6.5 DALL-E Image Generation

Quick Answer: Azure OpenAI creates images via the Images API with DALL-E 3 (and the newer GPT-image-1 family). DALL-E 3 supports exactly 1024x1024, 1792x1024, 1024x1792, n=1 per call, quality standard/hd, style vivid/natural. DALL-E 3 auto-revises the prompt and returns revised_prompt. Content filters screen both the prompt and the output image.

Generating an Image

from openai import AzureOpenAI

client = AzureOpenAI(api_version="2024-10-21",
                     azure_endpoint="https://my-openai.openai.azure.com/",
                     azure_ad_token_provider=token_provider)

resp = client.images.generate(
    model="dalle3",                 # deployment name
    prompt="A serene mountain lake at sunset, digital art",
    n=1,                            # DALL-E 3 supports only 1
    size="1024x1024",               # or 1792x1024 / 1024x1792
    quality="hd",                   # or standard
    style="vivid"                   # or natural
)
image_url = resp.data[0].url                  # valid ~24 hours
revised = resp.data[0].revised_prompt          # what DALL-E actually used

On the Exam: The returned url is temporary (expires roughly 24 hours) — download and persist the image (e.g. to Blob Storage) if you need it long term. You can also request response_format="b64_json" to get the bytes inline instead of a URL.

DALL-E 3 Parameters

Trap: Options offering 256x256 or 512x512 are testing whether you remember those are DALL-E 2 sizes, removed in DALL-E 3. Likewise, asking for n=4 in one DALL-E 3 call fails.

Prompt Revision

DALL-E 3 rewrites your prompt before generating, adding descriptive detail, artistic direction for vague requests, and safety-oriented phrasing. The revised_prompt field returns the exact text used, so you can log it, display it, or detect when the model substantially reinterpreted your intent. This behavior is unique to DALL-E 3 and is a favorite exam fact.

Dual Content Filtering

Unlike text generation (filtered once), image generation is screened 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, and you can receive a content_filter finish reason on either side.

Responsible-AI Restrictions

• No photorealistic images of identifiable real people (celebrities, public figures).
• No reproduction of copyrighted characters or trademarked logos.
• Refuses prompts for misleading, deceptive, or otherwise harmful imagery.
• A blocked request returns an error / content_filter flag rather than an image.

Scenario: an app sends "a photorealistic portrait of [a named celebrity]". The prompt filter rejects it under the real-person rule — the fix is to describe a fictional or non-identifiable subject, not to raise any parameter.

Reading the finish reason

When a request is blocked, you do not get a usable image. The Images API surfaces this either as an HTTP error carrying a content-filter code or, in chat-style flows, as a content_filter finish reason. Your application must handle this gracefully — show a friendly message and let the user revise the prompt rather than silently failing. This dual-stage screening means defensive code should anticipate a block at either the prompt or the image stage, not just on submission.

Persisting generated images

Because the returned url expires (~24 hours), production apps follow a consistent pattern: generate the image, immediately download the bytes (or request response_format="b64_json"), and store them in durable storage such as Azure Blob Storage with your own access controls. Relying on the temporary URL for anything beyond an immediate preview is a frequently-tested mistake. If you need several variations, loop and call the API multiple times with n=1 each, since DALL-E 3 never returns more than one image per request.

DALL-E 3 vs. GPT-image-1

Quality vs. style, decoded

The two creative knobs are independent and often confused. quality (standard vs. hd) controls how much rendering effort and detail go into the image — hd costs more and takes longer but sharpens fine detail. style (vivid vs. natural) controls the aesthetic: vivid pushes dramatic, saturated, hyper-real results, while natural yields more muted, photographic output. A scenario asking for "realistic, true-to-life product photos" points to style=natural; one asking for "eye-catching, dramatic marketing art" points to style=vivid. Neither knob changes the allowed sizes or the n=1 limit.

Responsible-AI summary for images

These are enforced by the service, not optional toggles a developer can disable, and they are the most likely responsible-AI image questions on the exam.

On the Exam: Remember the dual filter (prompt AND image), the fixed three sizes with n=1, the revised_prompt field, the temporary URL that must be persisted, and the real-person / copyright restrictions. If a scenario requires editing an existing image with a mask, that points to GPT-image-1, not DALL-E 3, which only generates from text.