Nobody has the source file.
Every long-lived project eventually hits the same small crisis. Something has to change on an image — a date on a poster, a price on a label, a version number in a screenshot — and nobody can find the file it was made from.
The artefact outlives the source, reliably. The designer has left. The agency that made the flyer closed. The design file lived in a project nobody was paying for, so it is gone. The screenshot came from a build that no longer exists, on a phone that has since been recycled. What is left is a flattened PNG, and the flattened PNG is now the authoritative record.
The flattened artefact
This is the point where text stops being text. In the source file a word is an object with a font and a size. In the export it is a patch of pixels that happens to resemble letters, sitting on a background that was composed to sit behind it. Changing it means repainting that background and redrawing the letters in a typeface you probably do not have.
The classic route is manual: fill to erase, then font matching, then perspective correction, then colour matching against lighting that varies across the image. For a poster that is a skilled hour. For an app screenshot with a status bar it is worse, because UI chrome is unforgiving. I have watched people rebuild an entire layout from scratch rather than change three letters, and it is not obviously the wrong decision.
What I use now
EditTextImage takes the boring route. You upload the image, type the text as it currently reads, type the replacement, and it regenerates that region in context while leaving the surrounding font, colour, size and perspective alone. Output comes back at 2K and the round trip is put at thirty to sixty seconds, which matches the shape of the job: a one-line change to an asset that already exists.
The jobs I reach it for are the ones where a rebuild is disproportionate. App-store screenshots localised into a dozen languages, where rebuilding every mockup is a day and swapping the string in the capture is not. A flyer whose date moved. A label whose weight changed. Ad creative where you want to test a second headline without booking anybody. It is also the only sane answer to “we only have the JPEG”.
It is credit-based rather than subscribed: $14.90 for 40 edits, $26.90 for 100, $39.90 for 180. Credits do not expire, commercial use is included, and a generation that fails is detected and refunded automatically. One free credit comes with signup, which renders at 1K rather than 2K — enough to judge the font matching against your own material before paying for anything.
The honest trade-offs
The site has a section titled “When is EditTextImage not the right tool?”, which is not the usual marketing instinct, and the answers are specific rather than hedged.
- Small text is the hard case. Lettering only a few pixels tall leaves no room to redraw a glyph cleanly, so the advice is to find the highest-resolution version you have and check the result closely.
- Partial words are unreliable. You are asked to type the whole word as it appears and the whole word as it should read, because given only a fragment the model cannot always tell where the change should stop. It shows up most in connected scripts such as Arabic and Persian.
- Uploads are single images up to 10 MB — PNG, JPEG or WebP. There is no PDF or multi-page handling, so you export the page you need first.
- Aspect ratios are bounded on the better model: between 1:3 and 3:1. Panoramas and long strips fall to Fast, which has no shape limit.
- It will not design anything. No canvas, no layout, no blank page — that is a job for Canva, and the site says so and links to a comparison rather than pretending otherwise.
The part I respect most
This category sits one step from document fraud, and most tools in it quietly look the other way, because that traffic converts. This one does the opposite. The generate button makes you confirm the edit is not intended to deceive, and the acceptable-use policy is a wall rather than a disclaimer.
The banned list is specific: government IDs, passports, visa applications and the letters supporting them, diplomas, transcripts, contracts bearing seals, tax documents, bank statements, currency. Payment-confirmation screenshots, refund-confirmation emails, transfer slips, fake order-confirmation pages. Edits that put false words in the mouths of real people, including fabricated social posts and fake quotes on news screenshots. And email or chat screenshots altered to change who sent what and when — which the policy correctly identifies as the mechanics of refund, romance and fake-support scams.
There is nuance where you would expect a blunt rule. Editing a screenshot of your own past message for clarity in a presentation is fine; editing it to mislead a third party is not. Violations terminate the account without refund, and request logs are kept and may be shared with law enforcement, including in the intended victim’s jurisdiction. That is a company writing down where its revenue is not, in a category where the money is precisely there.
None of that makes it cleverer than its competitors. It makes it usable by people who need to know what they are holding. The source file is still gone, though. Somebody should have put it in version control.