Text in a picture
Lesson 4 of 13 in the course “AI Images and Slides”. Free, with animations and a quiz.
Why letters used to be mush
A model draws letters the same way as everything else: as shapes resembling what it has seen. That’s why older models turned “Liya’s Candles” into gibberish. By 2026 it got much better. GPT Image 2 in ChatGPT came out in April and, according to OpenAI, is noticeably more accurate with dense text, interfaces and diagrams in many languages. Nano Banana 2 in Gemini writes legibly and can translate lettering right inside the image. But mistakes still happen, especially in long and small text.
Put the text in quotes
The model can’t guess what to write, so give it the exact text in quotes and say where it goes: “The text ‘Liya’s Candles’ large at the top centre, a simple white sans-serif font”. The shorter the phrase, the lower the risk of mistakes: models write 2–5 words cleanly almost every time. And add “no other text”, or the model sometimes invents captions and logos nobody asked for.
Cyrillic and plan B
Most models learned mostly from Latin script, so Russian letters come out worse than English ones: similar letters get mixed up, soft signs go missing. Plan A: a short phrase, a model that’s good at text, and checking every letter. Plan B is safer for anything important: ask for a picture with no text but with empty space for the title, and add the lettering yourself in Canva or any editor. That way Liya keeps her own font and zero typos.
Infographics and diagrams
New models can turn notes into a diagram: Google says outright that Nano Banana 2 makes infographics and data visualisations from text. Liya asks for “how I make a candle in 5 steps”. But the model may draw six steps instead of five, mix up the order or the numbers on a chart. So the rule is iron-clad: check every number and label on an infographic against your own data, and fix mistakes with a follow-up line in the same chat.
A poster people actually read
For the fair poster Liya thinks less about beauty and more about what a person gets in two seconds. One main headline of up to 5–7 words, one picture in the centre, and the date and place in small print. In the prompt she asks directly: “leave empty space at the top for a headline and at the bottom for the date”. Then the model keeps details out of those areas, and the text sits on a clean background instead of over the candle flame. The fewer the elements, the better the poster reads from afar.
Don’t redraw your logo every time
If you ask the AI to draw the logo in every post, it’ll come out slightly different each time and the brand starts to “drift”. The right way: make the logo once, ideally as a vector, for example in Recraft, and then insert the finished file. One more trap: an AI logo can accidentally resemble someone else’s trademark. Before printing it on her tins, Liya runs a reverse image search and checks that no such brand exists.
Check yourself
- The poster says “Liya’s Candels” instead of “Liya’s Candles”. What now?
Correct answer: Ask to redo the exact text in quotes, or add the lettering herself in an editor. A typo in the brand name hurts trust. Exact text in quotes or your own lettering in an editor fixes it. - AI drew a “Candle prices” infographic, and one price is wrong. What matters most?
Correct answer: Check every number against her own table and fix it. The model draws numbers as images and can distort them. A wrong price in a post means trouble with buyers. - Liya wants her logo on every post. What’s the best way?
Correct answer: Make the logo once and insert the finished file. A brand is recognised by a consistent logo. The finished file is always identical and error-free.