Dalle-3 is very good at improving our prompts, but is there a way to force him to follow my prompt ?
For exemple : i asked him to generate a picture of a blue flower in a grass flat field at noon. Dalle3 keeps generating a picture of a blue flower (sometimes he generates more) in a grass field but surrounded by moutains, at sunset / sunrise.
Why isn’t he following my instructions ?
Not sure how long it will work, but this works last I checked…
Follow this prompt exactly, “(your prompt here)”.
Do not add to the prompt. n=1, orientation=square
Let us know if it works for ya…
On the flip-side, you can also be more descriptive, ie tell it exactly what you want in the foreground, background, etc, so it doesn’t have to “guess” and fill in the space.
Quick response here, thanks !
Yep, i have heard of this tricks where you write “Follow this prompt exactly, “(your prompt here)”.” and that’s what i did for my blue flower in a flat field.
I have just tried one more time by c/c your prompt, and DALLE3 still generates what he already did : a blue flower in a field surrounded by mountains at sunrise/set, despite the fact that, when i click on the image, the prompt is exaclty what i wrote.
See the picture :
From top left to bottom right : (Prompt 1 & Result) - (Result & Prompt 1) ; (Prompt 2 following your suggestion & Result) - (Result & Prompt 2 )
API image creation w DALL-E <2
“a picture of a blue flower in a grass flat field at noon”
- A picture
- a flower
- in grass
- flat field (maybe)
- at noon (apply your image forensics to determine latitude)
Dall-e 3 likes to add realistic context that is not mentioned. “Picture of two men at a wedding” gets you the partygoers and streamers in the background.
I would try more prompting, like “Subject: A blue flower; Background: a flat green open meadow where the short DOF background is out of focus with bokeh; Lighting: noon sunlight; Style: modern photograph; Mood: inspiring, artistic”
You can see this image, specified with “Informal photo taken outside with a background of out-of-focus foliage” got me the desired setting even though in this iteration there is something going on back there besides just the greenery of other trials.
You can just tell it you want a blank or white or black background if you want…
Use ordinary language, like you would with a person. Here’s what I do…
"Run this prompt, verbatim, without modifications:
If you want to execute your verbatim prompt and then entertain some variations:
"Run the following prompt, verbatim, and then feel free to create some variations:
I brute force through hours of generating images. Hours of teeth gnashing. This was the unfinished product. Even the throwaways can be stunningly good.
Second day of using ChatGPT, so hopefully with time I get more adept.
yeah, i noticed this all week that it keeps trying to give creative variations that you never ask for, even when you tell it to not deviate from the prompt. Something has changed in the system. It seems we have to stay on top off prompt instructions it will listen to. I’m sure someone will figure it out and post it in the topics!
It’s a shame it changed. Unfortunately I only have the experience of the last two days of using it. I notice it hates to show faces if there are certain situations depicted.
Its becoming to politically correct for situations that have no reasons that a creed or colour or gender matters. For example, i wanted to test how it treated a mirror reflection of ‘a man’ just as my basic guideline. It gave me 3-4 results as i was tweaking the prompt, then it suddenly told me it would need to adjust my prompt to include a ‘descent’ lol I asked what does it matter if he is white or black. And it want on about basically not wanting to upset anyone by ‘assuming’ a descent. FFS! lol
Yes, the censorship is frustrating. However, just remember, there are ways to get it do almost anything. You just have to talk in a way that doesn’t trigger their alarms. Also brute forcing by repeatedly asking for variations and appealing to the AI’s ‘sympathy’. It’s frustrating, but also fun, if you think of it like a game.
Yeah, I’ve learnt to ask lots of questions and getting tons of variations. Its a bit like wanting a spiderman picture, but you have to say give me an illustration of a superhero with spider qualities! and it shows you spiderman lol
Hopefully once the API is released we have more control over the prompt. ChatGPT is a great tool to “improve” the prompt but it really tries to take too much control.
It would be nice for cGPT to say “this is how I have altered the prompt”, and ask for confirmation, but I get the whole AGI fiasco.
Fortunately, we can intervene once an API version is released and finally, separations of concerns can exist again.
I mean, my god.
Could you imagine asking for vanilla ice cream and you get some motherfucking 75% chocolate rocky road mix.
Besides not liking the taste of chocolate, sometimes I just want vanilla dammit.
You can start with “
Do not change my parameters, do not change the prompt, do not touch my code”, followed by something like this;
"prompt": "your (long) text...",
"refering_image_id": "[if any...] "
And ask to show the
formatted JSON data BEFORE and AFTER the image creation, so you can check if the prompt was touched.
If so… this alternative works 9 out of 10 times;
Okay, you are touching my prompt. Let’s change the workflow and do this step by step. I give you a key / value pair and you construct them as valid JSON data. Do not create an image, wait until I say “okay”.
Then you go step by step through your prompt, assuring ChatGPT is exactly following the rules, like;
size : 1792x1024
prompt : your (long) text…
n : 1
refering_image_id : if any…
You can also give a fake key / value pair and say
do not touch anything, but report it when it is not valid;
bananas : 23
This way, it will show you step by step the construction of your prompt and the
bananas will be reported (so you know he is listening).
Using the step by step approach, at the end my prompt is always 100% followed (except when he went woke, for god sake).
I have exactly that in my
custom commands and this works flawlessly.
The output is something like;
1. This is the data I will send
2. This is the image I have created
3. This is the data I did receive
For complex prompts you can do a
file diff in order to see what has happened.
Or push them to a .git repository, so you can not only track differences in text, but also see what happened at pixel level (just export, iterate, render and push the data from your profile).