When an AI photo edit changes the wrong object, a longer prompt is not always the best fix. Sometimes the missing information is location: which handle, which reflection, which patch of background? A visual annotation can make that editing brief easier to understand.
This week, OpenAI's image-editing documentation drew attention for a new markup input that names GPT Image 2.5 and 2.6. The useful takeaway is a concrete input format for visually guided edits—not proof that a new model has launched. Here is how to prepare a clear markup brief, avoid common file mistakes, and decide whether the resulting edit is actually usable.
What the documentation confirms
As checked on October 31, 2026, the OpenAI image-edit API reference describes markup as a transparent PNG overlaid on the first input image. It must match that image's dimensions after orientation, contain transparent pixels, and cannot be combined with a mask in the same request. The description explicitly says it supports GPT Image 2.5 and 2.6.
EvoLink's October 29 reporting also documents the parameter and the model-name reference. That report distinguishes a documentation appearance from a public release. The model choices shown in the API reference we checked do not establish a public 2.6 model ID. Do not guess an ID, assume account access, or promise a release date based on the markup description.
This guide is an editorial workflow, not a hands-on benchmark of GPT Image 2.6. It also does not imply that AI Photo Generator currently exposes OpenAI's markup parameter. Check the controls and supported models in the tool you use before preparing a dedicated overlay.
Markup, masks, and text instructions have different jobs
- Text describes the intended outcome. Specify the change and what must remain intact: “Remove the loose ribbon; preserve the box, label, and camera angle.”
- Markup provides a visual reference. The documented input is a transparent overlay on the first image. It is not a promise that a particular arrow or color has a standardized meaning.
- A mask is a separate editing input. Follow the selected tool's masking rules rather than treating a drawn circle as an interchangeable mask. OpenAI's reference explicitly disallows submitting markup and mask together.
Neither a tidy annotation nor a preservation instruction is a pixel-lock guarantee. If the job requires exact retention of a logo, legal label, or product component, inspect it carefully and use conventional compositing where necessary.
Prepare a clean visual editing brief in five steps
1. Choose one change that can be checked
Start with a narrow task, such as removing one distracting prop from a product photo. Avoid combining background replacement, packaging redesign, relighting, and object removal in the same first test. If the output fails, you need to know which instruction failed.
Write a one-sentence acceptance rule before generating: “The loose ribbon is gone, the product remains the same shape, and every letter on its label is unchanged.” Use source photographs you own or have permission to edit.
2. Normalize the source before drawing
Open the original in an image editor, make sure its orientation is correct, and record its displayed pixel dimensions. Create a transparent layer on that same canvas. For example, if the oriented source is 1536 by 1024 pixels, the overlay must also be 1536 by 1024 pixels.
Do not resize or rotate only one file afterward. The requirement concerns dimensions after orientation, so relying on the camera's stored orientation tag without checking the displayed result can cause confusion.
3. Mark the target without covering the evidence
For an initial experiment, draw a thin outline around the unwanted object, keeping the subject and its important details visible. Use one annotation rather than several competing colors. This is a proposed briefing technique, not documented syntax that every model is guaranteed to understand.
Export the annotation layer alone as PNG with transparency. Keep the original source separate. A screenshot of the photo with a circle drawn on it is not the same file structure as a transparent overlay. Reopen the PNG to confirm it has transparent areas; merely using a .png extension does not create transparency.
4. Explain the annotation in plain language
Pair the visual marker with a specific edit request. For a product scene, try:
The outline on the overlay identifies the loose ribbon beside the box. Remove only that ribbon and reconstruct the tabletop behind it. Keep the box geometry, printed label, colors, camera angle, and existing shadows consistent with the source. The outline is an editing instruction and must not appear in the finished image.
For an interior photo, adapt the same structure:
The outline identifies the cable on the floor beside the chair. Remove the cable and continue the existing floor texture. Preserve the chair legs, rug border, furniture positions, and lighting. Do not include the annotation in the output.
These examples separate the target, replacement, and preservation requirements. They are starting prompts, not claims of tested success.
5. Review against the original, not from memory
Inspect the source and result side by side at full resolution. Check the edited area first, then deliberately look outside it. An apparently clean removal can still alter a nearby shadow, soften label text, or change an object's proportions.
- Target: Was the intended object changed, and only in the intended way?
- Edges: Are there halos, duplicated textures, or broken lines?
- Identity: Are product shape, face details, and distinctive markings preserved?
- Text: Does every important character still match the source?
- Cleanup: Have all visible annotation marks disappeared?
How to tell whether markup helps your workflow
Run a small controlled comparison rather than trusting a single attractive example. Use the same source image, model, output settings, and edit instruction. Compare a text-only request with a text-plus-markup request where supported. Make three attempts per condition and retain all results, not just the best-looking one.
Record whether each result passes your acceptance rule, how much manual cleanup it needs, and the total generation cost. Six outputs are not a scientific benchmark, but they can reveal whether preparing an overlay is worthwhile for this specific job. If text alone already passes reliably, an extra annotation step may add work without improving the deliverable.
Common failures and the next sensible step
- The API rejects the overlay: Check PNG format, transparent pixels, matching oriented dimensions, model support, and whether a mask was also supplied.
- The wrong object changes: Simplify the annotation and explicitly identify the target relative to a stable nearby object.
- The drawn mark appears in the result: Clarify that it is guidance, not scene content. Reject outputs that retain it.
- Important untouched details drift: Reduce the edit scope, or finish with a conventional editor that preserves the original layers. Do not assume more retries will protect critical details.
The practical takeaway
The news is useful because it points to a more explicit way to communicate local edits. The sensible next step is not to wait for a rumored model: it is to prepare better source files, define one measurable change, and test the editing controls actually available to you.
When using AI Photo Generator, start with the image-to-image options and model choices currently shown in the app. Apply the same clear change-and-preserve brief. If a dedicated markup input is unavailable, do not assume an ordinary image upload implements it; use the supported editing workflow instead.