope we have is that it can be done if we, the user, assists by inputting  what to draw, what colors to match, where to attach, scale etc. by click, click, clicking for each question it asks.

Example:

1. What to draw: We describe. It uses Chat GPT to create the prompt.

2. Colors: Click, click

3. Insert where: click, click, click.

4. What point to Connect to object: click

These are ALL variables!

And the fill can not contain background because we probably will have to resize and blend, etc.