.Net: Bug: AzureOpenAITextToImageService: gpt-image-1 support gaps — wrong model/deployment handling and missing execution settings (quality, output_format)

[AdriaanLarcai]

When using gpt-image-1 through AzureOpenAITextToImageService, two categories of problems surface:

1. Model vs deployment handling

• Observed: Using an Azure OpenAI deployment name other than “gpt-image-1” triggers HTTP 400 from the image API, consistent with the request body passing the deployment name as the model-id.
• Expected: For OpenAI’s non-Azure endpoint, model should be the literal “gpt-image-1”. For Azure OpenAI, the SDK should target the Azure deployment without passing the deployment name as model-id to the OpenAI-style JSON payload. In other words, avoid sending:

  {
    "model": "my-gpt-image-deployment",
    "prompt": "..."
  }

  and instead either:
  • OpenAI (non-Azure): "model": "gpt-image-1"
  • Azure: use the Azure.AI.OpenAI client with the correct deployment routing, not a model-id string in the JSON payload.

2. Missing gpt-image-1 execution settings

• OpenAITextToImageExecutionSettings lacks:
  • quality: “low” | “medium” | “high”
  • output_format: “png” | “jpeg” | “webp”
• Expected: gpt-image-1 supports these parameters; SK execution settings should expose them and map them to the underlying request.

Repro steps

• Configure SK with AzureOpenAITextToImageService and a valid Azure OpenAI deployment for gpt-image-1 (e.g., “my-gpt-image-deployment”).
• Call GetImageContentsAsync and inspect the outbound payload or capture the exception.
• Actual: HTTP 400 invalid_request due to model-id mismatch when the deployment name is used as the model in the payload.
• Attempt to set execution settings for quality=“medium” and output_format=“jpeg”; these aren’t supported by OpenAITextToImageExecutionSettings in current SK releases.

Actual result

• 400 error when using non-default deployment names, due to the deployment name being passed as a model-id in the request.
• No way to specify gpt-image-1 parameters like quality and output_format via SK execution settings.

Expected result

• Azure connector uses Azure.AI.OpenAI v2 client routing via deployment, not a model-id field in the body.
• Execution settings include gpt-image-1 parameters and are correctly mapped:
  • quality: low | medium | high
  • output_format: png | jpeg | webp
• Requests succeed for any valid Azure deployment name bound to gpt-image-1.
• OpenAI (non-Azure) connector sends "model": "gpt-image-1".

Environment

• Semantic Kernel: 1.66.0
• Connector: Microsoft.SemanticKernel.Connectors.AzureOpenAI.AzureOpenAITextToImageService
• Runtime: .NET
• Azure OpenAI: gpt-image-1 deployment
• OS/IDE: Windows 11 / Visual Studio 2022

Impact

• Blocks proper use of Azure OpenAI deployments for gpt-image-1 when deployment names differ from the literal model string.
• Prevents callers from using supported quality/output_format options, limiting functionality and cost/perf tuning.

Proposed changes

• Model/deployment handling

  • Azure: Use Azure.AI.OpenAI OpenAIClient.GetImageGenerationsAsync (or current equivalent) with the deployment routing, avoiding putting deployment name into “model” in the JSON payload.
  • OpenAI (non-Azure): Set "model": "gpt-image-1" explicitly.

• Execution settings additions

  • Extend OpenAITextToImageExecutionSettings to include:

    public sealed class OpenAITextToImageExecutionSettings
    {
        // New
        public string? OutputFormat { get; set; }     // "png" | "jpeg" | "webp"
        
        // Fix  GetGeneratedImageQuality to support "low", "medium" for gpt-image-1
    }

  • Map these properties to the underlying request body for gpt-image-1:

    {
      "model": "gpt-image-1",
      "prompt": "...",
      "quality": "medium",
      "output_format": "jpeg",
      "size": "1024x1024"
    }

  • Maintain backwards compatibility: default quality to “high” and output_format to “png” if not specified.

• Tests

  • Azure: Valid deployment with non-“gpt-image-1” name should succeed; no “model” mismatch in payload.
  • OpenAI: Requests include "model": "gpt-image-1".
  • Execution settings: When set, the service produces images with requested quality and format; serialization verified.
  • Negative tests for invalid values produce clear, actionable errors.

Acceptance criteria

• Azure deployments work with arbitrary deployment names mapped to gpt-image-1 without 400/404 caused by model-id misuse.
• OpenAITextToImageExecutionSettings exposes quality/output_format and these are correctly honored in requests.
• Docs/samples briefly mention new properties and Azure vs OpenAI model/deployment behavior.
• Unit/integration tests cover both Azure and OpenAI paths.

References

• OpenAI docs: Image generation parameters (quality, formats) Image generation - OpenAI API
• Prior SK Azure image issue referencing AzureOpenAITextToImageService: .Net: The API deployment for this resource does not exist
• Microsoft Learn image options overview (quality modes) How to Use Image Generation Models from OpenAI - Azure OpenAI

[markwallace-microsoft]

@AdriaanLarcai thanks for creating this issue. Are you interested in working on a fix?