Image Generation Workloads: Prompt, Edit, and the Model That Went Away

Every other workload in this topic consumes content. Image generation produces it, and that inversion changes almost everything about how the work is scoped, evaluated, and governed. This note covers what image generation actually offers beyond the obvious text-to-picture trick, how you tell a generation scenario from an analysis one, and a piece of naming that trips up anyone revising from older material.

First, the naming correction

If you learned this subject even a little while ago, you learned the name DALL-E. Retire it. The image generation model dall-e-3 was retired on 4 March 2026 and is no longer available for new deployments; existing deployments do not function. Its predecessor dall-e-2 was retired earlier still, with gpt-image-1 given as the suggested replacement.

Current image generation on Azure runs on the GPT-Image series of models from OpenAI, available through the Foundry Models catalogue. Naming a retired model as the current answer is the kind of error that marks out stale study material, so it is worth over-learning: the family name to reach for is GPT-Image, and different members of it trade resolution, speed, and cost against one another.

The workload is bigger than text-to-image

The mental image most people hold — type a sentence, receive a picture — describes only the simplest case. These models accept text and image inputs, and that second input is what makes them useful in production rather than merely fun.

Generation from a prompt alone is the familiar mode: concepts, illustrations, and placeholder assets conjured from a description. Editing takes an existing image and changes it according to instructions, which is a completely different business proposition because it operates on assets you already own. Inpainting refines that further by supplying a mask alongside the prompt, so that only the marked region is regenerated while the rest is preserved — swapping the background behind a product while the product itself stays untouched. Variations produce alternative renderings of an existing image, which is what turns one approved asset into a set.

Face preservation appears as a distinguishing capability across the series, and it is a good illustration of how these models are differentiated: some members target realistic, consistent depictions of people, while a smaller, cheaper member is explicitly better suited to non-portrait creative imagery.

Recognising a generation scenario

The tell is the direction of information flow. Analysis workloads take content that exists and produce facts about it. Generation workloads take a description and produce content that did not exist. If a scenario mentions marketing concepts, product mock-ups, storyboards, illustrations for articles, localised creative variants, or synthetic imagery for prototyping, you are in generation territory.

A subtler tell is how success is judged. Extraction has a right answer you can check against the source. Generation has no source to check against — output is assessed for fitness and appeal, which is judgement, not accuracy. That is why generation workloads almost always keep a human in the approval loop by design, and why they are rarely candidates for straight-through processing in the way document extraction is.

Operational shape and the guardrails

A few practical properties are worth carrying. Requests can produce several images at once, which supports the natural workflow of generating options and choosing among them rather than expecting one perfect result. Output arrives as encoded image data in the response rather than as a link to fetch later. Resolution and aspect-ratio support vary by model, and the newest members reach substantially higher resolutions than the earlier ones.

Governance deserves particular attention here because generation creates new material rather than describing existing material. Content filtering applies to both the prompt you send and the content that comes back, and the reason for the two-sided check is exactly that a benign-looking prompt can produce something unacceptable. Access to some models in the series is gated behind an application process, which is itself a signal about how the capability is regarded.

Mistakes to avoid

• Naming DALL-E as a current option. It is retired; say GPT-Image.
• Thinking only in text-to-image. Editing, inpainting, and variations are where most business value sits.
• Expecting deterministic output. The same prompt need not produce the same picture, so version and archive what you approve.
• Assuming filtering is only about the prompt. Both input and output are evaluated.
• Confusing generating an image with understanding one. They are opposite directions and use different tooling.

What to carry forward

Image generation is the producing member of the workload family. Ask whether the scenario needs something new or something explained; if it is new, reach for the GPT-Image series, expect editing and masking to matter more than pure prompting, and plan for human approval rather than automation.