Overlaid text is one of the easier things to remove from a photo — easier than most watermarks. There’s one case where it’s impossible, and it’s worth knowing which is which before you start.
Why text isolates well
Two properties help.
Sharp, high-contrast edges. Text has crisp boundaries with a clear distinction from what’s behind it. A tool can determine precisely where it starts and stops. A faint semi-transparent watermark spread across a whole image is genuinely harder to select cleanly.
Small footprint. Text usually covers a modest share of the frame, leaving plenty of surrounding context. Coverage is the main constraint on reconstruction quality, and text rarely tests it.
So text over ordinary photographic content — a wall, a shirt, grass, a blurred background — typically comes off invisibly.
The case that never works
Text over other text.
A caption across a shop sign. A date stamp over a document. A watermark across a screenshot of a screen. In all of these, what’s underneath is itself text — and text cannot be reconstructed.
Inpainting predicts how visual content continues, which works because texture is statistically predictable. Letterforms aren’t. There’s exactly one correct arrangement of pixels for a particular word, and no way to derive it from surrounding pixels.
The model will produce something. It’ll look like writing at a glance and dissolve into nonsense when you read it. That’s not a tool failing — the information was destroyed when the overlay was applied. See how AI watermark removal works.
The test: if you can’t read what’s under the text by looking at the photo yourself, no tool can recover it. It has no more information than you do.
The outline and shadow
The most common cause of a disappointing result, and easily avoided.
Overlaid text almost always carries a stroke, outline or drop shadow — added so it stays readable against varied backgrounds. Those extend past the visible letterforms.
Select only the characters and you’ll leave a faint ghost tracing exactly where the text was. Include everything that came with it, plus a few pixels of margin.
What to expect
| Text sits over… | Result |
|---|---|
| Sky, wall, grass, water | Invisible |
| Blurred background | Invisible |
| Clothing or skin | Very good |
| Busy detailed scene | Very good |
| A face | Risky — subtly uncanny |
| A sign, document or screen | Gibberish |
| Repeating pattern | Alignment may drift |
| Most of the image | Not viable |
Common jobs
Date stamps. Among the easiest cases — small, high-contrast, usually in a corner over fairly uniform content. Nearly always comes off cleanly.
Camera timestamps burned in by older devices. Same as above.
Captions added by a social app. Usually large with a heavy shadow. Easy to isolate; check the shadow is included.
Text watermarks — a username or URL. Typically semi-transparent, so some of the original survives in the blend and can be partly recovered. Often easier than opaque text. See remove watermark from photo.
Screenshots with UI text. Fine over a plain panel, impossible over content that is itself text.
Steps
- Open the highest-resolution copy you have.
- Select the text plus outline and shadow, generously.
- Reconstruct — with an inpainting tool, not a blur tool. See remove watermark without blur.
- Zoom to 100% and check the boundary for halos and colour shifts.
- Export at full resolution, maximum quality.
On iPhone
MarkOff’s tap-to-select approach suits text well, because the hard edges make boundary detection reliable — tap the text and the app picks up its full extent including the outline.
Reconstruction is AI inpainting with no blur step. Photos are processed in the cloud, where a larger model than a phone can run gives better results, and your image is used only to carry out the removal you asked for. Video is processed on-device.
Before you remove it
Text on a photo is often a credit. A photographer’s name, an artist’s signature, a source attribution.
Removing text from your own photo is your call. Removing a creator’s signature from their work and republishing it strips exactly the attribution that would have credited them — and in the US, removing copyright management information can be a separate legal claim from infringement itself.