Azure Translator v3 PDF translator Issues

Hi,

Thank you for the assessment.

One practical question: how is this investigation initiated? Does this thread get forwarded to the product team, or should I open a formal Azure support request? If so, which support category should I select?

I have the source PDF and both outputs ready to attach wherever useful.

Thank you.

Hi,

Thank you for the guidance. I tested the same sample PDF using the 2026-03-01 batch API and I can share the results.

Test setup

• Endpoint: {endpoint}/translator/document/batches?api-version=2026-03-01
• Asynchronous batch translation, managed identity authorization
• Source: Italian (auto-detected), Target: English
• Same DWG-derived native PDF used in the previous samples
• No additional options passed (translateTextWithinImage not enabled, since I do not want image content translated)

Results

There is a clear improvement compared to the v1.1 output:

• Layout and page structure are preserved significantly better
• Text translation quality is good

However, the translated text still shows changes in font and colour compared to the source document. The layout preservation is much stronger than in v1.1, but the typography of the translated text is not retained.

And the core issue remains: the company logo is still being translated and re-rendered, even though it is an embedded image and not selectable text. The typography of the logo is lost in the output, exactly as it was with v1.1.

The job status response is worth noting:

"status": "Succeeded",
"summary": {
    "total": 1,
    "success": 1,
    "totalCharacterCharged": 2070,
    "totalImageCharged": 0
}

totalImageCharged is 0, which indicates no image translation was billed. Yet the logo was still re-rendered in the output. This suggests the logo is being processed as part of the standard PDF pipeline rather than through the image translation feature.

Question

Since translateTextWithinImage was not enabled and no image scans were charged, I expected embedded images to be left untouched, consistent with the FAQ statement that only the digital portions of a document are translated.

Could you clarify:

1. Why is the embedded logo still being re-rendered in the 2026-03-01 PDF path when image translation is neither enabled nor billed?
2. Is there any way to prevent specific embedded images from being processed, or is this inherent to the Document Intelligence-based PDF pipeline?
3. Is there a roadmap item for image or region preservation in PDF translation?

For context, this affects every technical drawing we translate, since each drawing carries the company logo in the title block. The layout improvements in 2026-03-01 are valuable, but the logo re-rendering remains a blocker for production use.

You can find the v1.1 output, and the 2026-03-01 output attached for comparison if that helps.

Thank you.

sample1_2.pdf

Hi, 

Thank you for the detailed response. Here are the requested details: 

Document Translation API version: v1.1 

I am using the Microsoft Translator V3 connector in Power Automate. The operation-location header returned by the service confirms the endpoint being called: 

https://...cognitiveservices.azure.com/translator/text/batch/v1.1/batches/{id}} 

Workflow: Asynchronous (batch). The flow calls StartDocumentTranslation, receives an operationID, then polls GetDocumentsStatus until the document status is Succeeded or Failed. 

Source and target languages: Source is set to Auto-detect, since our documents often contain more than one language. Target varies per request (en, de, fr, it and 12 others), selected by the user from a SharePoint metadata column. For the purpose of test I will provide all documents in italian translated to english. 

PDF content: Both. The document contains selectable text and embedded images. I verified this with Ctrl+F in the PDF: the text inside the title block tables is searchable, while the company logo is not confirming the logo is an image, not text. 

Sample files: You can see the documents attached to the reply. 

Additional question about the 2026-03-01 API version 

You mentioned that the 2026-03-01 Document Translation API supports PDF translation using Azure Document Intelligence with the goal of preserving PDF layout and structure. Two questions: 

1. Would this version be expected to resolve the image/logo re-rendering I am seeing, or does it only address text layout reflow? 

Is this version accessible through the Power Automate Microsoft Translator V3 connector, or does it require calling the REST API directly? 

Summary of the observed behavior 

I will share two sanitized sample sets, because they behave differently and the comparison may be useful: 

Sample 1: native PDF exported from DWG (CAD drawing) 

Never scanned 

Fonts changed throughout the document, including in areas that were not translated 

The company logo, which is an embedded image and not searchable text, was re-rendered with a substitute font. 

Sample 2: native PDF exported from WORD 

Also shows font substitution and text shifting 

However, the embedded images are fully preserved and untouched 

Sample 3: WORD version of Sample 2 

This version is to show that when it is not PDF, the translation is completed perfectly. 

 In Sample 2 the images behave exactly as the FAQ describes: digital portions translated (with a different font), images left alone. In Sample 1 the logo is clearly re-rendered despite being an image. This is why I suspect the processing path differs depending on PDF structure, possibly with OCR being applied to the DWG-derived file. 

LAST UPDATE: While I was trying to prepare a sample of the PDF derived from a .dwg file, I translated it multiple times and the problem of the logo seemed resolved, but i do not understand why. I will still provide that file as a sample no1, and I appreciate if you can explain the main reason why it got fixed, so I can use the same approach for translating documents like that.  

That being said, I still lose the font, and the information like headers, titles (the ones that are written with bold/italic) in native PDFs. I would appreciate it, If you can provide me an approach I can use for native PDFs that allow me to provide an optimal output.  

Thank you for your help.