Participant

Your art generator is the worst I have ever used. It is very good at generating art exactly opposit

Forum|Forum|1 year ago
April 17, 2024
6 replies
596 views

Your art generator is the worst I have ever used. It is very good at generating art exactly opposite to the text, and seems deliberately defying and going against the text input.

For example, I input "A dog looking up to a high statue with a human on top", and your app manages to put a dog on the top of the pedestal with a human looking up, EVERY TIME!

Desktop

Donald2415339450as

Participant

@droopydog500 , you just furthered the point we are trying to make. The fact that we have to read any guide that teaches us how to use this specific AI becuase it does not behave inuitively like every other AI model strives to, is proof of how far behind it really is. I'm not opposed to learning how to use a tool, but if you created a created a can opener that only worked if you used your feet and only after 3pm, it wouldn't get much traction. Negative prompting is a pretty important feature.

droopydog500

Community Manager

Hi @Donald2415339450as,

The model does not understand the concept of a word being negated by the word in front of it. So using "not [X]", "no [X]", "exclude [X]", "minimise [X]", "with no [X]", and "without [X]" does not work. As you observed, it ignores the negation word and adds to the image or video the thing you are trying to avoid. You basically need to find "positive" words and phrases telling the model what you want to see rather than telling it what you do not want to see. Such as "flat land" rather than "no mountains" or "desolate" rather than "no people". If I am not sure how to describe something I want to exclude in a positive way, I will ask Google Gemini or ChatGPT for ideas. Just be aware that when they give prompts, they are usually too wordy and conversational, so you have to edit them, but they are good for ideas.

Your prompt is too long, too conversational, like a story. The model will not understand it well. The most important object should be as far left as possible and everything related to that object should be grouped together.

I recommend this document: How to create stunning visuals with Adobe Firefly Image Model 4: Tips, Tricks & Visual Examples and Writing effective prompts to further help you in crafting good prompts.

My best,

droopy

Adobe Community Expert (not an Adobe employee)

Donald2415339450as

Participant

i don't know... not sure how much more specific i can be:

swarm of 6 black hornets like the style image, with no yellow (just dark), flying around in the image frame at different angles with central wasps being slightly large to appear closer. Straight on camera angle. The wasps should be flying at different angles and should not all look exactly like the style image, but only use that as a reference. please make it look more natural, like a swarm. The wasps should each only be approximately 1/16th to 1/20th of the image size. please make them look like they are flying, some towards, some away, some up, some down and all at various 3D angles. make these photorealisic wasps. pretend they are swarming around an invisible nest at the center of the image but DO NOT add a nest. Please try harder to make this look like a very realistic cluster of flying wasps. No shadows and none of them should look like they are sitting or standing. and please add realistic wings to they each have 4 wings. obviously, not all 4 wings will be visible for each position.

And i got this:

@mj

Community Expert

Firefly is fussy.

This may help too: https://helpx.adobe.com/in/firefly/using/tips-and-tricks.html?linkId=100000250655968&linkId=100000253415044

mj

iMSD

C

Chiu36810288jtueAuthor

Participant

Thank you. I tried your prompt. There is still one with a dog on the pedestal. But the rest at least have humans on it. I think I got your point.

Dave Gustafson

Inspiring

Your prompt is confusing, the way it is structured, sounds like a dog, looking at a statute, with a person on top of that. So you are stating 3 figures instead of 2, this is what is causing the issue.

I fixed your prompt:

A dog looking at a "human statue on a pedestal"

My first and second generations produced these images right off the bat.

My prompt is concise and reads evenly, with no confusion as to what is going on.