4

I am using ClipDrop/Stable Diffusion to generate some images with this specific prompt.

big number 4 at the center of the image with forest in the background

and this is the result.

results of big number 4 at the center of the image with forest in the background

As you can see, my big number 4 at the center of the image is entirely ignored in all of the results no matter how many times I regenerate the prompt.

I tried changing the number from 4 to different numbers like 1. Still, no number comes up in the image.

How to emphasize it so that it is absolutely going to be in the image?

Edited according to SPArcheon's suggestion: It turns out that some styles give me the correct results. Right now, I am out of daily quotas to try more. I tried no style and anime style. With no style, it gave me the number. With anime style, it gave me like the ones in the question.

3
  • Related about prompt weight: What do (parentheses) and [square brackets] do in image prompts?
    – Andrew T.
    Commented Jul 19, 2023 at 8:28
  • Based on a quick look at the page, you may also want to specify the style you are using. I suspect that different styles may be mapped to different models.
    – SPArcheon
    Commented Jul 19, 2023 at 9:53
  • @SPArcheon Yes. I checked, and it turns out that some styles give me the correct results. Right now, I am out of daily quotas to try more. I tried no style and anime style. With no style, it gave me the number. With anime style, it gave me like the ones in the question.
    – holydragon
    Commented Jul 19, 2023 at 10:18

1 Answer 1

6

This is a blatant example of how difficult it is to have a good answer to similar issues.

As you saw firsthand the site allows you to select multiple styles but not every style works the same way. Since you are using an actual online generation service you can't even know how the thing is configured under the hood. Are the styles just prefixes that get added to the prompt you specify? Or is each style mapped to a different model file? Are they using publicly available well know model whose behavior is "documented" or do they use custom closed models?

Since all these bits of information aren't available you can only use an empirical try-and-see approach: try a prompt, see what you get, change a little, retry (very convenient when you are using a paid service. Imagine a pay-per-use license where it is not clear how to use said service and you have to try- and pay - again, again and again just to get how things work...).
But you already did without much luck.

Sadly this is no longer just a matter of "prompt syntax / format" but more of a "fight against the model biases" one. The model seems to be biased toward a landscape artwork and trying to insert text directly works poorly. In this case you have to test your specific use case and try to bruteforce the model into doing what you want.

I tested the site you linked as an anonymous user.

Rewriting your idea into a physical entity instead of an abstract one helps.

(((a statue shaped like the number 4))) in the middle of an enchanted forest

enter image description here

1
  • 1
    It works sometimes though, like 1 image out of the generated 4 images will have a proper number statue in it like that. Better than my previous prompt anyway. Thank you!
    – holydragon
    Commented Jul 19, 2023 at 14:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.