# Where does an image model fit in an internal app?

> Put the image step on the approval, inspection or quality screen your team already opens. A named person accepts a version before the status moves.

- URL: https://www.rebels.software/guide/image-models-in-internal-apps
- Published: 2026-11-10
- Author: Bart Szczepański
- Tags: Applications, Image generation, Quality checks

Put an image step on the internal record a person already reviews: an approval, an inspection or a quality check. A vision-capable model may draft a note from the photo, and a GPT Image model may produce or edit the picture stored beside it. The record moves only when a named person accepts a version, and the app logs the call and that decision.

## Where the step sits

Use the admin panel or the operational dashboard your team already opens. A goods-in line, a quality hold, a document approval or a field photo on a ticket is a record with a status. The image step sits between the attached photo and the person's decision, in the status flow that screen already uses.

Show four things at once: the source photo, the model output, the decision buttons, and who last ran the call. The attached file stays the source. The model result sits beside it, still in review, until a person accepts, sends the record back or rejects it.

## What the model may check, and what it may produce

Ask for one job per call. The [images and vision guide](https://developers.openai.com/api/docs/guides/images-vision) splits the work by what you want back.

| | A check | A picture |
| --- | --- | --- |
| Result | Text: a draft note | An image: a new version or an edit |
| Model | A vision-capable language model | A GPT Image model |
| API | Responses or Chat Completions | Images API |
| On the screen | Under the source photo | Beside the source photo |

A check comes back as text the reviewer can edit. Vision models can describe an image, read visible text, and answer questions about objects, shapes, colours or textures. OpenAI says to [account for the model's limitations](https://developers.openai.com/api/docs/guides/images-vision) when you use those answers. On an inspection photo, expect incorrect descriptions, approximate counts, trouble with small text, and trouble with rotated or upside-down text. Specialised medical images are called out as unsuitable. For small print, such as a label, the guide recommends a higher `detail` setting, including `"detail": "original"` when the model supports it. The note can pre-fill the form. The hold stays open until a person closes it.

A picture is pixels. GPT Image models generate and edit images from text and image inputs, and inside an internal app that is usually an edit of the source photo. The [image generation guide](https://developers.openai.com/api/docs/guides/image-generation) says to review the result before using it, and lists limits that matter here: complex prompts may take up to 2 minutes, text in the image can be unclear, and precise placement can slip. Plan the screen for a slow call, and keep any pass or fail in the record fields.

GPT Image 2.6 Sunburst and Flare do this through the Image API only. The [Sunburst](https://developers.openai.com/api/docs/models/gpt-image-2.6-sunburst) and [Flare](https://developers.openai.com/api/docs/models/gpt-image-2.6-flare) pages list `v1/images/generations` and `v1/images/edits` as supported, and the Responses API as not supported. Which model to call, and the parameters, is in [how to use GPT Image 2.6 in business apps](/guide/gpt-image-2-6-in-business-apps) and in [what GPT Image 2.6 changes for business apps](/blog/gpt-image-2-6-what-changes-for-business-apps).

When a reviewer must point at a region, use markup. They draw on a separate transparent PNG, the same width and height as the source after orientation, and the app sends that layer as `markup` on the edit. Markup applies to the first input image, it cannot be combined with `mask`, and the marks stay off the photo. The prompt says what must change and what must stay.

## Step by step

1. **Name the screen, the photo and the decision.** One sentence: which record, which photo, which status changes on accept. Example: a goods-in receipt stays on hold until a reviewer accepts the pallet-label photo.
2. **Pick check, picture, or both as separate calls.** A draft note about the label is a vision call. A clearer crop of a circled region is an image edit. When you need both, show both results on the record.
3. **Keep the source photo unchanged.** Later runs add versions. They never replace the file from the warehouse, the inspector or the requester. The screen shows source and candidate side by side.
4. **Put the person on the same screen.** Accept, send back and reject are buttons on the record. Anything other than a plain accept needs a short note. The model has no button that moves the status.
5. **Treat a blocked call as a state.** The image generation guide says not to retry an `image_generation_user_error` unchanged, including `moderation_blocked`. Leave the record in review and show a plain message.
6. **Store the version the person accepted.** Later steps may use that file, with the reviewer's name and the time. A further edit is a new version waiting for the same decision.

The reviewer compares each result with the source photo, then accepts a version. Measure accuracy on your own records.

## What to log

Log enough to reconstruct the decision later.

- Record id, photo id and a hash of the source file.
- Model id, and whether the call was vision analysis, image generation or an image edit.
- The instruction you sent. For an edit, note the prompt and whether markup or a mask was attached. Keep names and raw email addresses out of the prompt.
- A `safety_identifier` when the app acts for one user: a hash of the username or email, 128 characters or fewer, not the address itself. The image generation guide recommends this on generation and edits.
- The response `usage` object, so token use per accepted record stays on the record.
- Request id and, on failure, `error.code`. For `moderation_blocked`, also store `moderation_stage` and `categories` when they are present. The reviewer sees a short message. Your team sees the details.
- The decision: who, when, which version, and the note.

A photo of a person is personal data under the [GDPR](https://eur-lex.europa.eu/eli/reg/2016/679/oj). Send it to an external image API only when that processing is allowed and the file may leave your network. The same boundary for documents is in [how to prepare company documents for private AI search](/guide/prepare-documents-for-private-ai-search).

## Common mistakes

- **Letting the draft close the record.** A person still accepts the result before the status moves.
- **Overwriting the source photo.** When the edit is wrong, the original still has to be there.
- **Painting the decision onto the picture.** Text in the image is a known weak spot. A pass or fail belongs in fields you can search.
- **Flattening the reviewer's marks onto the photo.** Markup is its own transparent PNG at the source size.
- **Retrying a blocked request unchanged.** `moderation_blocked` needs a different prompt or a different input.
- **Logging the picture alone.** Keep the model id, the request id, `usage` and the person's decision with it.

## Next step

Pick one screen you already have, one photo type and one decision. Write the statuses, the calls and the log fields above. Run them on a small set of real records, with a person accepting every one, before a status can move on its own.

If you want that step on a screen your team already uses, [we can walk through one workflow with you](https://calendly.com/bartlomiej-jacyno/intro?utm_source=blog&utm_medium=text_cta&utm_campaign=image-models-in-internal-apps).

We start with your biggest problem and the data behind it. [Book a free call](https://calendly.com/bartlomiej-jacyno/intro)
