GPT Image 2.5 does render Chinese characters inside an image, but no official page promises they will come out correct, and the distance between those two statements is where most guides turn vague.
What the official record does contain is a limitation and a set of prompt constraints, and both are more useful than a promise.
What the official record says about Chinese text
Start with what is written down, because the surrounding coverage is confident in ways the source material is not.
- The launch announcement has ten sections and none of them discuss text rendering or typography quality
- It names no language at all: not Chinese, Japanese, Korean or any other non-Latin script
- The closest detail is in the alt text of one sample image, which mentions Japanese lettering, and that describes a picture rather than a capability
- The eight-page system card mentions typography zero times and names no language either
- Its one relevant capability sentence is that the model “improves infographic accuracy and layout”
- Claims that OpenAI specifically tuned Chinese rendering have no counterpart in either document
None of this means Chinese comes out broken. It means the ceiling is undocumented, so you plan around prompt rules instead of a guarantee.
The four limitations OpenAI does document
The API guide carries a limitations list, and it is the only place where the company discusses in-image text quality at all.
- Text rendering: “Although significantly improved, the model can still struggle with precise text placement and clarity”
- Latency: “Complex prompts may take up to 2 minutes to process”
- Consistency: the model “may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations”
- Composition control: “the model may have difficulty placing elements precisely in structured or layout-sensitive compositions”
Read the first bullet twice. Improved is not the same as reliable, and dense small type is exactly where clarity gives way.
Put the exact characters in quotation marks
The official Codex image generation page gives the clearest instruction on this, close enough to quote as it stands.
- Keep in-image text short and “put the exact text in quotation marks”
- Preserve the capitalisation you want instead of letting the model choose
- Describe font style, size, colour and placement as explicit constraints
- “For an uncommon name, spell out the letters when accuracy matters”, which covers rare characters and place names too
- State whether any other text is allowed, so stray lettering does not appear
- The prompting guide repeats it: put literal text in quotes or capitals, and spell tricky words letter by letter
The official example prompt shows the shape, and it is worth copying: “Add only the title “SPRING WORKSHOP” in large, bold, white sans-serif letters, centered in the top third of the image. Keep the title on one line.”
Keep the string short and declare the language
Length is the variable you control most directly, and language is the one almost nobody writes down.
- Short beats complete: a headline and a short subtitle survive far better than a paragraph
- One line of text per instruction, so each string gets its own placement
- Name the language explicitly, as in “all text in Simplified Chinese”, rather than assuming it
- Give the script as well as the language when the two can differ, since that ambiguity is yours to remove
- Small dense type is the documented failure mode, so raise the size and cut the word count
Nothing here is specific to Chinese: the same rules decide whether a German headline or a Korean label arrives intact.
A workflow for posters, cards and interface text
A repeatable order beats a longer prompt, because you can see which change moved the result.
- Start from a template so the layout is already correct, as we describe in GPT Image 2.5 templates
- Write the copy outside the tool and paste it in, so the characters are never retyped by the model
- Split the layout request and the text request when a frame has missed twice
- Ask for one text block per round rather than the whole poster set
- Reserve the zones in the prompt before you add words, which leaves fewer placement decisions
- Keep the quoted strings in a file when you build a series, so every run pastes the same text
When the characters still come out wrong
Rewrite the text instruction rather than the whole image, because the rest of the picture is usually fine.
- Fix the characters in place with a comment instead of regenerating, which is what GPT Image 2.5 comment editing is designed for
- Change one variable per attempt: the string, then the position, then the font constraint
- Shorten the line before changing anything else, since clarity is the documented weak point
- If two characters keep swapping, spell them out letter by letter in the prompt
- Expect a second pass, because the prompting guide itself says “small wording/layout tweaks usually improve legibility”
A worked example: a short Chinese headline on a poster
Text rendering is the thinnest part of the official record, so the workflow has to be defensive. Take a poster that needs one Chinese headline and nothing else.
- Write the characters between quotation marks, never as a description: 「春日限定」
- Declare the language and the script in the same sentence, for example: the headline reads 「春日限定」 in simplified Chinese; render exactly these four characters.
- Keep the string short. Four characters is a safe length; a full sentence is not
- Say where the text goes and how big it is: “Centred in the upper third, the widest element in the frame.”
- Render once, read the result character by character, and change one character at a time if it is wrong
The figure above is the useful reference point: an ornate frame only works because the lettering inside it stays readable. When a string comes back wrong, the cause is usually that the frame was described in far more detail than the characters were.
If the text has to sit inside a layout you have already drawn, the GPT Image 2.5 Sketch guide covers supplying that structure.
Frequently asked questions
Does GPT Image 2.5 still garble Chinese text?
No official page gives an accuracy figure, and the API guide warns that text placement and clarity can still be unreliable. Treat every string as something to check rather than something guaranteed, and plan a correction pass for dense layouts.
Why do some articles say OpenAI fixed Chinese text?
The launch announcement and the system card contain no language-specific claim at all. Statements circulating in coverage do not appear in either document, so read them as third-party impressions rather than an official position.
How long can the text in one image be?
The official guidance is to keep in-image text short. A headline plus a short subtitle is realistic, while a paragraph of dense body copy is the case the API guide flags as unreliable.
Can I specify Traditional Chinese characters?
Yes. Name the script explicitly, the same way you name the language, and quote the exact characters. Script and language are separate constraints, and stating both removes an ambiguity the model would otherwise resolve on its own.
Text inside an image is not a special Chinese problem, it is the strictest case of one rule: quote it, shorten it, place it explicitly, and correct it in place.