How to change the text in an image with AI: edit, erase or split
Three ways to change the words in an existing picture: an edit prompt, erasing the text and setting your own, or splitting it into layers. When each one works.

A poster says 12 May and the event moved to 19 May. A product shot has last year's label. A menu board needs a new price. You have the picture, it is right in every other way, and you need different words in it. There are three ways to do that in the catalog, and they suit different pictures. Picking the right one first saves more than any prompt trick.
Route 1: ask an edit model to swap the words
An edit model takes your picture and a sentence and returns a changed picture. At least one of them lists editing text and graphics in its own description, and the typography-focused models have edit versions of their own. This is the right route when the words sit inside a scene (a shop sign, a T-shirt, a label on a bottle) and have to keep their perspective, their material and their lighting.
Replace the text "12 MAY" with "19 MAY". Keep the font, the letter size, the colour, the position and everything else in the picture exactly as it is.Three habits make the swap land more often:
- Keep the length close. "12 MAY" to "19 MAY" works far better than "12 MAY" to "SATURDAY 19 MAY, DOORS AT EIGHT". New words that need more room force the model to redraw the layout around them.
- Match the case. Capitals for capitals, lower case for lower case. Say so if the model drifts.
- One line per round. Change the date, check it, then change the price. Two changes in one sentence double the places it can go wrong.
The general rules of edit instructions are in the guide to writing edit prompts. Edit models often have a quality setting, so test the swap at a low step and only spend the high one on the version you keep.
When the rest of the picture must not move
An edit model redraws the whole frame, so even a good swap can shift a shadow or soften a face elsewhere. Two settings exist for that. One typography-focused edit model has an edit precision setting: its high value restores the pixels you did not ask to change (leave the size empty when you use it, as its own help text says). The same model accepts a mask, an image of the same size where black marks the area to edit and white the area to keep. With a mask over the date, nothing outside it is open to change. Another edit model has an input fidelity setting for keeping logos and faces from the source; leave it on high.
Route 2: erase the text and set your own
For flat graphics (a flyer, a banner, a social card) the cleanest result is often not to let any model write the new words. The catalog has a tool that separates a flat graphic into layers, pulling the text away from the background. Keep the clean background, open it in any design tool, and type the new words yourself in a real font. Spelling is then guaranteed, and next month's change takes a minute.
This is the same logic as leaving space for text when you generate a picture from scratch: the model makes the picture, you set the type. It is the best route whenever the words are long, legal, or change often.
Route 3: split an ad into editable layers
One more tool goes further: it splits a flat advertisement into image, text and vector layers and returns the whole layer document as JSON. Under its advanced settings you choose how the text comes back: traced letterforms that keep the exact shapes, or editable text in the nearest matching font. That makes it a tool for a design or developer workflow rather than for a quick fix, since the result is a document to build on, not a finished picture.
Which route for which picture
- Words inside a photo or a 3D scene: Route 1, with a mask if the rest must stay put.
- A flat graphic with a short change: Route 1 is quickest; Route 2 if the first try is off.
- A flat graphic you will update again and again: Route 2, once, then edit the type in your own tool.
- A whole ad to rebuild in a design tool: Route 3.
What none of them can promise
No route here matches a specific brand font by name; the closest is "the nearest matching font", and nearest is not the same. Long sentences, small print and rare words still come out misspelled more often than a short headline, so read every letter at full size before you publish; the post on fixing hands and text covers why. Curved or folded surfaces, like a label wrapped round a bottle, are the hardest case for every edit model. And do not use any of this on a picture that serves as proof of something: a receipt, a ticket, a certificate or a screenshot. Changing the words there is not editing, it is forgery.
Each route shows its exact price on the Generate button in the studio before you start, so it costs little to try the edit route first and fall back to erasing if the letters will not behave.
Questions
How do I replace one word in an AI image?
Open the picture in an edit model and write: replace the text "old" with "new", keep the font, size, colour and position. Keep the new word close in length to the old one.
Can I keep the rest of the image from changing?
Yes, with a model that accepts a mask or has a high edit precision setting. Both limit the change to the area you mark.
What if the new text keeps coming out misspelled?
Erase the text with a layer separation tool and type the words yourself in a design tool. Spelling is then guaranteed.


