FireFly really poor at generating people images, and following scene instructions.

[Art-ist]

The product doesn't not illustrate the verbal description. Believe that you should allow a user-uploaded picture or one of the previously generated pictures to be uploaded (or check) as the starting point for further refinement. There needs to be a way of ranking the 4 choices provides from worst to best, and have the best used for the next iteration of the image. People generation is by far the weakest link of this program. The people are in no way realistic. Some have three feet, some have donkey legs, nobody has a face. Useless for creating a scene with people in it. I am extremely unimpressed by this program in its current state 9/2023. When a newer release that learns how to read and follow instructions is available, I would like to evaluate it again. For now, it's crap and a time-waster. 

[Smitty32336246pzon]

I agree that the program is not good at rendering realistic features, you can correct the image with the edit option but it uses one credit fore each edit and an image can take many edits to make it look good.  And at least in the Beta version it could crash before you are finished.

[Anubhav M]

Hello @Art-ist, @Smitty32336246pzon,

We are sorry for your difficulties. Please let us know what is happening when you try generating your image. I recommend sharing a screenshot for more details and investigation. You can also share feedback directly on images you generated in the app by clicking on the thumbs up or down > Feedback > Select all applicable > Submit Feedback.

For help with using a reference image when using Text to Image, kindly check out this community post: (https://community.adobe.com/t5/adobe-firefly-ideas/p-upload-a-reference-image-s-for-text-to-image/id....

Thanks,

Anubhav

[Stan33165354ch79]

Absolutely, first generation of Stable Diffusion was better than firefly with faces, and don't get me started about hands and fingers. To me, all the images look pretty much like poor cut and paste. Not good for the king of all things graphic related when free open source software like Stable Diffusion xl can outperform your paid offering. 

[tand78023760]

Adobe firefly AI generation is really a waste of time.. just wanted it to generate some background/ logo design but, 99% of the time it generates some really weird shapes and very complex elements/ texture, even after giving reference. Hope adobe will do something to improve this AI soon.

[InfiniteWhale]

I just want a sideways view of a street.  No combination of words can achieve whatcan be found instantly with a google search.  Adobe must exclusively employ incompetent people because everything they do is useless

[Barbara34794578m2tg]

Does anyone review our "thumbs-down" submissions? Is there context when you review feedback? I don't feel heard. I have sent detailed feedback, but tonight, my results are worse than ever.

[Barbara34794578m2tg]

I feel you. And it differs every single session. Sometimes signing out and giving it time to cool down helps. I could not get a man who looked angry for three straight sessions. It was... uncanny. His lips just would not 😠 Then i got a 😐 in my results and ran with it.

I also get so. Many. Improperly. Flagged . Responses. It seems completely random. Sometimes I can write "woman in custodian uniform puts arms out and hands up defensively" and something is blocked. It never says what. Woman? Custodian? Hands up? Then someone else can do it no problem. Or I can do it later.

The whole thing is a crap shoot.

[CMass]

If you're having issues with Firefly Image 2, please add your vote to this post: Firefly Image 2: Human are not realistic

[RobertJohnson353]

I have literally been trying to get my full prompt to work. Then I found something close that I was going to try and work with but when I tried typing in "black lips" or "black lipstick" it refuses to give my person the lips I'm looking for. And its not just been that. I tried to do a completely different prompt with a strapless bathing suit and "strapless" was not allowed. And I was just introduced to Bings text to image and omg it got my prompts almost perfect the first time. Sadly you only get a certain amount of credits per day. But jeez, i feel like in the last two months since I lasted used firefly its just gotten so much worse. 

[CMass]

Can you share your exact prompt and the results you got? Also, please share feedback within the app when a result doesn't appear as you intended. 

[Sergi22232848ieg1]

Faces look like something from a horror movie. And the scenery is bad too. 

[GeekFilter]

I had a similar issue with people, except for the fact that it doesn't listen to you when you say "no people". or if you say left-hand, no matter what you get the right hand so I thought OK maybe it's just transposed so I said right hand then I got no hands. This is like asking an alient to draw something!

[GeekFilter]

Also, I'm writing this when I have a massive migraine so hopefully it makes some sort of sense (and although it could be, Firefly is not the cause of this headache!) and apparently in 2024 Adobe doesn't let you delete or edit your posts. I remember when this company was ahead of the curve on so many things, now it can't even keep up with Facebook!

[marenw]

Even worse when you ask for people running. They're all melting Picasso-esquely and have spaghetti legs. The latter might be because I asked for college students so naturally they're all in skinny jeans.

[marenw]

I forgot to mention, it seemed to do OK with people when I tried it in beta around December 2023.

[Mike Sirén]

Adobes AI is a joke. You cannot even generate realistic car, or human, or anything. How they even can say its AI because child can draw better.

[Mike Sirén]

"Formula one car" promt Adobe vs Midourney. 

[InfiniteWhale]

Then they have the absolute brass nexk NERVE to make it a credit system, so not only are we having to rent software that gets increasingly worse, we have to pay extra for broken features.  I almost can't believe Adobe hasn't failed as a business, they are so bad at EVERYTHING! 

[Jan239468033oau]

It seems to be the issue with just about any image generator out there including Midjourney. AI image generation is still faaaaaaaaaar away from being used in real client works. AI image generators are for now a wish or a nice-to-haves but the quality of images generated using those technologies remains incredibly poor and primitive. Personally, I can't believe people hop on paid plans for image generation.