How to Edit Photos With ChatGPT (and What It Still Refuses)

Yes, for a good range of edits. You upload an image, type what should be different — "make the sky overcast", "warm the whole thing up", "remove the bin on the left" — and the edited picture comes back. For casual, non-portrait work it is genuinely good, and it keeps getting better with each image-model update.

The limits show up the moment the photo has a recognisable person in it, which is most real photo work. There are three of them, and they are worth knowing before you spend an afternoon fighting the chat window.

In early November 2026 OpenAI updated the image model behind ChatGPT to GPT Image 2.6. The widely reported changes are sharper, more legible text inside generated images, better use of a reference image you attach, and short looping clips. OpenAI's image generation guide is the primary source for what the current model can be asked to do, and the broader history of its image models is on the DALL-E article; treat anything else as secondhand.

What the update does not change is the shape of the three walls below. A newer model renders text better and follows a reference more closely, but a general assistant still applies a broad refusal on photos of identifiable people, still tends to regenerate the whole frame, and still returns a screen-sized image. Those are product and policy choices, not model limits a version bump removes.

It declines most edits to an identifiable face

A general-purpose assistant refuses the majority of requests that alter a photo of an identifiable person. That is a reasonable default for a chatbot and a dead end for anyone editing their own photos — retouching a portrait, changing an outfit, tidying a background behind a person.

It regenerates the whole image, not one part

Ask it to change the shirt and the background quietly shifts too, because it produces a new, similar-looking image rather than your image with one change. That is whole-image regeneration, and it is the main reason multi-step edits drift away from the original.

It hands back a screen-sized file

The output is a fixed, modest resolution that looks fine in a message bubble and is useless for a print, a poster, a product listing or a page layout.

Whichever tool you use, the wording decides the result. Name the thing and the change, and nothing else.

• Good: "remove the car on the right", "change the jacket to dark green", "make the light look like late afternoon".
• Avoid: long stylistic paragraphs. They push the model toward regenerating the whole picture, which is the opposite of an edit.
• Build in steps: make one change, check it, then give the next instruction, so each edit is small and correctable.

If you want a library of edits worded this way, our AI photo editing prompts post has copy-paste examples with before-and-after results.

When the job is a photo of a real person — you, a client, a model you shot — a general assistant is the wrong tool, not because the model is weak but because the product declines the request. A photo editor built for this can ask instead that you have the right to the image, and then apply narrower rules.

That is what LazyKiwi's prompt-based photo editor is for. It works the same way — you type the change in plain English and the edited photo comes back — with the three walls addressed: edits to people are supported on images you own or have permission to use, the change is masked to the region you named and composited back so faces, framing and the untouched background survive exactly, and the export goes up to print resolution. It is an independent tool, not ChatGPT and not made by OpenAI.

• Works on real faces for images you own — the portrait and clothing edits a chat assistant usually declines.
• Region-masked, not regenerated, so the parts you did not mention are literally untouched and edits can be stacked.
• Full-resolution export, and a batch mode to run the same instruction across a folder of photos.

The content line is the same everywhere and worth stating plainly: no impersonation, no sexual content, nothing involving minors, and no fabricating a real event that did not happen — regardless of who owns the photo.

A sentence is the long way round whenever the job already has a name. These give more control in fewer steps than describing the same thing to any model: