Microsoft Foundry: gpt-image-2.5-sunburst blocks simple photo edits at output moderation (other)

Idan Slonimsky 0 Reputation points
2026-11-20T09:51:55.37+00:00

I’m using gpt-image-2.5-sunburst in the Microsoft Foundry playground to edit a family photo. Simple edits, including adding a text caption or removing background bystanders without changing the foreground subjects, consistently fail with a safety block. I’m already using the least restrictive guardrail settings available to me, and the JPEG has no embedded EXIF metadata.

The request returns HTTP 400 with this error (request ID redacted):

{"error":{"message":"Your request was rejected by the safety system. If you believe this is an error, contact us at Azure support ticket and include the request ID [redacted].","type":"image_generation_user_error","param":null,"code":"moderation_blocked","moderation_details":{"moderation_stage":"output","categories":["other"]}}}
  1. Does Sunburst’s output moderation run independently of the guardrails configured on the Foundry deployment?
  2. What does the other category mean for image edits? Is there any way to get a more specific reason for the block?
  3. My subscription doesn’t let me open an Azure technical support request. Is there another way to report a suspected false positive and share the request ID privately, without posting the family photo?
Azure OpenAI in Foundry Models
0 comments No comments

1 answer

Sort by: Oldest
  1. AI answer

    2026-11-20T09:53:37.46+00:00

    Submit the suspected false positive through Filters Feedback in the Foundry Images playground, including the request ID and error details but not the family photo or other private information.

    Report the false positive without a support plan

    1. Reproduce the failure once in the Images playground.
    2. Select Filters Feedback at the top of the playground.
    3. Choose the applicable content-filtering issue.
    4. Include:
      • Model/deployment: gpt-image-2.5-sunburst
      • Operation: image edit
      • HTTP status: 400
      • Error code: moderation_blocked
      • Stage: output
      • Category: other
      • The complete request ID
      • Region and approximate UTC timestamp
      • A sanitized version of the edit instruction
      • That multiple benign edits fail and that the least restrictive available deployment filter is already selected
    5. Do not attach the family photo, names, faces, or other sensitive information.

    Microsoft documents Filters Feedback specifically for filtering issues in the Images, Chat, and Completions playgrounds and instructs users not to include private or sensitive information.^1^

    This is a feedback/reporting channel, not a technical-support case, so it might not provide an individual investigation or response.

    Output moderation versus deployment guardrails

    The error identifies where the rejection occurred:

    moderation_stage: output
    

    This means an image was produced internally and then rejected during output evaluation; Microsoft explicitly documents that a generated image itself can be filtered.^2^

    Sunburst also has built-in Responsible AI protections, while Azure applies input and output moderation and additional safeguards across image-generation models.^3^ Deployment guardrails configure the documented output filters, but selecting the least restrictive setting does not establish that every built-in safeguard has been disabled. There is no supported per-request bypass for this output rejection.

    Meaning of other

    Microsoft’s public image-edit schema documents detailed results for categories such as hate, sexual, violence, and self_harm, including severity and filtered status.^4^ It does not document a finer breakdown for the other value shown in your response.

    Consequently:

    • other should be treated as a nonspecific safety classification, not as evidence of a particular violation.
    • The API response contains no documented field that expands it into a more precise reason.
    • The request ID is the key correlation value for Microsoft’s internal investigation.
    • Absence of EXIF metadata does not narrow the category, and there is no reason to repeat that test.

    Conditional check: photorealistic images of minors

    If the family photo contains a minor, Microsoft documents a separate safeguard: photorealistic images of minors are blocked by default. Enterprise-tier customers are automatically approved; other customers can submit the documented request for access.^3^

    That access request is distinct from both Filters Feedback and an Azure technical-support ticket. Use it only if the source photo contains a minor and the subscription is not already covered by automatic enterprise-tier approval. The other response itself does not prove that this safeguard caused the block.

    If an individual investigation is required

    Azure technical support requires both:

    • A qualifying support plan.
    • Subscription-level Owner, Contributor, Support Request Contributor, or equivalent Microsoft.Support/* permissions.^5^

    If support should be available through your organization, ask the subscription administrator or owner of the organization’s Microsoft support agreement to verify entitlement and either grant the required role or open the case on your behalf.^6^ The case can contain the request ID and sanitized reproduction details first; do not post the photograph publicly.


    References

    1. Configure content filters
    2. Quickstart: Generate images with Azure OpenAI in Azure AI Foundry Models (rest-api)
    3. How to Use Image Generation Models from OpenAI - Microsoft Foundry | Microsoft Learn
    4. Azure OpenAI image and audio REST API reference (2025-04-01-preview)
    5. Create an Azure support request
    6. Microsoft Cloud Portals: Self-Help When a Support Request Cannot Be Opened
    AI-generated content may be incorrect. Read our transparency notes for more information.

    Was this answer helpful?

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.