You upload a photo of yourself, ask for a portrait in a different setting, and get a beautiful picture of someone who is almost—but not quite—you. If you make AI images from reference photos, you know the frustration.
OpenAI introduced ChatGPT Images 2.5 on 8 September 2026, with a particular emphasis on keeping subjects recognisable, changing only the parts you ask to change, and carrying those decisions through later edits. Those are the right problems to tackle. The interesting question is how to put them to work in an ordinary creative session, not how polished the launch examples look.
This guide is about that workflow. It is not a hands-on benchmark or a claim that every face, hand, label or edit now comes out perfectly.
What changed in Images 2.5?
OpenAI says the new model handles reference-photo likeness, natural light and texture more reliably than Images 2.0. It also says generation latency is up to 50% lower—a company comparison, not a speed test we ran. For editing, the claim is that a request to change one part of an image is less likely to unexpectedly redraw everything else, including over several rounds of changes.
There are new ways to communicate the change you want, too: Sketch lets you draw a rough visual guide; Templates provide starting layouts for common formats; and image comments can point to a specific area. OpenAI also added prompt sharing, so someone else can try an idea with their own details. The ChatGPT release notes say Templates are not yet available in Work mode and that existing image-generation limits have not changed. Interfaces vary by platform and rollout, so do not assume every button in a demo will appear in your account today.
For most PromptyBox readers, the practical improvement to test first is simpler: can you keep the person while changing the scene?
Give each reference image one job
A face photo and a mood-board image are different instructions. If you upload both without saying which one controls identity, an image generator may borrow the wrong person's face from the scene reference. Be explicit about their roles before describing the new scene.
Use image 1 only for the subject's exact face, likeness and skin tone. Use image 2 only for the room, clothing, pose and lighting. Make a new full-colour portrait of the subject from image 1 in the scene suggested by image 2. Do not copy the person from image 2. Keep the face recognisable and the anatomy natural.
That opening does more useful work than a long list of adjectives. You can add your camera angle, clothes and colour after it. If the generated face still drifts, make your next edit about the face alone rather than rewriting the entire scene.
Original editorial illustration of a reference-photo workflow—not a screenshot or an Images 2.5 test result.
There is another small but important habit: do not describe the reference person's hair, facial hair or facial structure unless you deliberately want to change them. The uploaded photo already communicates those details. A text description can accidentally compete with it. Our editorial portrait prompt guide goes deeper on setting, lens, pose and light without losing the person.
Edit one thing at a time
Imagine you like a product portrait except the backdrop is too warm. Starting over risks changing the object, camera position and reflections. A tighter follow-up gives the model less room to drift:
Change only the background from warm cream to muted blue. Keep the bottle, its label-free shape, size, position, cap, reflections, lighting direction and camera framing exactly as they are. Do not add any other objects or text.
You can apply the same pattern to a portrait: change the jacket, not the face; move the subject, not the room; soften the shadows, not the colour palette. OpenAI says Images 2.5 is better at honouring that kind of focused instruction over multiple edits. Your result is still worth checking at full size—especially fingers, jewellery, typography and repeated details.
Original editorial illustration of a single-variable background change—not a before-and-after benchmark from ChatGPT.
If you have several changes, make them in a sensible order: first identity and composition, then wardrobe or props, then colour and finishing details. Save a version you like before the next round. A later edit can still undo something you meant to keep.
When to use Sketch, Templates and comments
Some ideas are easier to point at than to explain. A rough rectangle for a product, a pencil line for the horizon or a circle where the subject should stand can make the composition clearer. OpenAI's launch post says you can type @Sketch to draw a visual guide and then add written instructions. The release notes specifically describe the mobile Sketch and comment flows; availability and placement may differ elsewhere.
Templates are useful when the final canvas matters—say, a poster or product photo—but treat the template as a starting composition, not as the whole creative brief. Comments are useful after a generation, when you can point to one area and say what to change. If you do not see these controls, a normal uploaded reference image plus a precise text instruction still works as the basic workflow.
A ten-minute test you can actually judge
You do not need to accept a product demo or our opinion. Try the same reference photo with three requests:
- Identity test: place the subject in a plainly different room, keeping face and skin tone. Look for recognisable features, not just a similar mood.
- Focused-edit test: keep the resulting image, change only one visible item, then compare the face, pose, light and framing with the previous version.
- Second-edit test: change a different item and check whether the first edit survives.
Keep the original photo and all three outputs side by side. If the model changes a detail you explicitly locked, name that detail in the next prompt. If it keeps happening, use a cleaner reference photo with the face unobstructed and fewer competing subjects. A good result is not merely attractive; it follows the brief.
Only upload photos you have the right to use, especially when another person is identifiable. OpenAI says its Images 2.5 safety approach continues to use C2PA metadata and invisible watermarking, but provenance tools are not a substitute for consent or an honest caption. If an image is illustrative rather than documentary, say so.
The bottom line
The most promising part of Images 2.5 is not the shinier image. It is the possibility of saying, “Keep this person, change that scene,” and then refining the result without starting over. OpenAI says the model is better at both steps; the useful proof is whether it holds your own reference and your own edits together.
Start with a clearly labelled face reference, ask for one scene, and edit one variable at a time. For more examples, browse our AI image prompts or read how to keep characters consistent across AI images.
Sources and disclosure
- OpenAI: Introducing ChatGPT Images 2.5 — launch features and company performance claims.
- OpenAI: ChatGPT release notes — feature availability and limits.
- OpenAI: Images 2.5 system card — safety and provenance approach.
The three images in this article are original editorial illustrations. They are not screenshots, controlled comparisons or evidence of Images 2.5 output quality.
