Fixing text in an AI-generated image is an inpainting problem, not a prompt problem

TL;DR: If the picture is right and one word is wrong, do not regenerate. Remove the glyphs, inpaint the surface behind them, then re-render the replacement. That is a local editing problem, not a prompt problem.

AI image generators are good at moons, gardens and lighting. They are still unreliable at spelling. A poster comes back with “MOONLIGT” instead of “MOONLIGHT”, a menu swaps a letter, or a book cover has a word that looks almost right until you read it.

The instinct is to reroll the prompt or the seed. That is the wrong default when the composition and style already work. A new generation changes the moon, the plants, the light and the layout, and it can introduce a different spelling error. You spend another round of credits and review time to fix one word.

Why this is not an OCR problem

OCR reads the lettering. It does not remove it, and it does not draw a replacement. A text layer from OCR would sit on top of the old pixels, not replace them.

The operation has three parts:

1. Locate and remove the old glyphs.
2. Rebuild the surface behind them (inpainting).
3. Re-render the new string with matching weight, spacing, colour and alignment.

Missing any one of them gives a visible failure: a ghost of the old word, a clean patch with no text, or a new word that clearly belongs to another typeface.

A practical workflow

1. Upload the image. JPG, PNG and WebP up to 10 MB.
2. Enter the current string exactly as it appears, including the misspelling. You can also select the text region in the preview.
3. Enter the replacement.
4. Generate, inspect at 100%, then download.

The example below is a generated poster. The headline was misspelled. Only the headline changed; the moon, garden, lights and layout are the same pixels.

What I check before accepting the result

• Letter edges at 100%. Zoom in; soft edges and duplicated strokes show up here.
• Background continuity. Grain, gradients and shadows should not restart at the edge of the replaced word.
• Baseline and spacing. The new word should sit on the same line and keep the same tracking.
• Colour and shadow. A flat fill over a lit surface is the usual giveaway.
• The rest of the image. Compare the untouched area against the original.

Limits

• The exact typeface is not recoverable from a flattened image. Weight, spacing and colour can be matched; the font file cannot.
• Rebuilt backgrounds are approximations, especially over faces, logos and detailed objects.
• Curved, decorative or very small text may need a second attempt.
• Dense non-Latin scripts vary more than Latin text; check at full size.
• This is one image at a time. It does not open PSD, AI or layered files, and it is not a batch pipeline.

When to reroll instead

Regenerate if the composition is wrong, if the image has many errors, or if you want a different style. Editing the words only makes sense when the picture is right.

Try it

The workflow above is the AI image text editor. The first edit is free; a Google sign-in protects the credit, and no payment method is needed. For ordinary screenshots, the same image text editor works the same way. If the problem is a plain typo, the fix a typo in an image page is a narrower entry point.

I work on Reword Image, the tool used in the example. The before/after image is a real product example, not a mockup.