How to stop DALL-E generating text into images?

Is there any effective way to prevent DALL-E from adding fonts/text/characters into the images?
I know negative prompts aren’t working well, so what is a good way to deal with this?

1 Like

Have you tried using the terminology “image only without typography”?

Yes, but sadly it has zero effect on the outcome. For example, here’s the image it generates for the following prompt:

Blockquote Create a vector artwork image in a 1.91:1 aspect ratio and in 1200 x 628 pixels, for a blog post paragraph for a blog post discussing the Introduction to auto loans and covering the topic of auto loans. The image should be image only without dypography, and it should contain at least one person or object related to the paragraph title. The image should contain only vector illustrations. Use primarily a blue color palette. Use a clean, solid white background.


A quick fix for DALL-E text issues is to refine your prompts.

For instance, ask ChatGPT to describe a specific image focusing on visuals.

Then, edit this description to exclude text references and use it as a DALL-E prompt.

While this workaround can be effective, it’s a temporary solution.

Think creatively about instructing ChatGPT to generate unique images for each topic.

Example Prompt:

  1. Describe: “Describe image [number] or use generation ID [ID].”
  2. Edit: Remove references to personalities, brands, and text-generating elements.
  3. New Prompt: Form a prompt for DALL-E based on this edited description.
  4. Generate: Submit the prompt to DALL-E.
1 Like

Try adding these 3 rules to your prompt:

  1. Focus on specific, visually representable elements.
  2. Describe actions and scenarios rather than abstract concepts.
  3. Avoid ambiguous language that could be interpreted as including text.

It helps! like tendencies of text is probably 10%

Hi, wondering if you managed to find sustainable solution for this?
For me also “do not include any text or letters” didn’t effect, my 2 cents is focusing only on the object rather then usage, assuming that by specifying “for blog post” (or “for a travel journal” in my case) makes model provide a “complete” solution as (again, assumption) these types of media often use text as well for emphasizing the idea through the artwork


Yes, similarly to what you suggested, the best solution I found was to ask GPT4 to write a new visual prompt for DALL-E while avoiding any mention of books, signs, titles, etc, and then use this prompt for DALL-E.

As long as the prompt doesn’t mention anything “blog” related or anything that assumes the use of text in the image, then it usually works fine.


The word “character” seemed to trigger it for me a lot…

Yeah, when you try to use a negative prompt (do not do this) it usually doesn’t work very well. Better to focus on what word is causing it… sounds like blog or journal could be it for you.

It’s very disappointing that negative prompts have no use yet on DALL-E 3. Stable Difussion uses negative prompts very accurately… Hopefully, DALL-E 4 fixes this.

1 Like

As said above, I can also confim that “character” word seems much more effective than others. I appended existing prompt with something like “Do not use any character on image” and the results are much better now. Though still running into some images with letter-ish lines, but at least not full text