# Text in AI images: which models spell well and the prompts that help

> To get readable text in an AI image, put the exact words in quotes, say where they go and how they look, ask for them once and add "no extra text", then check every letter. GPT Image, Gemini, Qwen-Image and Ideogram 4 each document their text handling, and OpenAI's page says placement and clarity can still go wrong.

Mindket editors · Published 2026-11-06 · Updated 2026-11-06 · https://mindket.com/learn/ai-design/text-in-ai-images

## In short

- Write the exact words in a note first, then paste them into the prompt inside quotes.
- Say where each text sits and in what type style, and ask for it once.
- Choose the model by the script you need: Latin letters, Chinese characters or a layout with many text boxes.
- Compare every letter with your note at full size before the image goes anywhere.

## Which model for which text

An image model draws letters as shapes, and a wrong shape is a misspelling. OpenAI's [image generation guide](https://developers.openai.com/api/docs/guides/image-generation) lists text among the limits of its model: "Although significantly improved, the model can still struggle with precise text placement and clarity." The [prompting guide](https://developers.openai.com/api/docs/guides/image-prompting) adds that the `quality` setting "controls how much visual detail the model can render, such as texture or small lettering".

Four families of models document how they handle text. The table gives what each maker's own page says, and each row is that maker's claim, not a measurement of ours.

| Model | What its own page says about text | Where the page says to use it | Licence or access |
| --- | --- | --- | --- |
| GPT Image 2.6 (Flare and Sunburst) | Put the required wording in quotes, describe its position and typography, and compare medium or high quality for small text. The generation guide says placement and clarity can still be a struggle. | OpenAI's Image API. A community post of 4 November 2026 introduces ChatGPT Images 2.6. | API Organization Verification may be needed |
| Gemini: Nano Banana Pro and Nano Banana 2 | Google's quickstart calls Pro the flagship with "complex typography rendering". It says Pro and 2 can "generate and translate complex text, diagrams, and infographics in dozens of languages". | The Gemini API, as `gemini-3-pro-image-preview` and `gemini-3.1-flash-image` | Billing gives higher rate limits |
| Qwen-Image | The model card reports "exceptional performance in text rendering, especially for Chinese". A newer model, Qwen-Image-2.0, launched on 10 February 2026 according to the README, which says it supports 1k-token instructions for posters, slides and infographics. | Qwen-Image: weights on Hugging Face. Qwen-Image-2.0: the README points to Qwen Chat online and a blog, with no weights link | Apache 2.0 licence for Qwen-Image. The README states no licence for Qwen-Image-2.0 |
| Ideogram 4 (released 3 June 2026) | The model card claims "best-in-class multilingual text rendering" and takes text as JSON elements with optional boxes. | Gated weights on Hugging Face, and ideogram.ai online | Ideogram 4 Non-Commercial licence |

## Write a prompt that gets the words right

OpenAI's [image prompting guide](https://developers.openai.com/api/docs/guides/image-prompting) has an ad prompt for a streetwear brand called Thread. It quotes the words, says how often they appear, says where they go ("integrated into the ad layout") and forbids anything else.

Copy its pattern and swap in your own tagline.

````text
Give me a cool in culture ad / fashion shot for a brand called Thread.
It's a hip young street brand. The ad shows a group of friends hanging out together with the tagline "Yours to Create."
Make it feel like a polished campaign image for a youth streetwear audience: stylish, contemporary, energetic, and tasteful.
Use clean composition, strong color direction, natural poses, and premium fashion photography cues.
Render the tagline exactly once, clearly and legibly, integrated into the ad layout.
No extra text, no watermarks, no unrelated logos.
````

1. Write the words in a note first and proofread them there. The model can only copy what you gave it.
2. Put each text in double quotes inside the prompt, then say where it goes and how it looks, for example a bold, rounded headline across the top.
3. Say how many times it appears. OpenAI's example asks for the tagline exactly once.
4. End with the exclusion line from the example: no extra text, no watermarks, no unrelated logos.
5. For small or dense text, spell unusual words and brand names letter by letter, and compare medium and high quality. Both tips are in OpenAI's prompting fundamentals.
6. Zoom to full size and compare each word with your note. When one word is wrong, follow up with the correct spelling and the details to keep, as [the image prompts lesson](https://mindket.com/learn/ai-images-video/ai-image-prompts) shows.

## Pin text to a spot with Ideogram 4's JSON prompt

Ideogram 4 takes its prompt as JSON, a text format of labeled fields. Each piece of text becomes its own element with the literal words in a `text` field and a description of how it looks in `desc`. The [prompting guide](https://github.com/ideogram-oss/ideogram4/blob/main/docs/prompting.md) gives this business card, which has two text elements.

An element can also carry a `bbox`, a box written as `[y_min, x_min, y_max, x_max]` on a 0 to 1000 scale from the top-left corner. The guide marks it optional. The `color_palette` list takes uppercase hex colors, up to 16 for the whole image.

Two Ideogram pages disagree about plain sentences. The model card says plain-text prompts work but do worse. The prompting guide says plain text sent straight to the model will not work and should first go through a "magic prompt", an LLM step that expands a short prompt into this JSON. Both say JSON gives the best results.

The licence matters for paid work. The [model card](https://huggingface.co/ideogram-ai/ideogram-4-fp8) lists Ideogram 4 under its non-commercial licence, so read it before you put the weights behind a client project.

````json
{
  "high_level_description": "A clean, modern business card layout for a tech company.",
  "style_description": {
    "aesthetics": "minimal, professional, geometric",
    "lighting": "even, diffuse studio lighting",
    "medium": "graphic_design",
    "art_style": "flat vector design, generous whitespace, sans-serif typography",
    "color_palette": ["#FFFFFF", "#F0F0F0", "#333333", "#0066FF", "#00CC88"]
  },
  "compositional_deconstruction": {
    "background": "A solid off-white card surface with subtle paper texture.",
    "elements": [
      {"type": "text", "text": "ACME TECH", "desc": "Bold dark grey sans-serif company name across the upper third of the card."},
      {"type": "text", "text": "hello@acme.tech", "desc": "Small blue sans-serif contact email near the bottom of the card."}
    ]
  }
}
````

**First-place picks in ContraLabs' blind typography test** Ten professional designers picked the best of four models. These are the figures on Ideogram's own model card, for its own model. They add up to more than 100, and the card does not say why. It gives no date for the test.

- Ideogram 4: 47.9%
- Nano Banana 2 (Gemini 3.1 Flash Image Preview): 30%
- FLUX.2 [max]: 15.5%
- `Grok Imagine 1.0`: 15%

Source: [Ideogram 4 model card on Hugging Face](https://huggingface.co/ideogram-ai/ideogram-4-fp8) (read 2026-11-06)

## Chinese, Japanese and other scripts

For Chinese, Qwen-Image is the model whose own pages lead with it. Its [model card](https://huggingface.co/Qwen/Qwen-Image) says "especially for Chinese", and its [README](https://github.com/QwenLM/Qwen-Image) shows a Chinese edit prompt for a photo of a woman. In English it reads: this girl looks at the TV screen in front of her, and the screen shows the words "Alibaba". The README pairs it with an input photo.

Google's quickstart says Nano Banana Pro and 2 can generate and translate text in dozens of languages, without naming which. Check a script you rely on by sending a short sample first.

````text
这个女生看着面前的电视屏幕，屏幕上面写着“阿里巴巴”
````

- To translate text already inside an image, OpenAI's guide uses this prompt: "Translate the text in the infographic to Spanish. Do not change any other aspect of the image." It then tells you to check the translation and any words left in the original language.
- For a Chinese or Japanese sign, write the characters in quotes exactly as they should appear and state the script in the prompt.
- Qwen-Image runs in ComfyUI, which the README says supports it natively. The [ComfyUI tutorial](https://mindket.com/learn/ai-images-video/comfyui-tutorial) covers that tool.

## Common mistakes

- Describing the text instead of quoting it, so the model has no exact words to copy.
- Asking for the same line twice or for text on three objects. Each extra text is one more thing to misspell.
- Checking spelling in the chat preview. Zoom to full size first.
- Settling for the first try. Generate several and keep the one with the right letters.
- Skipping the licence line. Ideogram 4's weights are non-commercial, and Qwen-Image's model card says Apache 2.0.
- Baking changing copy into the picture. A price list that changes weekly belongs in a design app, with the AI image as the background.

## Try it yourself

Use real copy from a post or poster you need: a short headline and a longer line under it.

- Draft both texts and have a second pair of eyes check the spelling.
- Run the Thread-style prompt with your headline and subline in one app, and in a second app if you have one.
- Compare each letter with the note at full size and write down every miss.
- Run it again with the hardest word spelled letter by letter, at a higher quality if the app offers one.
- Ask for the finished image in a second language and check it for words left in the first.

> You are done when every word matches the note at full size, in both languages.

## Next: one design, many sizes

A poster is one picture. A launch needs a square post, a tall story and a wide banner that all look like the same brand. [The next lesson](https://mindket.com/learn/ai-design/ai-social-graphics) covers sizes and a repeatable brand look.

The first lesson, [how to make a logo with AI](https://mindket.com/learn/ai-design/ai-logo-design), is where the spelling check started.

## Check yourself

1. Which prompt line follows OpenAI's advice for exact text?

   A. A banner that says something about a sale
   B. Render the tagline "Yours to Create." exactly once, with no extra text
   C. Make the lettering perfect

   Answer: B. OpenAI's example quotes the words, says how many times they appear and rules out extra text.

2. You need small text on a dense infographic. What does OpenAI's guide suggest comparing?

   A. Fewer words in the prompt
   B. A square size
   C. Medium or high quality

   Answer: C. The guide suggests comparing medium or high quality for small text, dense information or several fonts.

3. In an Ideogram 4 JSON prompt, what goes in the text field of a text element?

   A. The literal words to render
   B. A description of the font
   C. The box coordinates

   Answer: A. The text field holds the literal words, while desc describes how they look and bbox gives their place.

4. You want to use open weights for a paid client poster. Which fact decides whether you can?

   A. The model's parameter count
   B. Whether the model is also hosted online
   C. The quality setting
   D. The licence on the model page

   Answer: D. Ideogram 4's weights carry a non-commercial licence, while Qwen-Image's model card says Apache 2.0.

## Frequently asked questions

### Why does AI still misspell words in images?

No page gives a full reason, but OpenAI names small text, dense information and multiple fonts as the cases to compare at medium or high quality. Shorter text in one font is less to check, but every word still needs a check at full size.

### Which AI image generator is best for text?

No page ranks them for you. Ideogram reports a blind typography test where designers picked Ideogram 4 first most often, and Google, OpenAI and Qwen describe their own text handling. Run your own words, in your own language, through two models and compare the letters.

### What if I cannot read the script in the image?

Ask a reader of that script to check the result before you use it, and send a short sample first. A wrong stroke or accent looks right to someone who cannot read the language, and a maker's claim does not cover your wording.

### Can I fix one wrong word without redoing the image?

You can try a follow-up. OpenAI's guide says to write "change only X" and list the details to preserve, so name the wrong word, give the right spelling and list what to keep. It warns that repeated edits can still change details you meant to keep, and says to composite the approved edit into the original if part of the image must stay pixel-identical.

### Can I run Ideogram 4 on my own computer?

The model card lists a 9.3 billion parameter model in two builds, nf4 for CUDA and fp8 for all hardware. The weights are gated: accept the licence on Hugging Face, create an access token and log in with `hf auth login` before the code can download them.

## Sources

- [OpenAI API docs: Image prompting](https://developers.openai.com/api/docs/guides/image-prompting) (read 2026-11-06)
- [OpenAI API docs: Image generation](https://developers.openai.com/api/docs/guides/image-generation) (read 2026-11-06)
- [OpenAI Developer Community: Images, upgraded, introducing ChatGPT Images 2.6](https://community.openai.com/t/images-upgraded-introducing-chatgpt-images-2-6/1412901) (read 2026-11-06)
- [Google Gemini cookbook: Get started with Nano Banana](https://github.com/google-gemini/cookbook/blob/main/quickstarts/Get_Started_Nano_Banana.ipynb) (read 2026-11-06)
- [Hugging Face: ideogram-ai/ideogram-4-fp8 model card](https://huggingface.co/ideogram-ai/ideogram-4-fp8) (read 2026-11-06)
- [GitHub: ideogram-oss/ideogram4 prompting guide](https://github.com/ideogram-oss/ideogram4/blob/main/docs/prompting.md) (read 2026-11-06)
- [Hugging Face: Qwen/Qwen-Image model card](https://huggingface.co/Qwen/Qwen-Image) (read 2026-11-06)
- [GitHub: QwenLM/Qwen-Image README](https://github.com/QwenLM/Qwen-Image) (read 2026-11-06)

[Browse skills](https://mindket.com/skills)