Guide
Does the language of your prompt matter?
Usually answered with a shrug and an assurance that modern models are multilingual. They are, up to a point, and the point is further away than you would like. Here are the numbers from 5,533 generations on this platform.
The measurement
Script matters more than language
Every generation on this platform carries the prompt that produced it, so the question is answerable rather than arguable. Grouping 5,533 of them by the writing system of the prompt and comparing how many finished:
| Prompt written in | Generations | Completed |
|---|---|---|
| English | 4,474 | 15% |
| A Latin script, not English | 568 | 11% |
| A non-Latin script | 491 | 3% |
The split people expect is English against everything else. That is not what the data says. Spanish, Portuguese, French, German and Italian land close to English — a gap worth knowing about but not one that decides whether you get a picture. Cyrillic, Arabic, Persian, Thai, Chinese, Japanese and Korean land at roughly a fifth of the English rate.
Two honest caveats
What these numbers do not prove
The absolute rates are depressed by something unrelated. Much of this period overlaps a stretch when the upstream providers were out of credit and almost everything failed regardless of what was written. That is why 15% is the English figure rather than something respectable. It affects all three rows roughly equally, so the comparison holds even though the absolute numbers are poor — and in the weeks since, with that fixed, English sits around 74% and Latin-script non-English around 64%, which is the same gap at a healthier level.
The non-Latin figure is the weakest of the three. It is a large sample over the whole period, but only ten non-Latin generations have run since the providers were funded, and none completed. Ten is far too few to quote a rate from. The direction agrees with the long-run number; the size of the gap at today’s reliability is genuinely not established, and anyone telling you otherwise from this data is overreading it.
Why
Captions in one language, tokens in another
Two mechanisms compound, and neither is about the model disliking your language.
The captions were English. A text-to-image model learns what words mean from the text attached to training images, and that text is overwhelmingly English. A French prompt still lands near the right meaning because French shares a great deal of vocabulary and structure with English, and much of it appeared in training anyway. A Thai prompt shares none of that surface, so the model is working from a far thinner and noisier association between those words and those pictures.
Non-Latin scripts cost more tokens. Text is split into tokens before the model sees it, and the splitters are fitted mostly to English. A word in Latin script is often one or two tokens; the same word in Korean, Thai or Arabic can be five or more. So an identical description is several times longer in tokens, it fills the attention budget sooner, and the end of it is dropped — the same mechanism as a long prompt being half ignored, arriving earlier than you expect.
What to do
Translate if your script is non-Latin; otherwise worry about the light
Writing in a non-Latin script: translate into English first. On this evidence it is the largest single improvement available, and larger than any rewording. A machine translation is fine — the model is not grading your prose, it is matching words to pictures, and plain translated English matches better than careful writing in a script the captions never used.
Writing in Spanish, French, Portuguese, German or Italian: translating is worth a try, but it is not your main problem. The gap is a few percentage points. Lighting, camera and pose are worth far more — naming a source and a direction will change a picture more than changing its language.
Keep proper nouns and technical terms in their usual form. Camera language is close to universal. Terms like 35mm, f/1.8, golden hour, bokeh and rim light appear in English in captions worldwide, so leaving them as they are is better than translating them into something the model has seen less often.
State ages and relationships explicitly, whatever you write in. This matters more outside English, not less: screening has to judge a description from what it contains, and an inference that survives a translation is a weaker inference. Leaving it unstated is the most common avoidable cause of a prompt being refused.
One more thing worth saying plainly, because it is the useful version of this advice. Translating is a workaround for a gap in the models, not a verdict on your language. The gap is narrowing as training data broadens, and nothing above is a reason to write in English anywhere except the prompt box.
Asked often
Short answers
- Does it matter what language I write my AI prompt in?
- Less than people fear if you write in a Latin script, and considerably more if you do not. Across 5,533 generations on this platform, prompts in Latin-script languages other than English completed at 11% against English at 15% - a real gap but a narrow one. Prompts in non-Latin scripts such as Cyrillic, Arabic, Thai, Chinese and Japanese completed at 3%, roughly a fifth of the English rate.
- Should I translate my prompt into English?
- If you write in a non-Latin script, yes - it is the single largest improvement available, and larger than any change of wording. If you write in Spanish, Portuguese, French, German or Italian, translating is worth trying but is not the main thing holding a result back; the usual suspects of light, camera and pose matter more.
- Why would the script change the result?
- Two reasons compound. The captions these models learned from are overwhelmingly English, and the other Latin-script European languages share enough vocabulary and structure that a description still lands near the right meaning. A non-Latin script shares none of that surface, and it is also broken into far more tokens per word, so a long prompt fills the attention budget sooner and the end of it is dropped.
Keep reading
The rest of the series
- Why AI changes the face every timeThe mechanism behind the problem: no memory, and a prompt that names a category rather than a person.
- How to get the same character in every imageFive approaches, what each costs, and the cases where training a LoRA is still the right answer.
- Keeping a character consistent in AI videoWhy video drifts twice over, and the two techniques that fight it from opposite directions.
Where this comes from
Unlimited Studio is built around this one problem
Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.