6.5 Image Generation (gpt-image; DALL-E 3 Retired)

Retired: The AI-102 outline still says "Use the DALL-E model to generate images", but Microsoft retired dall-e-3 in Azure OpenAI on March 4, 2026 (existing deployments stop working) and dall-e-2 on January 27, 2025. Current image generation uses the gpt-image series.

Quick Answer: Azure OpenAI creates images via the Images API with a gpt-image model. GPT-Image-1 supports 1024x1024, 1024x1536, 1536x1024, quality low/medium/high (default high), and 1-10 images per call with n. Output is base64 only (b64_json, no URL). gpt-image models also edit images with a mask. Content filters screen both the prompt and the output image.

Generating an Image

import base64
from openai import AzureOpenAI

client = AzureOpenAI(api_version="2025-04-01-preview",
                     azure_endpoint="https://my-openai.openai.azure.com/",
                     azure_ad_token_provider=token_provider)

resp = client.images.generate(
    model="gpt-image-1",            # deployment name
    prompt="A serene mountain lake at sunset, digital art",
    n=1,                            # 1-10 images per request
    size="1024x1024",               # or 1024x1536 / 1536x1024
    quality="high"                  # low, medium, or high (default high)
)
image_b64 = resp.data[0].b64_json   # base64 only; no URL option
with open("lake.png", "wb") as f:
    f.write(base64.b64decode(image_b64))

On the Exam: GPT-Image-1 (like 1-Mini, 1.5 and 2) returns base64 image data, so your code decodes and stores the bytes (for example in Blob Storage); there is no temporary URL to persist.

gpt-image Models at a Glance

GPT-Image-2.5 models are also generally available. Trap: 256x256 and 512x512 were DALL-E 2 sizes, and 1792x1024 / 1024x1792 were DALL-E 3 sizes; both models are retired. The one-image-per-request limit was also a DALL-E 3 rule, while gpt-image models accept n from 1 to 10.

Editing with a Mask

gpt-image models accept text and image inputs. The edits endpoint takes an input image, a prompt, and an optional mask that marks the region to change; the mask must be the same size as the input image. This inpainting is the answer whenever a scenario edits part of an existing photo.

Dual Content Filtering

The Image APIs come with a content moderation filter that is applied twice:

Both filters use the same four harm categories as text — hate, sexual, violence, self-harm — at configurable severity. A prompt may pass yet still have its output blocked.

Handling a blocked request

When a request is blocked, you do not get a usable image. Your application must handle this gracefully — show a friendly message and let the user revise the prompt rather than silently failing. Defensive code should anticipate a block at either the prompt or the image stage, not just on submission.

Persisting generated images

Because GPT-Image-1, 1-Mini, 1.5 and 2 return base64 data, production apps decode the bytes and store them in durable storage such as Azure Blob Storage with their own access controls. To get several variations, set n (up to 10) instead of looping.

Historical: DALL-E 3

Older study material and the AI-102 outline describe DALL-E 3: sizes 1024x1024, 1792x1024 and 1024x1792, one image per request, quality standard/hd, style vivid/natural, a temporary URL, and an automatic revised_prompt. Those parameters belong to a model Microsoft retired on March 4, 2026, so don't treat them as current answers.

On the Exam: Remember the dual filter (prompt AND image), the gpt-image sizes and n range, base64-only output, and mask-based editing. If an item still offers DALL-E 3, treat it as the retired model; the current Azure OpenAI image models are the gpt-image series.