TranslateGemma 4B Image
Google's TranslateGemma 4B applied to images: it reads the writing in a photo, screenshot or scan and returns it translated into the target language as plain text.
Overview
This page runs Google's TranslateGemma 4B on images. You give it a picture that contains writing, such as a sign, a menu, a book page or a screenshot, and it returns that writing translated into the language you choose. Reading and translating happen in a single model call, with no separate OCR step.
TranslateGemma keeps the image understanding of Gemma 3, the model it was trained from, even though its translation training used no image data. Google's report finds that the text-translation gains carried over to images mostly at the 12B and 27B sizes, while the 4B model scored only slightly better than the base Gemma 3 4B. On the Vistra benchmark, which translates English text in images into German, Spanish, Russian and Chinese, the card lists a MetricX score of 2.57 for the 4B model, where lower is better. Each image is resized to 896 x 896 pixels and encoded as 256 tokens before the model reads it.
Choosing a variant: TranslateGemma 4B is the same model for text you already have as text, which skips the reading step. TranslateGemma 12B and TranslateGemma 27B score higher on text translation, and their pages take typed text only.
What you can build
- Travel and shopping features that translate photographed signs, menus and product labels.
- Translation of screenshots from apps, games or chats where the text cannot be copied.
- Reading foreign-language receipts, forms or notices that users upload to a support flow.
- Translation of short printed passages, such as a paragraph from a handout, cropped to the text.
Inputs
- An image with text in it, uploaded as a file or passed as a URL. The upload field takes JPG, PNG, WebP, GIF and HEIC.
- The language written in the image and the language you want back, both chosen from the same menu of 54 languages.
- Maximum New Tokens (advanced): the length cap on the answer, from 1 to 2,048, default 200.
Outputs
The translated words come back as a plain-text file. Google's image prompt tells the model to return only the translation, to leave out where the text sits or what else is in the picture, and to ignore symbols, pictograms and arrows. The image itself is left unchanged, and layout and fonts are not reproduced.
Recommended settings
- Crop to the text you need. Google's image benchmark used pictures that hold a single piece of text, and every image is scaled to 896 x 896, so small print in a large photo loses detail.
- 200 tokens is enough for a sign, a label or a short menu section. For a dense page, raise Maximum New Tokens toward 2,048.
- Set the source language to the language printed in the image.
Limitations
- Isolated words give the model little context, and a title or name on its own can come back mistranslated.
- Google's image benchmark covers only English text translated into four languages, so other directions have no published image scores.
Safety & compliance
Use is governed by Google's Gemma Terms of Use, and the Gemma Prohibited Use Policy lists what the model may not be used for. Google ran its child-safety, content-safety and bias tests on image-to-text translation as well and reports minimal policy violations. Photos of documents can carry personal data such as names, addresses or ID numbers, so handle uploaded images under your own privacy rules.
API quick start
Run translate-gemma-4b-it-image with a single API call.
{
"inputImage": "https://your-cdn.com/input.png",
"sourceLng": "Chinese",
"targetLng": "English",
"maxNewTokens": 200
}curl -X POST "https://api.wiro.ai/v1/Run/google/translate-gemma-4b-it-image" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"inputImage": "https://your-cdn.com/input.png",
"sourceLng": "Chinese",
"targetLng": "English",
"maxNewTokens": 200
}
JSONOther models you might like
Explore similar wiro ai models.
