One image · Lettering lifted · Surface rebuilt underneath
AI Text Removal
Every eraser tool can make text disappear. The hard half is the second half — what goes in the hole. This kind of edit is judged entirely on the repair, so that is what this page is about.
A shopfront with four separate pieces of text in it
A run we did on this page, chosen because it is the awkward case: a painted fascia board, a stencilled notice on glass, a chalked A-board and a metal parking sign, sitting on four different materials. Drag the divider and check the fascia first — it is flat paint, which is the hardest surface to repair invisibly. Every output is AI-generated, so treat this as what the tool does rather than a result you will get back identically.
Drag to compare before and after
What taking the lettering out actually has to do
Two operations, and the second one is where every cheap tool gives up.
Removing text from a picture is an inpainting job. First the lettering is identified and masked — every stroke, plus the drop shadow, glow or outline that usually travels with it. Then the masked region is reconstructed from what surrounds it, which means inventing plausible surface: the continuation of a brick course, the fall-off of a gradient in a sky, the way a wood grain runs.
Older tools solved half of this with a blur or a patch cloned from somewhere nearby, and both fail on anything structured. A model that has seen enough photographs can instead rebuild the region so the join is invisible at full size, which is the only standard worth measuring against. Two things it is not: it is not OCR, so you never see the words listed anywhere, and it is not a licence — removing a credit line from an image does not change who owns it.
Creative engine
How a removal run works
Upload the image, keep the default prompt or name the surface, then compare the two at full size.
Busy surfaces are easier than empty ones
Counter-intuitive, but true. Texture hides a repair; a smooth painted wall or a clean gradient sky shows every seam. If your text sits on flat colour, look hard at the patch at full size before you use the result.
Say what should be there instead
The prompt is allowed to describe the repair. "Continue the brick course" or "plain sky" gives the reconstruction a target and removes the guesswork that produces smeared patches.
Three habits that decide the result
None of them is about the tool. They are about what you hand it and what you check afterwards.
Upload the largest copy you have
Reconstruction works from surrounding pixels, so a downscaled screenshot gives the model less to rebuild from and a resized crop gives it artefacts to copy. Start from the original file if it still exists.
Name what should replace the lettering
Add one clause to the prompt describing the surface — brick wall, painted metal, open sky, wood grain running left to right. This single sentence is the difference between a rebuilt area and a smudged one.
Inspect at 100%, not at thumbnail size
Zoom to the repaired region and look for three things: a rectangle edge, a repeated patch of texture, and any straight line that jogs where it crosses the patch. If you find one, rerun and describe the surface more precisely.
When the words in the frame are the problem
Four jobs where the text was never meant to be part of the picture.
Reusing a photo in another language
A product shot with English packaging copy cannot ship to a market that does not read it. Stripping the lettering leaves a clean plate that a designer can typeset over, which is far cheaper than a reshoot.
Getting a clean background plate
Signage, posters and number plates all date a photograph and tie it to a place. Take them out and the same frame works as a generic backdrop — the same instinct as building a plate from scratch, applied to a photo you already have.
Clearing captions off a screenshot
Subtitle bars, watermark strips and UI toasts sit over the part of the frame you actually wanted. Removing them and rebuilding the underlying image beats cropping, which throws away composition.
Removing stamps from your own archive
Date stamps burnt in by an old camera, or a studio credit on a photo you commissioned and own. Editing a file you already own is ordinary retouching; removing a mark from someone else's work is not, and no tool changes that.
Prepping an image for a mock-up
A blank shopfront, a blank billboard, a blank label. Text removal produces the empty version that a mock-up needs, so the client sees their own words rather than somebody else's.
Text removal FAQ
What people check before running an image through it.
Yes by default — the standard prompt asks for all lettering, signage, captions and stamps to go. If you want to keep one of them, say so explicitly: name the piece of text you want preserved and the rest will be removed around it.
It is rebuilt, not blurred. The model reconstructs the surface from what surrounds the removed strokes, continuing brick courses, gradients, grain and shadow through the patched area. On busy texture this is usually invisible; on flat paint or clean sky it is worth checking at full size.
The tool does not decide that and neither does removing a mark. Editing a photo you own or licensed is ordinary retouching. Taking a credit or watermark off someone else's work does not transfer any rights to you, and we would rather say that plainly than pretend the question does not exist.
It should not, and that is the thing to verify. The prompt pins composition, crop, colour and subject, so only the patched regions are meant to differ. Compare the two files side by side at full size — if the whole image has been subtly restyled, rerun rather than accept it.
Yes. The removal is not reading the words, it is finding marks that sit on top of a surface, so chalk, script, stencils and foreign scripts are all treated the same way. Very low-contrast text on a busy background is the case most likely to be missed — point at it in the prompt if it survives the first run.
It goes over HTTPS, lands in your own history and nowhere else. Nothing is read out of it — this is not OCR, so the words you removed are never extracted, listed or stored anywhere as text. Delete the image and both versions go.
Kavel pricing — pay only for what you generate
Credits are shared across everything Kavel runs. You see the cost before each run, and a failed job is never charged.
Free
What's included
- 15 credits before you sign up
- +50 a week from daily check-ins
- = 232 a month — 46 images or 5 videos
- No card, no account
- Unlimited free engineGrok Imagine
- Photo editingNano Banana 2 Lite
Secure checkout by Stripe. Card details never touch our servers.
Starter
What's included
- 3,150 credits per month
- Up to 630 images or 78 videosGrok ImagineNano Banana
- Clean downloads, no watermark
- Every model unlockedVeo 3.1SeedanceNano Banana
- Cost shown before you generate
Secure checkout by Stripe. Card details never touch our servers.
Creator
What's included
- 7,200 credits per month
- Up to 1,440 images or 180 videosGrok ImagineNano Banana
- Clean downloads, no watermark
- Every model unlockedVeo 3.1Seedance 2Nano Banana 2
- No shared queue
Secure checkout by Stripe. Card details never touch our servers.
Studio
What's included
- 13,200 credits per month
- Up to 2,640 images or 330 videosGrok ImagineNano Banana
- Clean downloads, no watermark
- Every model unlockedSeedance 2.5Kling 3.0Veo 3.1
- No shared queue
Secure checkout by Stripe. Card details never touch our servers.
Charges appear as “KAVEL AI” on your card statement. You can cancel any time from Settings → Billing; cancellation takes effect at the end of the period you already paid for. Operator details, the full model list, and the refund window are on the about page.
Other things to take out of a photo
Same repair problem, different thing sitting on top of the picture.
Lift the lettering out
Upload one image, name the surface underneath, and check the patch at full size — or clean up a whole product shot instead.