How to fix text in an AI-generated image: 3 ways compared
Three ways to fix a wrong word in an AI image — re-run the prompt, repaint a selection with generative fill, or mark the words and keep only those pixels.

There are three ways to fix a wrong word in an AI-generated image: re-run the prompt, repaint the area with generative fill, or mark the exact words and keep only those lines from the model's answer. Only the third leaves the rest of the picture — the composition, the faces, the colours you liked — as the exact file you started with.
Which one is right depends on how much of the picture you are willing to lose. If it is a poster you spent an evening on, a product shot you finally got right, or a slide someone already approved, losing any of it is not a fix.
Way 1: re-run the prompt
This is the usual advice, and it treats the image as cheap. A regeneration is not a fix. It is a new picture that happens to be similar: the colour drifts, the composition shifts, the face is subtly different, and the typo is now some other typo.
The reason is that the thing that made the typo is the thing you are asking to fix it. Rendering legible text has been the most-reported weakness of image models since they appeared, and new models still announce better text as a headline result — as in OpenAI's introduction of 4o image generation. A re-roll is the same dice again, and it re-rolls everything else in the frame with them.
Way 2: generative fill on the words
You select a region, type a prompt, and a model repaints the selection. Adobe's own description of Generative Fill is precise about the job: adding, extending or removing content in an image. Removing a photobomber or extending a sky is exactly that shape of problem, and generative fill is very good at it. For words it is the wrong shape, for two reasons.
The whole selection is repainted, not the letters
Everything inside the selection is regenerated — the words and the ground they sit on. On a flat colour you may never notice. On a gradient, a texture, a rule running through the text or a product edge behind it, the ground comes back reimagined, and the seam lands wherever the selection boundary was drawn.
A prompt is a vague way to point at a word
"Change the date to Friday" has to be matched to pixels by the model, inside a selection that usually holds more than the date. Text is already the weak spot of image models, and an imprecise instruction gives it more room to go wrong.
Way 3: mark the words and keep only those lines
Edit text in an image keeps the model and changes the two things that go wrong:
- The words are marked, not described. Each piece of text to change is boxed on the image the model sees, and the instruction names the box and the new words.
- Only the marked lines are kept. The model returns a whole picture, but only the marked lines are taken from its answer and laid back onto your original.
Here is the first half of that in the editor: the headline boxed on the picture, and the new words typed against its number rather than described in a prompt.
And here is the poster with its headline changed to AFTER DARK:
What actually changed?
"Only the marked lines" is a claim you can check. Subtract one file from the other, pixel by pixel, and this is where they differ:
Every change is in the headline: the strokes of NIGHT BLOOM that were erased, and the strokes of AFTER DARK drawn in their place, overlapping where the two words cross. The moon, the vines, the gradient and every line of small type are untouched — under 4% of the frame changed, all of it inside the headline.
One honest footnote if you try this yourself on the images on this page: they are compressed for the web, and compression leaves faint noise along every sharp edge — at most 22 levels out of 255 here. The map above shows differences larger than that. The file the editor gives you is a lossless PNG, so there is no such noise to see through.
A regeneration would have touched every one of those pixels a little; generative fill would have repainted the whole selected band behind the headline.
Why this matters most for screenshots
A headline with last quarter's date, a button with the old name, a "coming soon" that shipped — in a screenshot, the picture around the words is already correct. Regenerating it asks a model to imagine what your product looks like, and it will confidently imagine something adjacent. Taking back only the words leaves the interface exactly as it shipped. How to edit text in a screenshot measures this on a real page.
Which should you use?
| Re-run the prompt | Generative fill | Mark the words | |
|---|---|---|---|
| Built for | a new picture | objects, regions, backgrounds | words on the image |
| How the target is given | the whole prompt again | a selection and a prompt | the exact words, boxed |
| What is regenerated | everything | everything inside the selection | only the marked lines are kept |
| Outside the target | changes | untouched | untouched |
| Where it can show | anywhere | the selection boundary | a faint edge on textured ground (whole redraw avoids it) |
The decision is which question you are asking:
- "Give me a better picture." Re-run the prompt. Nothing about the old one is meant to survive.
- "Make this region into something else." Generative fill. The region is meant to change, and plausibility is the standard.
- "Change these words and touch nothing else." Mark the words. The picture is meant to survive, and the standard is that nothing but the words moved.
When should you use the whole redraw instead?
Marking the words has one weakness. Where the new words sit on a textured background — a photograph, grain, a pattern — the edge between the model's ground and your original can show as a faint outline around the new text. For that case the result has a Whole redraw switch: the model's full picture, with no edge, at the cost of small details elsewhere shifting. It is the same answer, so switching costs nothing; whichever is on screen is what downloads.
Where it works, and where it does not
- Comfortable: typeset words on flat, graded or evenly textured backgrounds — posters, slides, UI screenshots, banners. A handful of words at a time.
- Harder: words on busy photographs, where the background behind the old letters has to be invented and fine texture is approximated rather than recovered.
- Not today: lettering that is itself the artwork — a hand-drawn logo, a title painted into an illustration. Replacing it gives you clean type where the drawing used to be.
On the typeface: nothing that works from pixels can recover the original font file, and identifying one is its own unsolved problem — services like WhatTheFont exist for exactly that. The new words are drawn to resemble the lettering around them. On a short headline that carries most of the resemblance; on a logo it does not, and a logo is the case to take elsewhere.
The steps
- Open Edit text in an image and upload the picture.
- Draw a box around each piece of text to change, and type what it should say.
- Press Redraw — about a minute, one price however many boxes are in the round.
- Compare against the original, switch to Whole redraw only if you see an edge, then download the PNG.
If the new words must look like the original lettering, that case has its own page. If you are weighing Canva's built-in version, the comparison with Canva Grab Text covers what it needs and where it stops.
Questions
- Does fixing the text change the rest of the image?
- Not by default. The model redraws the picture, but only the lines you marked are taken from its answer; outside them every pixel is the original. The Whole redraw switch uses the model's full picture instead, for the rare case where the default leaves a visible edge.
- When is generative fill the better choice?
- When the thing you want changed is an object or a whole region rather than words — removing a person, extending a sky, replacing a background. That is what it is built for.
- What kinds of text can the redraw fix?
- Typeset words on reasonably even backgrounds — headlines, labels, buttons, dates, prices. Lettering drawn into dense artwork is harder, and the result should be compared against the original before export.
- Will the new words match the original font?
- They are drawn to resemble the lettering around them, but no tool working from pixels can recover the font file, and an exact match is not guaranteed. Short headlines in clear lettering match best.
- Is any of the three ways pixel-perfect?
- No. A re-run changes everything. Generative fill leaves everything outside the selection alone but regenerates everything inside it. Marking the words keeps the least of the model's picture, but its edge can show on a textured background — which is why the editor offers the model's whole redraw as a fallback.