Precise Edits and Reference Images in ChatGPT Images 2.5

Why these two capabilities go together

OpenAI leads the Images 2.5 announcement with reference fidelity ("better at preserving the subjects in your reference photos") and precision editing ("better at editing only what you've asked for, while keeping the rest of the details the same"). In practice they are one skill: the model has to understand which pixels are load-bearing and which are up for grabs.

The prompt is where you tell it. A photo attached with no instruction could be a subject, a style, a composition, or a thing to edit. Images 2.5 guesses better than 2.0 did. It still guesses. This guide is about not making it guess.

Everything below the official claims is Depikt's method, built from the CHANGE / PRESERVE / MATCH framework we used with GPT Image 2 and adjusted for a model that is more likely to honor a preserve list.

Part 1: say what the reference is for

We use five reference intents in the Prompt workspace, plus a sketch mode. Each one produces a different clause.

Subject identity

The reference supplies who. Keep identity; ignore everything else about the photo.

Then describe the new scene as you would in any generation prompt. If you want the expression kept too, say so; by default we free it.

Style

The reference supplies how it looks. Copy palette, medium, and mood; ignore its subject.

Edit source

The reference is the image to modify. This is the precision-editing case and gets the full CHANGE / PRESERVE / MATCH treatment (Part 2).

Product or object

The reference supplies an exact object that must be reproduced faithfully: a bottle, a shoe, a logo mark. Identity rules apply to the object.

Composition

The reference supplies the framing and arrangement. Copy where things sit; replace what they are.

Sketch or layout

Images 2.5 ships with Sketch in ChatGPT: type @Sketch, draw, and the drawing becomes a visual guide. OpenAI's examples are a room layout and the contour of an outfit. A sketch is a composition reference with the subject drawn in. The prompt still carries the style and the specifics:

Part 2: the edit prompt

For an edit-source reference, write three lists.

CHANGE. The one thing (or the short list of things) that should be different. One change per turn is still the safest habit.

PRESERVE. Everything that must not move. Be literal: subject, pose, expression, clothing, framing and crop, background objects and their positions, color grade, grain. Images 2.5 is better at leaving unmentioned details alone, but the mentioned ones are the ones you can hold it to.

MATCH. How the new element should agree with the old ones: light direction, color temperature, depth of field, grain, perspective. This is the difference between an edit and a paste.

Comments as a focus tool

ChatGPT now lets you place a comment directly on an image. It is the best way to point at a region, and it does not replace the PRESERVE list. Use the comment for where and the prompt for what stays.

When to regenerate instead

If the change is larger than the preserved part, you are not editing, you are generating with a reference. Switch to a subject-identity or composition prompt and describe the whole scene. Edits are for surgery.

Part 3: multi-turn editing without drift

OpenAI says Images 2.5 follows instructions more reliably across multiple turns and that quality degrades less over a long session. Our routine for making the most of that:

1. Turn one is the anchor. Get the base image right before editing anything. If the face is wrong on turn one, no number of edits fixes it.
2. One change per turn. Two changes in one instruction means the model has to weigh them. Split them.
3. Restate PRESERVE every turn. Copy the list forward. Consistency improved; memory was not promised.
4. Say "same image" explicitly. "Edit the image from the previous turn" removes the chance of a fresh generation.
5. Check the anchor features. After every turn, look at the two or three features you care about most (face, product label, brand color). If one has drifted, go back one turn rather than trying to repair forward.
6. Restart after a big drift. Reattach the original as an edit source and redo the surviving edits in one prompt. Three good edits from a clean base beats eight edits on a degraded one.

Part 4: the Imago handoff

If you build the prompt in Depikt and open it in Imago (our ChatGPT handoff), the reference image does not travel with it; ChatGPT sessions do not share attachments. Depikt shows a re-attach note whenever the prompt depends on a reference. Attach the same image in the new session before you send the prompt.

Common failures and fixes

• The face changed. The reference was read as a style or composition reference. Add "SUBJECT IDENTITY only" and name the features to keep.
• The whole image was regenerated. No PRESERVE list, or the change was described as a new scene. Rewrite as CHANGE / PRESERVE / MATCH and say "edit the attached photo."
• The new object looks pasted in. No MATCH clause. Name the light direction, color temperature, and grain the new element must share.
• The product's label is wrong. Product references need the label text quoted in the prompt; the reference alone is not enough for text.
• Edits keep getting worse. You are five turns past a drift. Go back to the last good frame or restart from the original.

Paste any edit prompt into the Critique mode; it scores reference handling and edit preservation as essential dimensions and caps the overall score when either one fails. For finished prompts that use these clauses, browse the Prompt Library, or start from the 12 Images 2.5 recipes.

Sources

• OpenAI, "Introducing ChatGPT Images 2.5," September 8, 2026: openai.com/index/introducing-chatgpt-images-2-5
• OpenAI, ChatGPT Images 2.5 System Card: deploymentsafety.openai.com/chatgpt-images-2-5