Spelling mistakes in Dalle-3 generated images

I am using Dalle-3 using APIs to generate social media posts inside our application. I have often noticed the spellings are wrong in the generated image even when told explicitly in the prompt to use correct spellings. Can anyone suggest a better approach?

1 Like

Because of the way DALLE works, text is difficult. DALLE3 is an improvement over DALLE2, and I’m sure DALLE4 will be even better. Image text generation isn’t stable enough for production use, imho.

1 Like

Photoshop. :person_shrugging:

extra words for discourse

Yeah, this is a known area of improvement for us, we have some stuff in the works to address is so stay tuned! My honest suggestion today is to not use DALL-E to generate the text if it’s more than a few words and add the text after the fact using a tool like Canva.


This issue has been going on for so long that I started to think it was impossible to get the text right in images. But today, while browsing through the gallery, I noticed that many users’ artworks included correctly spelled text. So, I decided to give it a try. I casually created a statue and asked for the base to be engraved with ‘What exists is reasonable,’ and to my surprise, it was spelled correctly. Maybe you could give it more tries?

This issue is really annoying, even after giving correct spelling , it’s still giving wrong spelling . So most of the time , have to give up on using texts .

The question I have is “How do you stop DALLE 3 from producing text in images?”

The answer to discouraging text is not to just send text to the image creator like it was a story.

Instead, describe the contents of an image as if you were piecing them together yourself. Just typing something up in a style where I know it’s not going to put “frog” as text into the image:

“Create a storybook image, with the style of a line drawing that is filled with watercolor. The imagery should portray a frog that riding on the head of a crocodile, The fat frog is green, but also has colorful accents and large eyes. The crocodile is seen partially submerged in the water, with just the top of its head emerging, with the frog seated on the crocodile head. The background beyond the water will be a pastel jungle scene that comes down to the water’s edge. The image is wide format, filling the frame edge-to-edge.”


Thus what I have been using after trying different prompts and getting wrongs texts on images

The only reason I signed up and paid for Plus was to get access to DALL·E 3. Since it can’t create images with the right text, even after prompting it with the correct spelling it still doesn’t spell properly. So what’s the point in paying for it?

That’s like buying a model-T and asking why it’s not electric and doesn’t have cruise control yet! Small smile.

Seriously, though, if you haven’t seen/noticed the improvements in just the last 12 months, hang tight. It will get better.

1 Like

Works better if you tell the AI between you and DALL-E 3 to include less description, place the text in quotes, or just let your prompt go unaltered.

1 Like

Yeah! I’m inclined to agree.

@grandell1234 suggested photoshop. I’ve been using photoshop ai to do some cool, cool stuff… but text is still gobbledygook y’all. Even if you only, and explicitly, tell it to do text, or some text effect.

In fact, overall, Dalle does some wayy cooler fully rendered images. The stuff that @_j just posted is way more coherent than anything I’ve seen PhotoshopAI come up with, text-wise and at a single go.

If I want text, I have to add it later. In fact, that’s one thing I like about Photoshop AI better than DALLE, its how you make selections and have the AI fill the space… In a more-perfect universe I’d be able to make a selection in Photoshop and have DALLE do the rest, which is kind-of what it sounds like what’s happening with Sora and Premier Pro. :upside_down_face:

Actually, one thing I HAVE done is generate an initial image in DALLE, then pull it into Photoshop. Use Photoshop AI to make specific selections and changes. Then add any text or anything else using native Photoshop things.

I bet all this stuff with text has engineers at Adobe and Open pulling their hair out.

1 Like

Yeah, I think Photoshop Beta has Adobe’s latest image model, but it still lags on DALLE and even SD and MJ? It’ll improve, though, I’m sure.

The gen-AI Photoshop CROP tool has been great for expanding ebook covers to full-wrap. Gotta take it slow and it messes up a lot, but saves a lot of time.

Adobe has Firefly too. I believe that’s the tool that does Vector Text… but that’s not full images.

@_j’s example is one of the best I’ve seen. A few months back the DALLE team on Discord mentioned they were tinkering with text a bit.

1 Like