ChatGPT managed to add the Romanian words to the DALL-E images, but he failed to integrate the diacritics/accent marks well.
For example, instead of ã should be ă. For example, instead of “Picaturã” should be “Picatură”
Instead of “Analizâ” should be “Analiză”
INstead of “Sânâtate” should be “sănătate”
So you have to differentiate between the diacritics, because they are not placed correctly.
DALL-E cannot create text perfectly.
There was an official info on following FAQ page, but it was removed.
https://help.openai.com/en/articles/6781228-how-can-i-generate-text-in-my-image
It was saying:
How can I generate text in my image?
You might be tempted to instruct DALL·E to generate text in your image, by giving it instructions like “a blue sky with white clouds and the word hello in skywriting”.
However, this is not a reliable or effective way to create text.
DALL·E is not currently designed to produce text, but to generate realistic and artistic images based on your keywords or phrases.
While you can request text in your image descriptions, the results might be distorted, unclear, or not as expected, as it does not have a specific understanding of writing, labels or any other common text.
4o Image Generation (image_gen) sometimes shows some letters like “ă”, “â”, “ș”, or “ț” the wrong way in images. This happens even if the prompt uses the correct letters. It’s not your fault.
Romanian uses the Latin alphabet, but it has special marks called “diacritics”.
The model sometimes mixes them up, like using “ã” instead of “ă” or using the wrong kind of comma under “ș” and “ț”. This makes the words look strange or incorrect in Romanian.
OpenAI has said in their official system card that GPT-4o is great at making images with text:
It can follow detailed instructions, including reliably incorporating text into images.
But, it can still make mistakes, especially with less common languages or scripts with special letters. Romanian falls into that group.
There’s also a research paper on arXiv that tested GPT-4o and found similar issues.
It says the model has trouble with underrepresented languages and some non-English characters.
But interestingly, some languages is written correctly sometimes, for example Turkish:
Gözünün yağını yediğim, n’örüyon? Bu şehir çılgın!
yes, but AIs need to advance. 2 years ago, I was doing the same thing with images. I mean, if I know any language well, I should also apply correct text to the image. I don’t know why it would be difficult. First, the AI generates the image, then applies text somewhere above it, with another layer. I don’t know why that’s difficult. And I don’t know why it writes texts correctly if I ask it to, but in images it doesn’t write them well.
I think they are working on it, but it will take time.
