RetypeWords

Once text has been flattened into a picture, it stops being text. There is no layer to click, no font to swap, and no way to fix a misspelled headline except to rebuild the graphic — which is impossible when the source file belongs to a designer who left six months ago.

RetypeWords is built for exactly that moment and nothing else. You upload a raster image, it reads the words it can see, you correct anything it misread, you type what those words should say instead, and it regenerates the affected area so the replacement belongs to the picture rather than sitting on top of it.

That framing matters, because it is not what a design app does. Putting a text box over the old words in Canva AI leaves the original lettering underneath and a patched background that rarely matches. RetypeWords changes the pixels — and, as the vendor says plainly on its own homepage, that comes with consequences worth understanding before you trust a result.

The RetypeWords Summary, In Five Lines

• What it does: replaces text that is already part of a PNG, JPG or WebP image, rewriting it in place rather than layering a new text box over it.
• Where it shines: flat, clearly printed lettering on digital artwork — posters, product graphics, packaging mock-ups, UI screenshots, and the garbled typography that Ideogram and other generators leave behind.
• Where it does not: handwriting, script and hand-painted signs, and any lettering on a curve or in perspective. The vendor states this on the product pages rather than in a support reply.
• The catch to plan for: the model redraws the whole frame, so detail outside your selection can shift. Every result has to be compared against the original before you use it.
• How you pay: a credit per generation task, however many words you change in that image, with retries and downloads inside the same task included.

RetypeWords At A Technical Glance

Why Can't You Just Edit Text In A Finished Image?

A JPG has no idea it contains words. It contains pixels that happen to form letter shapes, and the moment an export happens, the type layer, the font reference and the background behind the letters all stop existing separately.

The manual repair is genuinely laborious: mask out the old lettering, reconstruct whatever was behind it, identify a font close enough to pass, match the weight and letter spacing, redraw any shadow or texture the letters were sitting on, then position the replacement so the optical rhythm still reads. On a photograph of a sign, add perspective and lighting to that list.

Inpainting tools such as Cleanup.pictures solve half of it — they will erase the words and rebuild the background convincingly — but they leave you holding a clean gap and no way to put new lettering back in the original style. RetypeWords is aimed at the whole round trip: remove, rebuild and rewrite in one pass.

The Service is not a design tool and does not place a new text layer over your image. It changes the pixels. — RetypeWords Terms of Service, dated 3 September 2026

How Does RetypeWords Rewrite Words In Pixels?

The workflow is short enough to describe in one breath. Upload the image, wait for the scan to box the words it recognises, fix any character the OCR misread, tick the boxes you want changed, type the replacement wording, generate, then compare the result against the original before you confirm it.

The OCR correction step is the one people skip and then regret. If the scan reads a price as 519 and you do not notice, the generator faithfully rewrites 519 — the model is working from what the scan reported, not from what your eyes can see in the picture.

Several boxes can ride along in a single task, which is both a convenience and the basis of the pricing: an image where four separate labels change costs the same as one where a single word does.

What Style Matching Actually Promises

Style matching here means a close visual approximation of the lettering in your image: weight, size, colour, spacing and the way the characters meet the background behind them. It is not font recovery, and the vendor is explicit that no original typeface file is identified or reproduced.

On ordinary sans-serif and serif type the approximation holds up well. On a distinctive display face — a logotype, a custom masthead, anything with hand-drawn character — the difference will be visible to anyone who knows the original. For brand assets, the honest answer is still to open the vector file.

A Fifteen-Minute Test Before You Trust RetypeWords

Free usage covers one image per device, which is enough to learn whether your particular kind of picture is inside or outside what the tool handles. Spend it deliberately rather than on a screenshot you do not care about.

1. Pick the hardest image you actually need to edit, not a clean one. If your real work is menu boards shot at an angle, test a menu board shot at an angle.
2. Before uploading, write down the exact replacement wording, character for character. Deciding it in the input box is how a stray capital or a missing hyphen gets baked into the result.
3. Upload, then read every OCR box the scan produced — including the ones you are not changing. A misread elsewhere in the image is a warning about how legible the tool finds your typography.
4. Correct the boxes you will edit, select them, enter the replacement, and generate.
5. Open the result at full size, not at thumbnail size, and check the edited region first: spelling, line breaks, and any digit adjacent to the words you changed.
6. Pass/fail step: now compare the parts of the image you did not ask to change — small print, logos, faces, background signage. If anything outside your selection has shifted in a way you would have to explain to a client, this image is outside what the tool handles, and no amount of retrying will settle it.
7. If the edit passes, retry once anyway within the same task. Retries inside a task cost nothing, and seeing the second version tells you how stable the result is.

Which Images Suit RetypeWords, And Which Fight Back?

The vendor publishes its own boundary, which is rare enough to be worth repeating: flat, clearly rendered lettering on digital artwork is what the tool handles well. Everything below that line is described on the site as not there yet rather than as a limitation you discover after paying.

In practice, the images that come back cleanly share a few traits — printed type rather than written, front-on rather than angled, and enough resolution that each character is unambiguous before the model ever sees it.

• Reliable: posters, banners, sale graphics, product images, packaging mock-ups, UI and app screenshots, and generated artwork where the design is right but the lettering is gibberish.
• Workable with care: photographed signage that faces the camera squarely, image-based menus, and documents or forms with crisp printed text.
• Unreliable: handwriting, script and brush lettering, hand-painted signs, text wrapped around a curve or receding in perspective, and type so small or so heavily compressed that the characters have merged.
• Wrong tool entirely: logos and brand lettering, where an approximation of a custom typeface is worse than no edit at all.

Handwriting, Curves And Hand-Painted Signs

These fail for the same underlying reason. OCR has to decide what a character is before anything can be rewritten, and irregular strokes give it far less to work with than a typeface does. A model then has to reproduce that irregularity convincingly, which is a harder problem again.

Curved and angled lettering adds a second difficulty: the replacement has to follow a path and a vanishing point rather than a straight baseline. Product photography of cylindrical packaging is the common case here, and it is the one most likely to look nearly right and be subtly wrong.

What Happens To An Image After You Upload It?

Uploads and results are stored for 48 hours and then deleted, and the vendor states plainly that images are not used to train models and are not sold. Downloading a confirmed result again inside that window costs no additional credit, but once the window closes the file is gone, so save what you need.

Producing an edit involves handing the image to third parties: cloud hosting, an OCR provider and an image model provider, each named by category in the privacy policy along with what it receives. Payment details go to Stripe directly. Usage analytics run through PostHog, and the policy states that images, OCR text and replacement text are not sent to it; session replay is configured to mask page text and block uploaded artwork and generated results.

There is also an opt-in worth knowing about: you can allow the team to inspect a sample of images and results for quality checks, inside the same 48-hour window, and you can uncheck it in the result panel. It is off unless you choose it.

The practical read: this is proportionate handling for a small product, and it is documented in more detail than most listings of this size manage. It is still a third-party service, so confidential contracts, medical records and unreleased commercial assets deserve the same caution you would apply to any upload.

Where The Acceptable-Use Line Sits

RetypeWords deliberately does not screen images by what they look like — invoices, forms, certificates and screenshots all have ordinary reasons to be edited, for mock-ups, localisation, redaction and teaching material. The policy draws its line at conduct instead.

What you agree not to do is present an edited image as a genuine, unaltered record to someone who will rely on it as authentic, or use one to obtain money, credit, employment or status you are not entitled to. Identity and travel documents, prescriptions and professional qualifications, and material fabricated for legal or insurance proceedings are all named as prohibited. Enforcement is reactive, acting on reports rather than pre-screening, and the vendor says so rather than implying it catches everything.

Edited files may also carry embedded provenance information marking them as machine-edited, and stripping that to pass a result off as an original is a breach in itself. If you publish AI-modified media somewhere that requires it to be labelled, that labelling is your responsibility, not the tool's.

How RetypeWords Charges, And What One Credit Buys

The unit is a generation task, not a word and not an image area. Change one headline or change five separate labels in the same picture, and the same single credit covers it — which quietly rewards planning every replacement before you hit generate rather than running three separate tasks.

A credit is deducted when generation starts, and the model around that is more forgiving than most: retries within the same task cost nothing, downloading costs nothing, and a task that ends without producing an image returns the credit automatically. There are up to three processing attempts in total including automatic recovery. One asymmetry to note — a generated image consumes the credit even if you decide not to download it, so discarding a result you dislike does not refund it.

Two purchasing shapes exist: recurring monthly credit plans in three tiers, and one-time credit packs with no renewal. Monthly credits are described as rolling over without expiry, and cancelling stops future renewals rather than clawing back credits already granted. Checkout runs through Stripe.

Free usage is real but bounded: one image per device without an account, with sign-up required to download, and free exports carrying a visible RetypeWords mark that paid exports do not. AInexfinder does not publish plan amounts because vendor pricing changes without notice — the pricing page is the only current source.