AIFont Text Rendering Ledger: did five image models spell the word right, in eight languages

Images from five text-to-image models (GPT Image, Nano Banana, Grok Imagine, Seedream, Qwen-Image), each asked to write a given word in a given lettering style, in English, German, Spanish, Portuguese, Simplified and Traditional Chinese, Japanese and Arabic. Every image is judged for whether the text is spelled exactly as requested; incorrect ones record the error type and what the image actually says. Each record carries the full prompt, the upstream model version at generation time and the generation settings. Nothing is retouched or cherry-picked: failed images stay in. Accuracy is meant to be counted on first attempts only, and data/summary.json is filtered that way. The lettering styles and their prompts are published at aifont.ai.