77Fox translates a screenshot on Windows in place: press Alt+A, drag over the area and click Translate on the screenshot toolbar. 77Fox recognizes the text on your PC, then paints each translated paragraph over the spot where the original was. If you only need the words, click Extract text and copy them. Both tools are part of the 77Fox Windows desktop app; the web console doesn't have them.
A settings page in a language you can't read, a game menu, a scanned PDF page, an overseas hosting panel, a chat screenshot from a client: the text is pixels, so you can't select or copy it, and an online translator wants the whole image uploaded first. That's the case screenshot translation is for.
Screenshot translation, Extract text or page translation?
| What you're looking at | Use | What you get |
|---|---|---|
| A web page whose text you can select | Page translation in the 77Fox browser extension | Translations inserted after the original text; the page content isn't rewritten |
| A screenshot of an app, game menu or admin panel | Screenshot translation | Translations painted over the matching spots; copy or save it as an image |
| Text inside an image, PDF screenshot or chat capture | Extract text | Plain text you can copy and paste |
| Translation next to the original | Full translation, inside screenshot translation | Each translated paragraph with its original underneath; Copy all translations takes every translation at once |
If the page text is selectable, page translation is less work. It lives in the 77Fox browser extension, which you install through the desktop app; see page translation in the browser extension and how to choose a page translation engine.
How to translate a screenshot in four steps
- Press Alt+A and drag across the area you want. Hover over a window or control to snap to it; the scroll wheel changes the snap level, and right-click or Esc cancels. Alt+A is the default; change it under Settings → Tools → Screenshot.
- Click Translate on the toolbar next to the selection (the toolbar shows icons only; hover to see each name). The first time, 77Fox asks you to download the text recognition model, about 31 MB.
- A translation bar appears on the outer side of the toolbar, showing "Auto → English" and the current engine. Once recognition finishes, translated paragraphs land in batches, with progress counted in paragraphs, such as 3/12. You can stop at any point; finished paragraphs stay, and you can translate the rest later in the same screenshot.
- Press Enter or Ctrl+C to copy the whole image, or Ctrl+S to save it as a PNG. To grab a few sentences, drag across the text on the image and press Ctrl+C; only the selected words are copied.
To switch the target language or engine, click it on the translation bar and 77Fox starts again from the original text. Don't have the app yet? Download 77Fox for Windows.
Copy text from a screenshot with Extract text
Click Extract text and 77Fox recognizes the words in your selection on your own PC. They appear in an Extract text panel beside the selection. Copy text takes everything, and lines that were only wrapped are joined back into paragraphs. You can also drag across words on the image, or double-click one, and press Ctrl+C to copy just that part. More on this in Extract text.
Recognition uses the PP-OCRv6 model running on your CPU, so no dedicated GPU is needed. Screenshot pixels are never uploaded, and the recognized text isn't written to logs or synced. The recognizer starts when you need it and exits after five idle minutes.
One recognition pass has three limits: 60 seconds, a selection of about 25 megapixels, and 300 text boxes. A selection that's too large gets the message "Selection too large. Select a smaller area." On a dense, tall screenshot, anything past 300 boxes is skipped, so capture it in several parts.
The recognizer was tested with Chinese and English samples during development. How well it reads other scripts depends on the model itself, and 77Fox makes no promise there.
How the translation goes back onto the image
77Fox doesn't retouch the picture. For each paragraph it samples the colors around the text, lays an opaque solid block over the original words, and writes the translation in black or white, whichever contrasts with the block. On flat app interfaces the block matches the background; on photos, gradients or patterns you'll see a flat patch.
Type size follows the original line height, and the translation tries to sit on the original lines. If it doesn't fit, the text shrinks and spreads into nearby empty space without covering other text. If it still doesn't fit, the Full translation panel opens outside the selection when the run finishes: every translated paragraph with its original underneath, plus Copy all translations. Don't want it popping up? Turn off "Show the full translation when it doesn't fit" in the screenshot translation settings.
Paragraph splitting decides how readable the result is. Divider lines, bullets and numbers, and differences in size or color between a heading and body text all count as paragraph breaks, so list items and table rows separated by lines usually get translated one by one instead of as one blob. Wrapped body text is still joined into one paragraph.
The bar switches between Original and Translation. Copy and save output what you're currently looking at, without the recognition boxes. Arrows and boxes you draw sit on top and are exported too.
Which translation engine to pick
Click the engine name on the translation bar to open a two-column picker: Recent and five source groups on the left, that source's models on the right. The default is the free channel.
| Engine group | Cost | Where the recognized text goes | Good for |
|---|---|---|---|
| Free channel (Google's public endpoint) | Free, the default | Google's public translation endpoint | The odd screenshot, zero setup |
| 77Fox Cloud | Needs sign-in and a plan (from £1.49/month); uses AI credits | 77Fox's cloud service, where a large language model translates it | LLM translation without managing keys |
| Own Key | No 77Fox credits; billed by your provider | The provider you configured, such as OpenRouter, DeepSeek or DeepL | People who already have an API account |
| Local | No credits | Stays on your PC | Contracts, client files, internal documents, or no internet |
| Local CLI | No 77Fox credits; uses the account you signed in with in that tool | Codex CLI, Claude Code, Grok CLI or OpenCode on your PC; you pick the model in 77Fox and the tool calls it with its own sign-in | People who already live in those command-line tools |
There are five local models to choose from, with different download sizes and GPU support; see choosing an offline translation model.
A few details:
- A source that isn't set up can't be selected. The right column shows one button that fixes it, such as Set up Key, Download model, Sign in or Open settings.
- Your own keys are shared with page translation and stored encrypted on your PC, so you don't enter them twice.
- When the free channel is rate limited, the bar counts down and Retry stays disabled until then, with Switch engine next to it. 77Fox never swaps engines behind your back.
- The source, model and target language you pick in a screenshot, plus your three most recent engines, become the next screenshot's defaults; page and prompt translation keep their own choices. You can also set the default under Settings → Tools → Screenshot → Screenshot translation (the gear on the bar opens it). See the screenshot translation overview.
What it can't do
- Vertical text isn't recognized, and neither is strongly slanted or perspective-skewed text.
- Complex tables with interleaved columns may come out in the wrong reading order.
- There are nine target languages, matching the 77Fox interface: Simplified Chinese, English, Japanese, Korean, German, French, Italian, Spanish and Hebrew. The default follows the interface language, or you can pin one in settings.
- It translates the screen at the moment of capture. When a game or video moves on, take a new screenshot; the overlay doesn't follow the picture.
- The cover is a solid block. Background texture isn't rebuilt and the original font isn't matched. Very long translations can be cut off at the minimum size; the full text is in Full translation.
FAQ
Is my screenshot uploaded?
No. Recognition happens on your PC and the pixels are never uploaded; the recognized text isn't logged or synced either. Extract text sends neither the image nor the text anywhere; the only network use is downloading the recognition model the first time. Translate sends only the recognized text, and only to the engine you chose. With a local model, even that text stays on your PC.
Does it work offline?
Recognition needs no connection once the model is downloaded. For offline translation, pick a local model and download it in advance; the free channel, 77Fox Cloud, your own keys and Local CLI tools all need internet. See choosing an offline translation model.
Can it read Japanese or Korean?
77Fox makes no promise for them. The recognizer was tested with Chinese and English samples, so try Japanese or Korean on one screenshot first. What recognition can read and the nine languages translation can produce are separate things.
Does it cost anything?
Recognition is free, and translation uses the free channel by default. 77Fox Cloud needs a plan, from £1.49 a month, and uses AI credits; see pricing. Your own keys, local models and Local CLI tools don't use 77Fox credits.
The recognition model won't download. What now?
The model is about 31 MB and comes from Hugging Face. On networks where Hugging Face is restricted, the download can fail; click Download models in the panel to try again. You can also download or delete the model under Settings → Tools → Screenshot → Screenshot translation.