Why does the text on my AI images come out garbled, and what are my friends using?
Posted 4 May 2026Standard
The text comes out garbled because most AI image generators were trained on photos and art, not on writing. The older or art-focused tools never really learned to render letters and still get them wrong. A handful of newer text-friendly tools are now reliable, and your friends are almost certainly using one of those. The two most common are Ideogram, built specifically for text in images, and ChatGPT's image generator (the current model is called GPT Image 1.5).
Why this question matters
Birthday cards, "happy 50th" graphics, funny memes, and social posts all have one thing in common: they need words sitting on the picture. AI image generators only learned to handle text reliably in the last eighteen months, and only on certain tools. If the tool you are using is not one of them, no amount of careful prompting will fix it.
How I'd approach it
For anything with words on the image, lean on one of two tools.
Ideogram was built specifically to handle text inside images and it is still the most reliable. The free tier is enough to test it, and the paid plan is around US$8 a month. If you mainly want birthday cards, posters, and graphics with words, this is the simplest answer.
ChatGPT's image generator is the other strong option. If you already pay for ChatGPT Plus, you have it included. Tell it the image you want and the exact words you want on it, and it will usually get the spelling right on the first try. The thing ChatGPT does that Ideogram does not do as smoothly is iteration in conversation. "Make the cake bigger and change the name to Margaret" works the way you would expect.
For the artwork itself, with no text, the leading art-focused tools (Midjourney is the best-known) still win on raw quality. A common pattern for a birthday card is: generate the artwork in an art-focused tool, then add the text in a second pass with a tool that handles text, or layer it on top in Canva or Apple Photos.
Two small practical points. First, spell-check the prompt itself. Image models will faithfully render your typo if it is in your prompt. Second, keep the text short. "Happy 50th, Margaret" works on any of these tools. A whole verse of a poem will struggle everywhere.
What I'd avoid
I would avoid fighting an art-focused tool to do text. Some have improved (Midjourney's latest version, V8 Alpha released March 2026, is better at letters than older versions were) but the dedicated text-handling tools are still ahead on this one specific job. The honest answer is to use the right tool for the job and stop trying to force one tool to do everything.
Got a question?
Send it through the feedback link. No signup, no list. I'll add it to the queue.