Back

How-to5 min read

How to review AI-generated images before they go live

Guide card: How to review AI-generated images before they go live

Keep each generated image on the admin record until a named person approves it, rejects it or sends it back for another version. Save the prompt, the model id, the version and that person’s name on the file. The customer page uses the version approved with an empty note.

Where the image call sits on a screen you already open is in Where does an image model fit in an internal app?. Which model to call, and the parameters, is in How to use GPT Image 2.6 in business apps.

Who reviews which image

Write down each image type before the first file is generated. Name the person who may approve it, and a second person when the picture can leave the company.

Image Who approves Second person
Internal diagram, no text, seen only inside the company The person who requested it The requester, alone
Product photo on a page, in a catalogue or in an email The owner of that page or campaign Required if the picture contains text, a price, a logo or a face
A real person’s photo, a trademark you did not supply, or a claim (a price, a date, a legal line) The page owner Someone who can confirm the claim

The reviewer field holds one person’s name. When a customer will see the image, the page owner approves it.

Step by step

  1. Name the type and the two people. Use one row of the table for the first image type you turn on. A packshot of a product you already photograph is a sound start: the product owner requests the file, and the page owner approves it.

  2. Put the three decisions on the record. Approve, request changes, and reject. An approval with an empty note is what lets the file move on to publishing. The other two need a reason from the list below, and either can carry a short note. OpenAI’s image generation guide says to review a result before you use it.

  3. Keep one reason list, so a week of decisions can be counted.

    • wrong product or wrong variant
    • extra object
    • wrong colour or material
    • likeness off
    • unreadable text, or text you did not ask for
    • a logo or trademark you did not supply
  4. Store the generation on the version. Each file is a version. A new run adds one and leaves the earlier files in place.

{
  "version": 3,
  "parent_version": 2,
  "prompt": "Packshot of SKU 1842 on a white background. Leave text and logos out of the frame.",
  "model": "gpt-image-2.6-flare",
  "size": "1024x1024",
  "quality": "medium",
  "file_hash": "sha256 of this file",
  "requested_by": "product owner",
  "decision": "changes_requested",
  "reason": "extra_object",
  "note": "Remove the second cup.",
  "reviewer": "page owner",
  "decided_at": "2026-11-13T09:40:00Z"
}

On an edit, store the source-file hashes and the request id. Unclear text is one of the limits in the image generation guide, which is why the exact prompt stays on the version.

  1. Show the next prompt before the model runs. On request changes, take the previous prompt and add one line from the reason, such as “keep SKU 1842, remove the second cup, leave text out of the frame”. The reviewer edits that text, then the call goes out. The new file is the next version. The previous version stays, and later decisions are appended under it.

  2. Count the reason codes once a week. When one code keeps coming back, change the template and stop pasting that fix into every free-text prompt. When unreadable text is the code you see most, stop asking the model to draw the label and the price, and put both in the page template beside the image.

When a file may leave the panel

The page uses the version whose latest decision is an approval with an empty note. These stay on the record until a person has dealt with them:

  • No decision yet, or the decision is changes_requested or rejected.
  • A failed call, including moderation_blocked. The image generation guide says not to retry that error unchanged. Leave the record in review and show a short message.
  • A picture of a real person. That photo is personal data under the GDPR. Putting it on a page is the same decision as publishing any other photo of that person.
  • Text the model drew, such as a price, a date, a product name or a legal line. Someone reads those words on the image first.
  • A logo or trademark outside the reference images you are allowed to use.
  • An approval with a note on it. The note means the accept was qualified. Close it with a new version, or keep the file inside the company.
  • The first batch on a new model id or a new prompt template. The earlier approvals covered the previous pair.

Automatic publishing, for one image type, runs on a fixed template. Structured fields are what change, such as the SKU and the background colour, on one model id a person has already approved against real records. Each week someone looks at a sample. A free-text prompt, and any new model id, keeps a person on every file.

One packshot, two versions

Picture a catalogue team building a packshot from two warehouse photos of one product. The product owner requests it. The page owner reviews it.

Version 1 comes back with a second cup in the frame and a price drawn on the label. The page owner requests changes, reason extra_object, note “Remove the second cup”, and takes the price out of the prompt because it already sits in the page template. Version 2 comes back with one product and no drawn price. The page owner approves it with an empty note, and the product page uses version 2. Version 1 stays, with the reason.

Common mistakes

  • The person who requested the image also approves it. They already know the prompt, so a second cup is easy to miss. When a customer will see the picture, the page owner is the name on the decision.
  • The new file replaces the old one. Version 1 remains, with the reason that led to the retry.
  • Reject is a blank box. The reason code is what you count at the end of the week.
  • The next prompt runs unseen. The reviewer reads the added line before the model is called.
  • A timer publishes the newest file. Point the job at the version approved with an empty note.
  • A model upgrade inherits the old approvals. A new model id goes back through a person on every image of that type.

Next step

Pick one image type you already generate in the admin panel. Write the reviewer, the reason list and the fields on the version. For one week, a person sees every file. Read the reason codes before any file of that type publishes on its own.

If you want that queue on a screen your team already uses, we can walk through one image type with you.

  • Applications
  • Image generation
  • Review