I’ve noticed that sometimes the text generator isn’t accurate. For example, it generates duplicate text or forgets letters or replaces some of them.
Example: “hAllo” instead of “hEllo”. Or “Hello” twice.
Even if you tell it to write only once, it will still sometimes duplicate the text or make typing errors.
No matter how hard I try to correct my prompt, I can’t get it to generate simple words and phrases without error.
Do you have any ideas of prompts ensuring accurate generation?
It’s best with a single word or maybe a very few, if you are trying to pass it an entire sentence you will get lesser performance, peoples names, company logo’s product names, that kind of thing it’s great at, but not inspirational statements or longer stuff.
I’m making a guide about using Bard, ChatGPT, Midjourney, and DALL-E, but it’ll take a while to finish. Here’s a helpful tip for now:
- Keep your prompts simple and clear.
- Make sure you know what you want to achieve when you create prompts.
- Also for text response role play tends to give better responses.
- for Images the technical details with clear goal to achieve help to get so much good images . Also One thing is to Learn from others prompts.
Clear Example Prompt:
Can you provide a detailed guide on crafting effective prompts for Bard, ChatGPT, Midjourney, and DALL-E, with an emphasis on clarity and simplicity?
Unclear Example Prompt:
"Tell me about those things, you know, Bard, ChatGPT, Midjourney, DALL-E, and the tip, like, what's up with that stuff and how it all works?"
Great, thanks for your feedback, hope it will improve soon
You could just add the words in yourself.
Yes, most computers pre-installed photo-editing systems have the ability to add text.
This is not really a workable answer. Sure, you can try to insert text with Photoshop or some comparable software, but it’s hard for anyone with beginning-to-intermediate Photoshop skills to make the text blend in as if it were part of the image. There are things like warping and shadows and partial obscuration of the text that would be much easier to hand off to an AI.
Sure. If the AI could do it. Of course “telling” something to do something will always be easier than doing it yourself.
Words have always been difficult and maybe in the future it will continue to improve but for now the most reliable solution is to add it yourself.
Also, adding text with appropriate shadows/warping/blending is definitely beginner-level Photoshop
Have you seen how ideogram.ai is doing with insertion of text in generative AI? It’s pretty stunning stuff. Better than Dalle-3 and Midjourney for sure. The rest of their functionality is not quite as good, but they demonstrate that it is possible. If you don’t believe me, check out their system (free signup) and look at the community pages. You have to wade through a bunch of annoying “name artwork,” but you’ll see some pretty impressive examples of generative AI doing a decent job with type.
This is super cool. Thank you for showing me.
I completely believe that soon enough these type of models will be fully capable of writing out text. For now though they aren’t and if it’s failing then the solution (in my opinion) is DIY.
I did try with some words and it was pretty good. Soon enough. I’m sure of it.
OK… Wow. This is impressive, and relatable. But, like you said, this after combing through a lot. So I would consider these “exceptions to the rule”.
So going back to the original question: For Dall-E 3 if the model is not writing the text correctly regardless of the prompt, the most reliable solution is to add it yourself. Or find someone who can do it.
Short text is usually better - a single word or two… make take a few gens to get it…
When DALLE3 comes to labs, we should be able to edit there, but at that point it might be better to just do typography with photoshop?
And of course the relevancy of the words to the image. I think the same rule carries through all AI models (for now):
Handle logical sequences yourself.
Oh my god I just realized I have Dall-E 3 now
I just tried out DALL-E 3 for the first time,
(A white sheep laying in a pasture basking in the light from the sun)
It is WAY better than DALL-E 2, I can’t wait for the API.
I started a thread for everyone to show-off their prompts / art, so we don’t derail this one too much away from “text” generation with image model.
No doubt. If it works, it works.
If it doesn’t then Photoshop is a more reliable option. Pretty straightforward.
Ah crap. Missed the I.
One of the things I like about Adobe Firefly is their stylized text tool…
You choose letters, and it applies the style to the letters…
You could likely do it with DALLE edit image endpoint if you had cutouts of the letters. I’ve thought about it, but never took the time to do it yet. I do have shield-shaped images coming out of DALLE2, tho, which works in a similar way.
Absolutely amazing! Is there a way we can view the full prompt?
DALL·E 3 is generating crazy responses, i found this article showing DALL·E 3 results
Some strategies I’ve found effective for enhancing text clarity and reducing grammatical errors include:
- Adjusting where the text is positioned: Sometimes moving it to a different spot within the image can make a big difference.
- Emphasizing specific text requests by enclosing them in quotation marks and prefacing them with descriptors, like “clearly”, “specifically”, or “highlighted”, multiple times within the same prompt.
- Clarifying the desired writing style: Offering detailed descriptions of the preferred style (e.g., modern, minimalist) seems to sharpen the model’s focus on the text.