How to do it, step by step
- Start a fresh chat for each photo Identity persistence lives inside one conversation. Starting a new thread loses the face the model was holding, and mixing two photos in one thread makes it blend them.
- Attach the largest version you have Upload from your photo library rather than a screenshot or a re-share. The model downsamples anyway, and starting from a compressed copy compounds the loss.
- Say what to change and what to keep Name both halves. "Remove the scooter on the left, keep her face, hair and the wall texture exactly as they are" beats "clean this up" every time.
- Mark the area rather than describing it Use the select tool to paint over the region you mean. Positional descriptions such as "the one on the right" are still read unreliably, and marking an area limits how far the redraw spreads.
- Change one thing per message Sequential single edits in the same thread hold together far better than one message asking for four changes. If a result drifts, go back and re-edit the earlier image rather than stacking on top.
- Compare at full size before you keep it Open the result and the original side by side and look at eyes, hands, teeth, text and any repeating pattern. These are where invented detail shows first.
Which model is doing the editing
ChatGPT has no separate photo editor. Image work runs through whichever image model is wired into the chat window, and that model has changed three times in eighteen months, which is why most advice about it online is already stale.
The current one is GPT Image 2.5, announced on 8 September 2026 and rolled out across ChatGPT on desktop, mobile and web. Through the API it comes in two variants: Flare, the faster default, and Sunburst, which is slower and holds fine detail better. Inside the app you do not pick one. It replaced GPT Image 2, which arrived in April 2026, which replaced the GPT Image 1 series before that.
Two changes in 2.5 matter specifically for editing rather than generating. The first is identity persistence: the model holds a representation of a face across a single conversation, so the same person can survive four or five sequential edits instead of drifting into a stranger after one. The second is localised editing, which is closer to true inpainting than the whole-frame redraw that earlier versions did every time you asked for a small change.
Everything it produces carries C2PA provenance metadata, and from this release an invisible SynthID watermark as well. The file is marked as machine-made whether or not you say so.
Six edits it handles well
Ranked roughly by how reliably it gets them right on the first attempt.
- Replace or remove a background. The strongest use by a distance. Cutting a subject out and putting them somewhere plausible is the job these models were tuned for, and it now handles hair edges properly.
- Take an object out. Bins, cables, signage, a stranger in the distance. Mark the area with the select tool rather than describing the object by position, which the model still reads unreliably.
- Change the light or the time of day. Turning flat midday into late afternoon is convincing, because it is inventing something soft rather than something precise.
- Put legible text into an image. Text rendering was the headline weakness of image models until 2026 and is now genuinely usable for a mockup, a poster or a card. Short text. Dense small text still breaks.
- Combine several uploads. Give it a product shot and a scene and ask for one placed in the other. Multi-reference work is where the newer models pulled ahead.
- Turn a sketch into an image. The Sketch feature added with 2.5 takes a rough drawing as the layout and renders it properly, which is a faster way to describe composition than any paragraph of prompt.
Notice what is not on that list: nothing that requires the rest of the photo to be left alone.
Where it goes wrong
It does not edit your photo. It draws a new one. This is the single thing to understand. Even with localised editing, the output is a generated image informed by your upload, so grain, sensor noise, lens character and the exact texture of a jumper are re-invented. On a photo you care about, compare the two at full size before you keep the new one.
Resolution is capped and rarely matches your camera. The newest model supports large outputs, up to a 4K long edge on the API, but in practice the app returns something far smaller than a 48 megapixel phone photo, and sizes snap to supported ratios rather than your original crop. Expect to lose pixels and to get a slightly different frame than you uploaded.
Repeated edits erode the parts you wanted kept. OpenAI's own documentation acknowledges that pixel-identical regions are not maintained across repeated edits, and that consistency across separate generations, meaning a new chat, is weaker still. Do your edits in one thread, in one sitting.
Text, layout and non-Latin scripts. Dense text placement, precise positioning of elements in a complex layout, and scripts such as Chinese, Arabic and Hebrew remain documented weak points. Earlier versions also showed a warm colour bias and a habit of cropping tighter than asked; check skin tones against the original rather than trusting the render.
Faces, real people and refusals
Image policy around real people is the most-changed and least-predictable part of the product. OpenAI loosened parts of it in 2025, then tightened elsewhere, and the practical result is that requests involving an identifiable person, including in some cases your own face in your own photo, are sometimes refused with no explanation and no appeal.
What tends to work: framing the request as a change to the scene rather than to the person. "Remove the bin behind her" passes more often than "edit this photo of my sister". What tends to fail: anything that reads as a likeness change, a face swap, or a public figure.
This is not a bug to be routed around. If a tool refuses to alter someone's face, that is largely the point, and pages offering jailbreaks for it are selling you a policy violation. The honest workaround is a conventional editor, which does not ask permission because it is not generating a new person.
One practical consequence: if your edit is about a person, a local editor on your phone is usually both faster and more likely to finish. If your edit is about everything except the person, ChatGPT is a reasonable choice.
ChatGPT against an ordinary photo editor
They are not competitors so much as different tools that happen to overlap on two jobs. Match the job.
| What you want | Better in ChatGPT | Better in a normal editor | Why |
|---|---|---|---|
| Remove a stranger from a holiday photo | Sometimes | Usually | Apple Photos Clean Up or Snapseed keep the original file, resolution and grain intact |
| Put your product on a marble worktop | Yes | No | Requires inventing a scene, which is exactly what a generative model is for |
| Fix an orange colour cast on skin | No | Yes | A white balance slider is exact and reversible; a redraw is neither |
| Add a headline to an image for a post | Yes | Depends | Text rendering is now good enough for short lines, and faster than a layout app |
| Retouch a face naturally | No | Yes | Refusals aside, generated skin loses the texture that makes a face read as real |
| Extend a vertical shot to landscape | Yes | Sometimes | Generative fill beats content-aware fill on sky, sea and plain walls |
| Anything going to print at size | No | Yes | The output resolution will be lower than your original and the detail is invented |
A useful rule: if you can name the slider you want moved, use an editor. If you can only describe the result in a sentence, use ChatGPT.
Where ShutterAI fits
Ours is a camera app rather than a chat window, so the overlap is small and worth stating plainly. ShutterAI's Clean Up removes people from a shot, Glow retouches skin without smoothing away freckles or texture, True to You corrects skin-tone white balance, and Stretch the Frame extends a crop. Those all run on your photo at its own resolution rather than through a chat prompt, which is the difference that matters for anything you intend to keep.
It is not released yet. ShutterAI comes to iPhone this autumn and is free to start; the early list is on the homepage. Until then the answers above are the ones to use.
Questions people ask
How to edit photos with ChatGPT?
Attach the photo in a chat, describe the change in plain words, and use the select tool to mark the area you mean. Work one change at a time in the same conversation, then compare the result against the original at full size.
Which model does ChatGPT use to edit photos?
GPT Image 2.5, released on 8 September 2026. In the API it comes as Flare, the faster variant, and Sunburst, which is slower and holds detail better. In the ChatGPT app the choice is made for you.
Does ChatGPT keep my photo's original resolution?
No. The output is a newly generated image at one of the model's supported sizes, which is typically well below what a modern phone camera produces, and the aspect ratio snaps to a supported one. Keep your original.
Why does ChatGPT refuse to edit a photo of a person?
Its image policy restricts edits involving identifiable people, and the filter sometimes catches your own face. Framing the request as a change to the background rather than to the person is more likely to pass.
Can ChatGPT unblur or upscale a photo?
It can produce a sharper-looking version, but it is drawing new detail rather than recovering lost detail. That is fine for a social post and wrong for a document, an identification or anything going to print.