Photo translator / Translate an image with ChatGPT
How it works
Translate an image with ChatGPT, and keep the layout.
ChatGPT reads the text in a screenshot, a menu or a banner and translates it as well as anything on the market. What it hands back is a paragraph. The three-column menu comes back as one column, the spec table comes back as a list, and the picture you actually needed is still in English. This page is about the missing half: getting those words back into the image, in the boxes they came from.
What ChatGPT does with an image, exactly
Upload a photo and ask for a translation and two things happen: it recognises the text, and it answers in prose. Both are genuinely good. But the answer is a message in a chat window, and it has no idea where on the image each line was sitting — the moment the words leave the picture, the layout is gone.
The obvious fix, asking it to produce the translated image, is worse. That is a different model doing a different job: it redraws the picture from scratch. The product changes shape, the logo drifts, the badge moves, the brand colour is nearly right. For a menu you are reading once, who cares. For an asset that someone approved, it is a new file that needs approving again.
So the useful division of labour is the boring one. The model is superb at the words. Nothing about your image needs a model at all — the text has a position, a size and a background colour, all of which are measurable.
The missing half: putting the words back
That is what this site is: OCR finds each text block and its rectangle, the background behind it is sampled and rebuilt, and the translated line is re-typeset into the same box — shrunk, wrapped and re-centred until it fits. Only the text pixels change. The crop, the artwork and the pixel dimensions are untouched, so the file is still the file you started with.
The translating itself never happens here. This site sends your text to no AI service at all; it writes the prompt, and you run it in the chat tab you already have open.
The round trip, in four steps
- Drop the image in and press “Detect text”. Every string is found and listed, each one tied to the rectangle it sits in. You can fix a misread line or drag a box the OCR missed.
- Pick your languages and press “Translate with ChatGPT”. A prompt goes to your clipboard with the text from your image already filled into a table, and a ChatGPT tab opens.
- Paste it, send it, copy the reply. Because the original column is already filled in, the model can only fill in the blanks — every row comes back matched to a row you sent.
- Paste the reply back. The grid fills itself as the text lands. Edit any cell, then export one PNG per language, or a ZIP of all of them.
Nothing in that loop is specific to ChatGPT. The prompt is plain text: Claude, Gemini, a local model or a human translator with a spreadsheet all produce something the grid reads.
Why you go to ChatGPT yourself
It would be less clicking if this site called an API for you. It deliberately does not, and the three reasons are the whole design:
- Your words never leave. A downscaled copy of the image goes out for text detection and nothing else about your file does — not the full-resolution original, not one character of your copy. Unreleased pricing and unlaunched product shots stay on your machine; here is the request-by-request list and how to verify it in DevTools.
- No quota to ration. Translation credits are the one thing free tools have to meter, and metering them is why they eventually ask for an account. The translating runs on your own ChatGPT plan, so there is nothing here to buy, and no sign-up.
- You can argue with the model. “Keep it under twelve characters.” “Use our glossary.” “That is too formal for a promo banner.” A hidden API call takes the first answer; a chat window lets you take the fourth.
When ChatGPT on its own is enough
Often. If you are standing in front of a menu and want to know what you are ordering, the plain-text answer is the answer — take the photo, ask, done. The same goes for understanding a foreign error message, a form or a receipt.
The image only matters when someone else has to look at it: a store listing, an app screenshot in a help centre, a promo tile for a new market, a slide going to a client. That is the case where a paragraph in a chat window is not a deliverable, and where a translated listing image or a batch of them is what you are actually being asked for.
Questions
Can ChatGPT translate the text in an image?
Yes, it reads the text and translates it well. What it returns is prose: a list of the lines it found, in the new language. It does not return your image with the words replaced, because it edits text, not pixels — and asking it to generate the image instead redraws the whole picture.
Why does this site not call the ChatGPT API for me?
Three reasons. Your wording never leaves your browser, so nothing you are translating is sent to this site or to any AI service by it. There is no per-user quota to buy or ration, because the translating runs on your own ChatGPT account. And you can argue with the model — ask for a shorter line, a different tone or your own glossary — which an API call fired behind your back cannot do.
Do I need a paid ChatGPT plan?
No. The prompt is plain text and the free tier answers it. Nothing here is tied to a model or a plan: any assistant that can fill in a table works, including Claude, Gemini or a local model.
What if ChatGPT reformats the table it sends back?
Paste it anyway. The reply is read as CSV, TSV, a markdown table, JSON or a fenced code block, and the chit-chat above and below it is discarded. Rows are matched back by their original text, so the only reply that cannot be used is one where the model rewrote the original column as well.
Can I use it without ChatGPT at all?
Yes. Download the CSV, Excel or JSON template — it already contains every string found in the image — fill it with your own approved copy, and upload it back. Or type the translations straight into the grid.