The old version of this article used DALL-E as a catch-all name for image editing. That is now misleading. In Azure OpenAI, DALL-E 3 was retired on 4 March 2026 and Microsoft directs customers to the GPT Image family instead (Microsoft Learn).
The useful workflow remains simple: provide an image you are allowed to use, describe the change, generate a variation, and inspect the whole result before publishing it.
Before uploading an image
Check four things first:
- Rights: you own the image or have permission to modify it.
- People: you have an appropriate basis for processing any identifiable person's image.
- Confidentiality: the file contains no client data, internal diagrams or hidden metadata that your chosen service should not receive.
- Brand risk: somebody is responsible for approving the final result.
An AI edit can change more than the area you mention. Logos, hands, text, faces and small product details deserve a deliberate check.
A reliable editing prompt
Describe the fixed parts before the change:
Keep the subject, camera angle, background layout and lighting unchanged. Replace the blue notebook with a plain red notebook. Do not add text, logos, extra objects or people. Produce a landscape image suitable for a blog header.
This does not guarantee pixel-perfect preservation. It makes the acceptance criteria visible.
Use an interactive image editor
The exact interface varies by product, but the workflow is normally:
- Start an image-editing conversation or tool.
- Upload the source image.
- State what must remain unchanged.
- Describe one edit at a time.
- Generate a variation.
- Compare it with the original at full size.
- Save only the version that passes your review.
Making one material change per round gives you a clearer failure signal than asking for a complete redesign in one prompt.
Use Azure OpenAI for a governed application
Microsoft's current image-edit API accepts an input image plus a text prompt. It requires an Azure OpenAI resource, a deployed supported GPT Image model and a supported region. Access and model availability can vary, so confirm them in your own subscription.
The API uses multipart form data rather than a JSON-only request. Microsoft documents PNG and JPG inputs, content filtering, authentication failures, rate limits and model-specific options in its current image generation guide (Microsoft Learn).
For production, avoid embedding an API key in a client application. Use an approved secret store or managed identity where supported, restrict who can call the service, and log enough context to investigate failures without retaining sensitive images unnecessarily.
Review the result properly
Use a simple acceptance checklist:
- Is the requested object or area changed?
- Are protected details genuinely unchanged?
- Is all visible text correct?
- Are people represented accurately and appropriately?
- Does the output meet the required dimensions and file format?
- Can you document the source, prompt, model and human approver?
Do not treat a plausible preview as proof. Open the downloaded file, check its dimensions and inspect it at 100 percent zoom.
The proof boundary
This workflow can produce an editable candidate image. It cannot establish copyright ownership, consent, brand approval or factual accuracy. Those remain human and organisational decisions. Provider terms, model availability and billing can change, so verify them at the point of use.
For more practical AI experiments with the controls left visible, join Microsoft Copilot Adopters.