Python: New Feature: gpt-image-1 support

[ymuichiro]

name: Python: gpt-image-1 support
about:

AzureTextToImage does not support gpt-image-1. The AzureTextToImage implementation in Semantic Kernel does not support the gpt-image-1 deployment. It currently assumes a DALL·E 3-compatible interface that returns a URL, which gpt-image-1 does not provide.

Below is a minimal code example:

async def image_sample() -> None:
    kernel = sk.Kernel()
    service_id: str = "image"
    kernel.add_service(
        AzureTextToImage(
            service_id=service_id,
            deployment_name="gpt-image-1",
            endpoint=AZURE_OPENAI_IMAGE_ENDPOINT,
            api_key=AZURE_OPENAI_IMAGE_API_KEY,
        )
    )

    service: AzureTextToImage = kernel.get_service(service_id=service_id)
    settings = service.get_prompt_execution_settings_class()(service_id=service.service_id)

    if isinstance(settings, OpenAITextToImageExecutionSettings):
        settings.prompt = "sky"
        settings.size = ImageSize(width=1024, height=1024)
        settings.quality = "low"

    if isinstance(service, AzureTextToImage):
        r = await service.generate_image(settings=settings)

        print(r)


if __name__ == "__main__":
    asyncio.run(image_sample())

Problems Identified

1. Outdated openai-python versionThe SDK version is outdated and needs to be updated for gpt-image-1 compatibility. See: openai/openai-python@cc2c1fc
2. Assumption of URL-based responses. The current implementation assumes a response format that includes an image url, which is not the case for gpt-image-1. See:

   semantic-kernel/python/semantic_kernel/connectors/ai/open_ai/services/open_ai_text_to_image_base.py

   Line 69 in 8d1b3fd

   raise ServiceResponseException("Failed to generate image.")

Suggested Fix

• Add response handling logic that distinguishes between url (DALL·E 3) and b64_json (gpt-image-1).
• Consider explicitly supporting gpt-image-1 via a new class or branching logic within AzureTextToImage.
• (OPTIONAL) Update openai-python dependency to the latest version.

I have confirmed that even without updating the openai-python SDK, image generation with gpt-image-1 works correctly as long as you do not need to specify the latest properties#. Therefore, the only essential modification required is to handle the response format appropriately#.

---

[moonbox3]

@ymuichiro thanks for filing, we'll have a look. You can always upgrade the openai package on your own with pip install openai --upgrade.