Document Translation image translation: translated text drawn too small and grey. Any way to control font size or contrast?

Question

Document Translation image translation: translated text drawn too small and grey. Any way to control font size or contrast?

BLACHERE Pierre 0

We use Azure Translator Document Translation (batch API, api-version=2025-12-01-preview) to translate the text inside images taken from Word and PowerPoint files. We submit each image as a

standalone file (PNG, or JPEG at quality 95) in a batch job, sometimes with a TSV glossary.

The translation itself is fine, but our users complain about how the translated text looks in the output images:

Text too small. The translated text is often much smaller than the original. It seems to be shrunk to fit the original text box, especially when the translation is longer than the source (e.g. English → French or German).
Text too faint. Text that was solid black or a strong colour in the original often comes out grey or faded, which makes it hard to read on light backgrounds.

We couldn't find any request parameters in the API reference that affect how the translated text is drawn.

Questions:

Is there a way to set a minimum font size or text colour, or to let long text wrap or overflow instead of shrinking?
How does the service choose the colour of the translated text? Could anti-aliasing or a low-resolution source image make it come out grey?
Is there a recommended minimum image resolution or text height for readable results? Would scaling up small images before submitting them help?

Has anyone else seen this, or found a workaround (pre-processing, post-processing, a different API option)? Thanks!

Sai Kiran Mudavath 345 Reputation points Microsoft External Staff Moderator

2026-11-05T03:39:52.73+00:00

Hi @BLACHERE Pierre

Welcome to the Microsoft Q&A.

We are currently investigating the issue you reported. We kindly request that you allow us some additional time to complete our investigation.

We will provide you with an update as soon as we have further information.

Thank you for your continued patience and understanding.
Hari Venkat Miriyala 0 Reputation points Microsoft External Staff

2026-11-05T04:08:19.67+00:00
Hello @BLACHERE Pierre

Thank you for sharing the detailed observations and examples. I completely understand your concern, especially since font size, colour, and readability are important when translated images are included in Word and PowerPoint documents.

Based on the current Azure AI Translator Document Translation documentation, here are the key points:

Font size, wrapping, and text colour

For the 2025-12-01-preview batch API, there are no documented request parameters to specify a minimum font size, translated text colour, wrapping, overflow, or rendering resolution.

The image translation capability is designed to translate text within the image while maintaining the original design and layout. As translated text can be longer than the source text, the resulting appearance may vary while the service attempts to preserve that layout.

Faded or grey translated text

The specific internal logic used to select the rendered text colour and anti-aliasing behavior is not documented as a configurable API option. Therefore, I would not recommend attributing the grey appearance specifically to anti-aliasing, JPEG compression, or source resolution without examining a reproducible sample.

Image resolution

Microsoft does not currently document a minimum DPI, image resolution, or text-height requirement for this image translation capability.

As a practical test, I would recommend using a high-quality source image and, where possible, testing a lossless PNG version. You can also compare the results using a higher-resolution source image. This may help determine whether source-image quality is contributing to readability, but it does not guarantee a particular output font size or colour.

Workaround when precise formatting is required

If your application requires strict control over typography, such as minimum font size, exact colour, wrapping, or placement, a practical approach is to manage the final text rendering separately after translation. This allows the application to control these formatting properties directly.

For further investigation, please retain a reproducible source image, the corresponding translated output image, source/target languages, the batch operation ID, and the request timestamp. These details will help us investigate any consistent rendering differences more effectively.

Microsoft documentation

Azure AI Translator Document Translation overview

The documentation also confirms support for translating image formats such as .jpeg, .png, .bmp, and .webp, as well as text within images in .docx and .pptx files through batch document translation.

I hope this helps clarify the current behavior and provides a practical direction for improving the readability of the translated images. We certainly appreciate you bringing these observations to our attention.
Hari Venkat Miriyala 0 Reputation points Microsoft External Staff

2026-11-05T10:11:30.42+00:00

Hi Pierre,

Greetings!

Once you’ve reviewed the details above, could you please check and let us know the update?

1 answer

Your answer

Sai Kiran Mudavath 345 Reputation points Microsoft External Staff Moderator

2026-11-05T03:39:52.73+00:00

Hi @BLACHERE Pierre

Welcome to the Microsoft Q&A.

We are currently investigating the issue you reported. We kindly request that you allow us some additional time to complete our investigation.

We will provide you with an update as soon as we have further information.

Thank you for your continued patience and understanding.
Hari Venkat Miriyala 0 Reputation points Microsoft External Staff

2026-11-05T10:11:30.42+00:00

Hi Pierre,

Greetings!

Once you’ve reviewed the details above, could you please check and let us know the update?

Answer 1

Azure Document Translation does not expose controls for minimum font size, wrapping, overflow, text color, weight, contrast, or output rendering resolution, so applications that require predictable typography must extract, translate, and render the text themselves.

Current image-rendering behavior

For 2025-12-01-preview, image translation uses the normal batch request. Its documented request fields cover source and target locations, languages, category, glossaries, and storage settings—there are no typography or image-rendering properties ^1^.

Consequently:

Font size: There is no parameter to set a minimum, permit overflow, or select wrapping behavior.
Color and contrast: The service does not document how it selects the rendered color, nor an option to override it.
Long translations: Microsoft documents that changes in translated-text length can affect layout and cause reflow or shifting. Font substitution can also occur when the source font does not support the target language ^2^.
Grey output: Anti-aliasing, source resolution, compression, or service-side reconstruction could potentially affect appearance, but the published behavior does not identify which one causes grey text. The output alone is insufficient to attribute a specific cause.

A TSV glossary controls terminology, not visual rendering.

Recommended production workaround

Use a controlled OCR → translation → rendering pipeline:

Run Azure Vision Read OCR on the original image to obtain text and its location.
Translate the extracted strings with Azure Translator.
Remove or mask each original text region.
Draw the translated text yourself, implementing:
- minimum font size;
- explicit foreground color and contrast;
- target-language fonts;
- line wrapping;
- box expansion or overflow;
- high-resolution antialiasing.
Add review or fallback handling where text cannot fit without obscuring important image content.

Microsoft explicitly describes sequencing Computer Vision text extraction with Translator for image-text scenarios ^3^.

A typical fitting policy is:

font_size = original_font_size
lines = wrap(translated_text, box_width, font_size)

while text_height(lines, font_size) > box_height:
    if font_size <= minimum_font_size:
        box_height = expand_box_or_enable_overflow()
        break
    font_size -= 1
    lines = wrap(translated_text, box_width, font_size)

draw_text(lines, color="#000000", font_size=font_size)

The OCR coordinates do not necessarily preserve the original font family, size, or exact background. Those properties may need to be estimated from pixel measurements or supplied from the originating Word/PowerPoint objects.

Resolution and preprocessing

Document Translation does not publish a minimum text height or recommended DPI specifically for standalone image translation. Do not treat an undocumented threshold as a service guarantee.

Azure Vision Read OCR provides a useful benchmark for a custom pipeline: images must be at least 50 × 50 pixels, and its documented minimum text height is 12 pixels for a 1024 × 768 image—approximately 8-point text at 150 DPI ^4^. OCR accuracy can also be affected by resolution, contrast, lighting, rotation, text size, color, and density ^5^.

Prefer these preprocessing changes:

Export directly from Word or PowerPoint at a higher native resolution.
Prefer lossless PNG for screenshots, diagrams, and flat-color text. JPEG is lossy, while PNG uses lossless compression ^1^.
Increase text pixel height at export rather than enlarging an already-small raster image.
Preserve strong foreground/background contrast.
Avoid sharpening or thresholding that damages thin character strokes.

Upscaling may improve OCR when the original text is genuinely too small, but interpolation cannot recreate missing detail and there is no documented evidence that it changes Document Translation’s chosen output color or font size.

Test the GA API separately

Image translation became GA in Document Translation 2026-03-01, including standalone images and images embedded in Word and PowerPoint ^6^. Test representative files against that version before retaining the preview API, because rendering may differ. The GA schema still does not document typography controls.

If the source .docx or .pptx files remain available, another test is submitting those originals with translateTextWithinImage: true rather than exporting each image first ^7^. This may preserve the surrounding Office layout better, but it does not provide font-size or color controls.

If GA produces the same unreadable output, retain the smallest reproducible source/output pair and open an Azure support request as a rendering-quality issue. Include the API version, region, source and target languages, dimensions, format, text pixel height, and whether PNG and high-resolution export produce the same result.

References