Datasets:

Funovate-Technology
/

aifont-text-rendering

license: cc-by-4.0
pretty_name: AIFont text rendering in AI images
language:
  - en
  - de
  - es
  - pt
  - zh
  - ja
  - ar
task_categories:
  - text-to-image
  - image-to-text
tags:
  - text-rendering
  - typography
  - evaluation
  - multilingual
  - image-generation
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: generations.jsonl

Mirror of Funovate-Technology/aifont-text-rendering (canonical copy, DOI 10.5281/zenodo.23815210). Corrections go to the GitHub repo's issues. Findings written up at aifont.ai/blog/best-image-generation-model-for-text.

Text in AI images: did the model spell it right?

Every image we generate for aifont.ai is checked letter by letter against the word the prompt asked for. This repo is that ledger: one record per image, with the prompt, the model, the word it was meant to write, and whether it wrote it.

 2,495   images, each with a text verdict
    66   lettering styles
     8   languages
     5   image models
   733   also reviewed by a person
    74   disputed, waiting for a native speaker

Correct / judged. First attempts with the current prompt of each published style, so every model got the same prompts:

Language	GPT Image	Nano Banana	Grok Imagine	Seedream	Qwen-Image
English	58/59	58/58	58/58	58/58	57/58
German	59/59	59/59	58/59	56/59	57/59
Spanish	59/59	59/59	58/59	57/59	49/59
Portuguese	59/59	59/59	55/59	59/59	50/59
Chinese (Simplified)	59/59	58/59	58/59	56/59	56/59
Japanese	59/59	58/59	52/59	51/59	45/59
Arabic	48/50	51/51	35/49	27/47	6/53

What a record looks like

{
  "id": 2172,
  "style": "witch-font",
  "language": "en",
  "target_text": "Potion",
  "model": "Qwen-Image",
  "model_id": "qwen-image-3.0-pro",
  "generated_at": "2026-10-28T17:05:33Z",
  "attempt": "first",
  "image_url": "https://cos.aifont.ai/research/attempts/witch-font/witch-font-en-qwen3pro-s1-92438.jpg",
  "gallery_id": 92438,
  "text_verdict": "correct",
  "error_types": [],
  "rendered_text": null,
  "reviews": [
    { "reviewer": "ai", "text_verdict": "correct", "native_speaker": false, "zoomed": false },
    { "reviewer": "maintainer", "text_verdict": "correct", "native_speaker": false, "zoomed": true }
  ],
  "human_reviewed": true,
  "disputed": false
}

Field by field in SCHEMA.md. The prompt text is in prompts.jsonl, keyed by prompt_hash.

How the images were made

Each lettering style has one prompt per language, written in that language, with a word in that language. The same prompt goes unchanged to every model.
Five models, called through the same service, at one fixed resolution and quality per model. model_id is the upstream model version at the moment of generation, read from the service's configuration, so a silent model swap shows up as a new id.
Nothing is edited, retouched or cherry-picked. Failed images stay in the ledger. When an image was wrong we usually generated it once more; that second image is marked "attempt": "retry".

How the text was judged

correct means every letter, stroke, dot and diacritic matches target_text. Upper or lower case alone does not make it wrong unless the prompt asked for one.
incorrect records what kind of mistake it was (error_types) and, for spelling mistakes, what the image actually says (rendered_text).
Every image was first judged by an AI assistant. A person then went through the ones most likely to be misjudged: every image the AI marked wrong or unsure, every Arabic image, CJK images the AI had only seen on a contact sheet, plus a random sample of the rest. The sample of 50 matched the AI on all 50, so the other AI-only verdicts are kept, with human_reviewed: false.
That person was not a native speaker of most of these languages. Where the AI and the person disagree on Arabic, neither is a native reader, so the record is unsure with disputed: true. Those are left out of the accuracy table.
reviews lists each reviewer's own call, so you can weigh them yourself. This is not a 100% verified ground truth, and it isn't presented as one.

Help fix it

If you read one of these languages and an image is judged wrong, open an issue with the record id and what the image actually says ("Wrong verdict" template). Native speakers are what this dataset needs most, Arabic first. Accepted corrections are added as a new review under your GitHub handle, and the next release recomputes the verdict.

Counting accuracy

Use first attempts only ("attempt": "first"). Retries exist because the first try failed, and extra samples were drawn after a correct one, so both lean toward correct. To compare models, also keep to the published styles ("purpose": "published") and the prompt version each style uses now: new styles were drafted on one model before all five ran. summary.json and the table above are filtered that way, and leave out disputed images. With about 60 images per cell, report counts, not percentages.

Files

File
`generations.jsonl`	One record per image
`generations.csv`	Same, flattened, without prompts
`prompts.jsonl`	Prompt text by hash
`summary.json`	Correct / incorrect by model × language, first attempts

Images are linked, not included. gallery_id is the image's public page on FontVibe, the service that generated it.

License

Data: CC BY 4.0. Credit "AIFont (aifont.ai)". Code in scripts/: MIT. See NOTICE.

Citing

DOI: 10.5281/zenodo.23815210 (all versions). See CITATION.cff, or GitHub's "Cite this repository" button.