P: Negative prompting

[EricEm]

In the text to image part, I didn't find a place to cancel prompts like in stable diffusion. Often we have artifacts with double faces or eyes in the wrong places. It depends on the channels I think but it could be efficient and necessary to add this function. Or maybe add a basic lexicon for prompts if we have to use No in front of cancels or others basic operations.

Thank you for your job, im a teacher and i think you are the best to create an democratic useful app for commons users

Eric

[n7best]

feel either in the prompt use --no or a UI option for that. Saw people bring that up in discord also.

[Tracy.H]

Thank you for your feedback. Right now, the model doesn't understand negative prompts, so adding it to a prompt will add it to your image. You can use the parameter "[avoid=xxx]" in your prompt where xxx is the thing you want the image to veer away from. (Merging all similar threads here so we can track the popularity of this feature request)

[SanDiegoDude]

I'm finding a lot of situations where the ability to say "no, I don't want this" would be super beneficial.

For example, the prompt I am working on right now, I keep getting people with hats on their heads and/or masks on their faces. I'd really like to say "no hats, no masks" to keep them from being produced.

Hopefully this is something already in the works!

[Ladev]

I experienced difficulty while defining what I do not want to be included in the image generated. For excample I want a still life background with shapes but no product. There is a keyword in stock photography to define images without people - "no one". So there should be a mechanism to better define what you want out of the image. So if I search group of animals and it shows a great result but there is a green frog I don't want in, it should be great to be able to dial it in better. At this time if I include something like no whairs it will actually add the item.
Love the product will use it for sure.

[VazhlaV]

1:1 ratio is more accurate at first shot compared to other ratios, only visible if change Ratio while keeping same text.

Firefly struggles with negative values as>

something-less

no-something

without   

[Tracy.H]

Thank you for your feedback. I am moving this to the Negative prompt thread but I have also noted your comment on 1:1 aspect ratio being more accurate than others. Thanks!

[wanderloster]

Ive seen this is other programs by just putting a hyphen in front of -pink for no pink. This is also the way to do it in google search and other such programs so it would feel intuitive for it to be like this.

[defaultorvincps6r0o]

I prompted an elaborate scene:
"Blonde sleek long-haired celebrity actress dressed in a very elegant dress walks up stairs for an award ceremony in a beautifully decorated theater filled with a lot of guests. Dress is neither white nor ivory. Oblique side view from a distance with audience partly in view. Both shoes match in style and color."
I tried several prompts but I have a hard time to obtain a dress that is not white even when I explicitly state "neither white nor ivory". When I use "colorful dress" it becomes better but all the dresses have a very strong bias of being created in white or contain white.

[Ladev]

Yes, that is a great strategy, and should be implemented.

[Personal information removed as per forum guidelines]

[Fred5E8D]

Needs to honor negative prompts such as "no text".

[Annajepa]

realy need the negative prompting.

[GraphikJules]

Not sure why this keeps duplicating my response... can we have delete option and not just edit please?

[GraphikJules]

Are multiple posts getting merged together? As I'm definitely not entering this comment multiple times. Also, can we have the option to delete in addition to edit our comments please?

On Mar 29, 2023 in the Discord community, Tim from Adobe said to use [avoid=PROMPT]

[GraphikJules]

On Mar 29, 2023 in the Discord community, Tim from Adobe said to use [avoid=PROMPT]

[Dean_Utian]

Not sure if this is a bug or just an enhancement but...

I created images of
<animal> wearing suit jacket no tie
Result - images with a tie in every one.

<animal> wearing suit jacket without a tie
Result - images with a tie.

There should be a way to tell Firefly to exclude something but with my tests above (and variations), I have not been successful.

[Wiher154]

I don't know if i'm doing it wrong but it seems at the moment Firefly doesn't interpret word "no" correctly. I suggest some function for excluding some elements from picture. At least for now it's not clear for me enough how to do so.

Thank you for your work and have a good day!

[Shane289354640if9]

Would love the abilty to upload an image and create from image or also have negative prompts so i can tell AI what i dont want to see.

[Tracy.H]

Thank you for the suggestions. We are already considering both of these feature requests.

[Tracy.H]

@Wiher154 Merging your thread here so we can track the demand for certain features.

[Guest]

"Egg-free breakfast for those of us with egg allergies."
I let the AI draw with this prompt and it produced this image, thank goodness Firefly is a drawing software and not a cooking software.

[An3]

This thread's response explains:

https://community.adobe.com/t5/adobe-firefly-beta-ideas/negative-prompting/idi-p/13676093

[LAB 2]

would be amazing to give some negative prompts to make the vision more clearer

Merging your thread here so we can track the demand for certain features.

[Guest]

Thanks An3. I now understand for what reason this phenomenon occurs.

[Anitaphotography.nl]

This would limit the amount of unwanted content.
You can type the things you don't want to see in there.

For example: people (no people in the image); straight lines (all lines will be curved); red (all red tones will be avoided) and so on.

Merging your thread here so we can track the demand for certain features.

[Eefun28832123w0kj]

Can a "negative prompt" be added to the “text to image” tool to avoid unwanted objects in the generated photos?

Merging your thread here so we can track the demand for certain features.