Translate Text from Images: 3 Methods and Where They Fail
Three ways to translate text from photos, five situations where recognition fails, and what to do with documents and amounts.
👁 117 reads
Translating text from images is one of those features that works better than most people expect. You point your camera at a menu, sign, flyer, or contract, and the text appears in your language. No need to type anything out, and with alphabets you can’t even input on a keyboard, it’s the only practical option.
Three Methods and the Difference Between Them
Camera in translator app. The fastest method for short items: signs, menus, packaging. Translation overlays on the live image, so you see it immediately. Works well when traveling and also functions offline if you download the language pack beforehand.
Photo sent to a chat. Slower, but significantly smarter. The model understands context, so it doesn’t literally translate phrases that shouldn’t be translated word-for-word, and you can ask follow-up questions: what does this mean in practice, is this an important clause, what’s this about. For contracts, forms, and official documents, this is the only method that makes sense. I describe how to do this in the guides to ChatGPT and Gemini.
Text recognition, then translation. With multi-page scans, it’s better to first extract the text, correct recognition errors, and only then translate. More work, but with twenty pages it’s the only way to get a readable result.
Where This Breaks Down
- Handwriting. Handwritten text recognition is still unreliable, especially in languages with diacritics. Treat the result as a suggestion, not as an accurate reading.
- Photo at an angle. A crooked frame ruins recognition more than poor lighting. Shoot perpendicular and up close.
- Text over graphics. Text on photos and colored backgrounds gets lost most often.
- Numbers and amounts. Errors here cost the most, and they happen regularly. Always visually verify amounts, dates, and numbers, regardless of the tool.
- Special characters. When translating to languages with diacritics, special characters can be lost in low-quality prints.
How to Get Good Results
Three habits solve most problems. Photograph perpendicular, filling the frame with just the text. With documents, photograph fragments instead of the entire page from a distance, because letter resolution determines everything. And with every important document, tell the model explicitly what you want: “translate literally and separately list all amounts, dates, and deadlines” will give you a result you can verify, instead of a smooth summary where you don’t know what was omitted.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →What to Remember with Documents
A photo of a contract, test result, or identity document sent to a chat ends up on someone else’s server. With documents containing personal data, it’s worth covering what’s unnecessary, or using a solution that runs locally, which I describe in the guide to Ollama. And an obvious thing that’s easy to forget: machine translation has no legal force. For government offices and courts, you need a sworn translator, and these tools are for understanding content, not for official translation.
More on this from QUANTSLANT
guides·ChatGPT 4: What Happened and What Replaced ItGPT-4 disappeared from ChatGPT, but the name stuck. Here's what happened to it, whether you lost something good, and which model to choose today.