Skip to content

Text in images

What it does

The assistant can make a new picture that carries a short text, such as a poster with a title, and it can edit a picture you attached, for example by writing a name into an empty field of a card or by correcting a date. Everything runs on Myra's own image service in the EU.

Image generation has to be switched on for your gateway, and the model must be able to use tools. Ask your administrator if the assistant only describes pictures in words.

How to ask

Ask explicitly for an image or picture. If you only say "poster", the assistant may build an HTML file instead of a picture (with qwen3.8-27b, 4 of 10 runs did with such a wording). Put the text exactly as it should appear. Below are three prompts you can copy and adapt.

A new picture with a short text

Create an image of a poster for our summer party with the words SUMMER PARTY in big letters.

Tested on: qwen3.8-27b, gemma-4-31b-it, mistral-small-4

Asks for an image, names the subject and gives the exact words.

Write into an empty field of a card

Write the name Emma in the empty field at the bottom of the card.

Tested on: qwen3.8-27b, gemma-4-31b-it

Not recommended for: mistral-small-4

Attach the card to the same message as this prompt. The prompt names the text and where it goes.

Correct a date in an attached picture

The date in this picture is wrong. Change it to 10.10.2026 and leave everything else as it is.

Tested on: qwen3.8-27b, gemma-4-31b-it

Not recommended for: mistral-small-4

Attach the picture to the same message as this prompt. It names the change and says that everything else must stay.

Each prompt was run on the three Myra-hosted models, ten runs each. A run counts as passed when the picture was really made or edited, and an automatic text reader found the requested text in it; for an edit, the other lines of the card or picture had to still be there. The automatic reader can itself make mistakes. What the picture looks like, and whether the letters are shaped the way you wanted, is not part of the check; that is up to you. Under each prompt you see the models it was tested on; the table shows how many of the ten runs passed:

Prompt Test case qwen3.8-27b gemma-4-31b-it mistral-small-4
A new picture with a short text a summer party poster with the words SUMMER PARTY 10/10 10/10 9/10
Write into an empty field of a card the name Emma written into the empty field of an attached card 10/10 9/10 1/10
Correct a date in an attached picture a wrong date in an attached schedule changed to 10.10.2026, other lines kept 9/10 10/10 0/10

Tips

  • Say "image" or "picture" in the prompt, and write the text in the exact spelling you want.
  • For an edit, name the place ("the empty field at the bottom") and say what must stay the same.
  • Look at the result and check the text yourself; the automatic check can make mistakes.
  • To put labels or text on top of a picture you attached (for example a chart), attach exactly that one picture to the same message and name the labels or numbers to add and say that nothing else may change; the assistant writes them onto the real picture in the code sandbox instead of drawing a new one. For a later change to the result, attach the picture again.

Limits

  • Without the word "image" or "picture" you may get an HTML file instead of a picture.
  • One source image per edit. With two or more pictures in your message, attach the one you want on its own.
  • An edit keeps the aspect ratio of the source image and produces at most about one megapixel.
  • Only PNG, JPEG and WebP pictures of at most 12 MB can be edited; a GIF cannot. Writing onto a picture you attached (the code sandbox path above) takes pictures of at most 8 MiB.

When something goes wrong

  • You get an HTML file instead of a picture. Ask again and say "Create an image of …".
  • It says it cannot edit because there are several images. Attach the one image you want to edit in a new message, on its own.