I have been using Dall-E3 to help my girlfriend with a project and she needs some text in the image because it is a logo with some small text.
We have encountered numerous times that out of 100 logos, only about 10 of them have the correct spelling as requested word for word. Then when you ask Dall-E3 to keep the image intact but to fix the typos, it changes the images into completely different images with spelling errors.
Has anyone been able to circumvent this issue with other prompting?
Yeah, I get that with every model, the text-to-image and text-to-video models do it, all, except for Adobe’s Firefly. Probably because Adobe fine-tuned their model with all those fonts, they “own”? YMMV.
From my personal experience I feel like the last update kinda made GPT less good at typing (personal feeling). With Dalle-3 first version i was having a good amount of images with correct text: it’s still possible now, but will require more attempts. Finger crossed we gonna have updates about text.
From what i found: short text tend to works better than longer text.
Putting the text in front of the prompt might gather better results.