Comment editing in GPT Image 2.5 lets you mark a spot on an existing image and describe the change there, rather than describing the whole picture again.
It is one of four creative controls that shipped with the model on 8 September 2026, and it is the one that changes how a revision round feels: the instruction travels with the place it applies to.
What comment editing does in GPT Image 2.5
An annotation turns feedback into a location. You open an image, place a comment or a selection on the part that should change, and describe what should happen in that area.
OpenAI’s announcement describes it as the ability to “place comments directly on images for more focused editing”, and groups it with the other new product controls rather than with the API. A revision becomes a decision about one region, which is easier to review and easier to accept or reject than a rewritten prompt.
- A comment or selection placed on the image itself
- A written instruction scoped to that area
- The rest of the frame intended to stay as it is
- Later rounds that build on the image you already approved
- A visible history, because the edit stays in the same conversation
How to annotate an image on mobile
The help centre documents the mobile flow on a full-screen image, and the order of the steps is the part people get wrong.
- Open the ChatGPT mobile app and select an image generated by ChatGPT Images to open it full-screen
- Select Edit to describe your changes in the conversation panel
- Select Select to use the selection tool and highlight the part you want to edit
- Use the slider to change the size of the selection tool, and Undo or Redo to revise your selections
- After you finish your selections, select Next
- In the conversation panel, describe the changes you want to make to the highlighted areas
On mobile the image controls also include Resize and Remove. The detail that matters is that the written instruction comes after the selection is confirmed, not before.
Editing an image in the web editor
The web flow uses an editor panel rather than a comment pin, and it exposes the region controls as buttons.
- To open the editor, select an image in your conversation or in Images
- Select Select to highlight the area you want to edit
- Select Aspect ratio to regenerate the image in a different aspect ratio
- Select Undo or Redo to revise your selections, or Cancel to start over
- Select Save when you are finished
The help centre adds one alternative: you can describe your edit in the editor text box without using the selection tool at all. The region is therefore optional on the web, while the annotation is the point on mobile.
Comment, Select and Sketch are three different controls
Write-ups elsewhere often merge these into one feature, and most confusion about results starts there.
- Comment places written feedback on a specific part of an image
- Select highlights a region so an instruction can be scoped to it
- Sketch gives the model a drawn reference for a new layout
- Only Comment is a revision tool; Sketch is an input to generation
- None of them is a pixel mask, so none of them freezes the rest of the frame
Knowing which control you are using tells you which kind of failure to expect. For the drawing path, see the GPT Image 2.5 Sketch guide; for what the model does with no annotation at all, see the GPT Image 2.5 overview.
Writing an edit note the model can act on
A comment is a work instruction, and it behaves best when it reads like one.
- Name the target: the label, the background, the left hand
- State the change precisely, and quote any visible text exactly as it should appear
- List what must be preserved: pose, crop, lighting, unmarked text
- Change one concept per round instead of stacking requests
- Close by stating that nothing else should change
A useful test is whether a retoucher could act on the note in five minutes. If the note says “make it more premium”, the request is really a new generation rather than an edit, and it should be treated that way.
A worked example: one detail on a portrait
The difference a located note makes is easiest to see on a frame that already works. Keep an approved version first, then annotate one detail on it.
- Choose the smallest detail that is actually wrong — a crease on a collar, a stray object on a table
- Place the comment on that detail, not on the subject’s face
- Write the instruction as a change plus a protection list: “Remove the crease on the collar. Keep the face, the lighting, the crop and the background exactly as they are.”
- Confirm the selection, then send the instruction from the conversation panel
- Compare the result against the approved version, not against the previous attempt
A vague note turns the same request into a regeneration: “make the portrait more professional” gives the model no single thing to hold constant, so it re-decides the pose and the light as well. The protection sentence is what keeps a small fix small. If the element that keeps moving is the person rather than the detail, the GPT Image 2.5 character consistency guide covers that case.
What annotations cannot guarantee
The help centre is explicit about the limit, and repeats the same sentence in both the web and mobile sections.
- “Highlights are not always precise, and edits may extend beyond the area you selected”
- Nearby elements can still shift, including hands, shadows and reflections
- Visible text is the highest-risk element, because a plausible render can still carry the wrong characters
- Generation may take a few minutes, depending on the complexity of the request
So keep an approved version, compare each result against it, and return to that version rather than stacking corrections on a failed output.
Frequently asked questions
Can I use comment editing through the API?
The comment control is documented as a ChatGPT interface feature, and the help centre does not describe an equivalent API parameter. On the API, the comparable path is an image edit request that supplies a reference image.
Can I annotate an image I uploaded myself?
The control applies to an image you are working on in the conversation, so the same discipline applies: edit only what you own or are licensed to change, and approve a source version before you start.
Does a selection change only the marked area?
It directs the request to the marked area rather than restricting it. OpenAI warns that highlights are not always precise and that edits may extend beyond the selection, so protect important elements in writing as well.
How do I keep the rest of the image unchanged?
Name what must survive: face, pose, logo placement, text hierarchy, crop and light direction. Then finish with a statement that no other change should be made, and compare the result against the approved version.
Comment editing is worth adopting because it matches how people actually give feedback: point at the thing, say what changed. Keep an approved source, change one thing per round, and review the whole frame before you accept the result.