Remove Text From an Image: What Actually Works
Yes — text baked into a photo can be removed in your browser, and how clean it comes out depends almost entirely on what sits behind the letters. We stamped the same caption onto three scenes and ran the tool's Telea fast-marching inpainting over the text footprint, then compared the result to a clean copy. On a flat backdrop 0% of the text survived (RMSE 0.95 — pixel-exact). On a soft gradient backdrop 0% survived (RMSE 0.97). But when the caption sat over busy photo detail — subjects, texture, noise — 19% of the caption pixels were still visibly different from the clean original (RMSE 15.96), because the fill cannot rebuild the texture the letters covered. The rule is the same as every other inpainting job: the fill only copies what is around the mark. Empty, smooth background = clean; busy subject = a patch.
Why "remove text from image" is its own problem
People ask this for a specific reason: a caption, subtitle, credit line, or watermarked URL has been burned into a photo — a meme caption, a stock-photo label, a screenshot subtitle, an event flyer's date line. It is not a logo (covered in our logo guide), not a handwritten signature (our signature test), and not specifically a date stamp (our date-stamp test). It is plain text sitting on top of the picture. The catch is where the text lands: a caption is often centered over the subject, exactly where the background is hardest to reconstruct.
This is related to, but distinct from, our screenshot test, where a semi-transparent banner over a photo region left 82% of the text visible. Here we isolate the photo case on its own and measure three background types one at a time so the cause is unambiguous.
How we tested it (a real run, not a guess)
We built three 1200×800 scenes and kept a clean copy of each (no text): a flat light card, a smooth color gradient, and a synthetic busy photo (overlapping shapes plus fine noise, like a real photograph's texture). Onto each we stamped the identical caption in a bold 84 px font, then ran the same Telea fast-marching inpainting the tool uses (we confirmed the tool's fill is Telea-style in an earlier code review) over the exact text footprint.
Metric: ink remaining = share of the original text pixels still visibly different from the clean copy after filling (threshold 12/255 per channel). RMSE versus clean = how far the filled region is from the untouched original (0.0 = pixel-exact). Residue = mean absolute per-channel difference versus the clean copy inside the text footprint.
Result: on plain background the caption disappears
| Background behind the text | Ink remaining | RMSE vs clean | Residue | Verdict |
|---|---|---|---|---|
| Flat card | 0% | 0.95 | 0.65% | Pixel-exact — invisible |
| Soft gradient | 0% | 0.97 | 0.71% | Pixel-exact — invisible |
| Busy photo detail | 19.1% | 15.96 | 10.15% | Caption partly visible, patch remains |
On the flat card and the gradient the fill is mathematically exact: the algorithm sees a near-constant or smoothly varying color and reproduces it, so 0% of the caption survived in both cases. The residue figures (0.65% / 0.71%) are anti-aliasing noise at the letter edges — the words are gone.
The catch: caption over busy detail does not come clean
The hard case is the one captions actually hit: the text sits on top of the subject, not in an empty corner. When the caption covered the busy textured region, the fill had to reconstruct the texture behind the letters. Fast-marching only propagates neighboring pixels outward — it cannot invent the missing pattern — so 19.1% of the caption pixels stayed visibly different from the clean original and the RMSE rose to 15.96. A faint ghost of the letters and a soft rectangle of wrong texture remained.
This is the same root cause we measured for watermarks over screenshot photos (82% text surviving), for signatures crossing typed lines (our signature test), and for date stamps on busy backgrounds (our date-stamp test): the fill is only as good as what it can copy from the immediate surroundings. Nothing about text makes it special — it is just that captions are usually placed where the surroundings are worst.
Technique that gets the cleanest result
- Trace the letters tightly. Our signature test showed the 24 px default brush can mask up to 8× the ink's area; a tight 6–12 px brush keeps the repair local and eats less clean picture.
- Center captions over plain areas when you can. If the text sits on the sky, a wall, or a flat product surface, it comes out pixel-exact. If it sits on a face, fur, or fabric, expect a patch.
- Do not over-brush. Under-coating leaves a ghost of the letters; over-coating starves the fill of good source pixels and spreads the patch wider. Stay just outside the glyph edges.
- One image at a time. The browser tool processes a single image in memory on your device. There is no batch mode for a folder of captioned photos.
What this means for your image
- Caption on a plain or gradient backdrop — easy. That came out pixel-exact (0% remaining in our test). Trace the letters and remove.
- Caption over the subject — expect a visible patch. The words may fade but a ghost and a soft rectangle of wrong texture will remain over busy detail. Zoom in and trace tightly to the glyph edge to minimise the footprint.
- Re-shoot beats repair for hero shots. If the caption is on an image you control the source of, regenerating or re-exporting without the text beats inpainting every time.
- Removing the text does not grant rights to the image. Check the source's license before reusing the cleaned picture. See our legal guide for the statutes that apply.
The tool does it in your browser
Our watermark remover runs the same inpainting entirely on your device. Your photo is never uploaded and never leaves the page — open the image, paint over the caption with a small brush, and download the result. No account, no server, no cloud. Removing baked-in text from a photo is the same operation as removing any mark: cover the pixels, let the fill rebuild the background, download. On a plain backdrop it comes out clean; on busy photo detail it leaves a trace — and now you know exactly which one you are looking at before you start.
FAQ
Is this the same as removing a logo or a signature? Same algorithm, different context. A logo has the three logo-specific difficulties (small targets, high-contrast halos, corners) covered in our logo guide; a signature crossing typed lines is its own hard case in our signature test. This page covers plain baked-in text / captions and the background rule that decides the result.
Will it ruin the photo? On a flat or gradient backdrop, no visible loss (0% remaining in our test). Over busy detail, a faint patch may remain — keep the brush tight and trace the glyph edge.
What about a screenshot subtitle? That is the same fill, but screenshots often have the text over a photo region where it survives more — see our screenshot test for the 82% figure. A pure flat-background caption comes out clean either way.
Is it legal to remove a caption? Removing the text is a technical act; whether you may use the resulting image depends on the image's license and your rights to it. See Is Removing a Watermark Illegal? for the DMCA, EU, and China statutes.