How to edit text in an image without the original file.

09/30/26 on images, tutorials, workflow

Last month I wrote about the day the source file disappears and the flattened PNG becomes the authoritative record. What I skipped was the practical question that follows within the hour: the text in that image is wrong, and it has to change today. Here are the four routes, in the order I try them.

Route zero: find the source

Five minutes of asking beats an hour of inpainting. The designer who left still has a backup drive. The agency that closed still has a successor with the archive. The screenshot came from a build that can be re-run. Exhaust this route first — every option below is a repair, and repairs are worse than sources.

Route one: AI text editors

The newest route and the one that handles the hardest case: the words sit on a background you cannot cleanly erase, and they have to stay in the same font, size and perspective.

The mechanics with EditTextImage, which is the one I keep coming back to:

Upload the image (PNG, JPEG or WebP).
Type the text exactly as it currently reads — the whole word or line, not just the letters you want gone, because the model needs to know where the edit stops.
Type the replacement.
It regenerates that region in context and leaves the surrounding font, colour and perspective alone. Output at 2K, thirty to sixty seconds a round trip.

This is the route for the flyer whose date moved, the label whose weight changed, and the app screenshot that needs a localised string in twelve languages. It is credit-based rather than subscribed, one free credit comes with signup, and a failed generation is refunded automatically. The details are in the previous post.

Route two: Canva’s Grab Text

If the image is heading to social media anyway, upload it to Canva and use Grab Text: it detects the text regions and turns them into editable text elements. Free-tier friendly, fast, and you can restyle while you are in there. It is happiest on clean, flat backgrounds — text sitting on gradients, textures or photographs is where it starts leaving artefacts, and that is exactly where route one earns its credits.

Route three: Photoshop

The skilled hour. Select the region, fill to erase (content-aware fill does most of the work on ordinary backgrounds), then retype the text and match the typeface. Photoshop’s Match Font helps identify the face from the image itself; without it, a font identification site and a few trial downloads. Full control, best result at any size, and completely disproportionate for changing one date. This is the route when the asset is going to print at scale or the legal team is watching.

Route four: rebuild the asset

Sometimes the edit is a symptom. If three strings need to change and two of them are structural, or the asset is brand-sensitive enough that a regenerated region needs review anyway, remaking it from parts is faster than patching. Screenshots especially: re-running the capture in a build that exists is minutes, and the result is a real source file again.

The “same font” question

The most common phrasing of this job is “edit text in image with same font”. The honest answer: the contextual regeneration in route one is the only one that just keeps the font, because it redraws the replacement in the context of the original glyphs. Every manual route needs a font identification step first, and the more exotic the typeface, the more the match degrades. Connected scripts — Arabic, Persian, Devanagari — are the hard case for every tool in this category, so check closely before sending anything to print.

The honest trade-offs

Small text is the hard case everywhere. A few pixels of letter height leaves nothing to reconstruct a glyph from; find the highest-resolution version of the image first.
Partial words are unreliable. Type the whole word as it is and the whole word as it should be.
Brand- and legal-sensitive assets deserve route four regardless of how good the patch looks. A repair you cannot diff is a repair you cannot audit.
None of these routes survives contact with a rasterised gradient behind a curved baseline perfectly. Test on the worst image you actually have, not the demo image.

The real fix is still putting design files where the code lives — version control, or at least a shared drive with an owner. Until then, route one covers most of the “nobody has the file” tax.