BGNINJAOpen the tool → ◐

---

Remove text from an image

By ShinobiTools Team · Last updated: July 2026

Paint over unwanted words and BGNinja's Magic erase fills in the background behind them. Great for stray captions, date stamps and watermarks on your own images. Free, no account.

Remove a background now →

How erasing text actually works

Deleting text isn't the same as covering it with a box: you have to rebuild whatever was behind the letters. BGNinja's Magic erase uses an AI inpainting model (LaMa). You paint over the text, and it reconstructs the surface underneath, wall, sky, table or skin, so the words simply vanish instead of leaving a patch.

• Old date stamps burned in by a camera
• Stray captions or subtitles baked into a screenshot
• A watermark on an image you own
• Signs, labels or writing in the background of a shot

What it does well, and what's hard

Magic erase uses an AI inpainting model (LaMa) that paints in what it thinks belonged behind your selection. It's genuinely convincing on calm, fairly even backgrounds (a wall, grass, sky, a plain floor); very busy or detailed scenes are harder for any inpainter, so the fill can look smudged. Erase in smaller strokes and run it more than once for the cleanest result.

Text sitting on a plain wall, a clear sky or an even surface disappears cleanly. Text laid over a detailed pattern or a face is much harder to rebuild convincingly, so check the fill and redo it in smaller strokes if needed.

1. Open BGNinja and switch to ✨ Magic erase.
2. Upload the image (convert HEIC to JPG/PNG first).
3. Paint over every letter, plus a little margin around them.
4. Hit ✨ Erase it, check the fill, and repeat on anything left.

Please only remove text you're allowed to

Rebuilding a background works best when nothing important sat behind the words, so keep your strokes tight to the text. And only remove text from images you own or have the right to change, watermarks on other people's work are there for a reason. To take away whole objects or people the same way, see removing a person from a photo and removing objects.

Flat backgrounds erase perfectly, busy ones don't

Whether text comes out cleanly has almost nothing to do with the text and everything to do with what's behind it:

• Text on a flat area — sky, a plain wall, a solid colour block, an out-of-focus background. Effectively perfect: there's one colour to reproduce and the fill matches it.
• Text on an even texture — grass, sand, concrete, fabric. Very good. Texture is exactly what the model is good at inventing.
• Text over structure — brickwork, window frames, tiles, a straight horizon. Usable, but lines can come back slightly bent, because the fill reproduces texture rather than geometry.
• Text over a face or fine detail — the hardest case. An invented patch of face is uncanny even when technically smooth. Small text over a cheek is fine; a headline across someone's eyes is not.

Two details that trip people up. Include the outline and shadow. Most overlay text has a stroke, a glow or a drop shadow, and if you only cover the letters you leave a ghost of their silhouette behind. Brush a little wider than the glyphs. And size decides difficulty: a small caption is a thin hole with plenty of surroundings to sample; a giant headline across the middle is a large hole with little context, which is why it comes back blurrier.

The special case is a semi-transparent watermark. It isn't sitting on top of the image — it has tinted every pixel underneath it. Erasing it means replacing that whole region with invented content, so a large diagonal watermark leaves a large invented area. The result can be clean, but nothing recovers the original detail: it was never in the file you have.

Sometimes covering beats erasing

Erasing invents plausible content, and there are cases where "plausible" is the wrong goal:

• Redacting personal data — an address, an account number, a name, a face in a screenshot. Here you specifically do not want an invented replacement that looks like real content. Use a solid block. Avoid blur and pixelation for anything that matters: both are transformations of the original data and, for short predictable strings, can sometimes be reversed. A flat rectangle can't be.
• Documents and records — an invoice, a contract, a form. Inventing content on a document is misleading even when it looks tidy. Cover, crop, or reissue the document properly. For cleaning up a scan's contrast, shadows and skew rather than its content, our sister tool ScanReviver is the right one.
• When the detail genuinely isn't recoverable — a large watermark over an important area. A crop is honest; an invented reconstruction quietly isn't.

Two extra notes on redaction, since it's the case with real consequences. First, redact and then export a flattened copy — a layered file or a PDF annotation can have the covering shape moved or deleted by whoever receives it, which has embarrassed larger organisations than yours. Second, an image can still carry metadata: location, device, timestamps. Removing text from the picture doesn't touch any of that.

And the obvious one, which we'd rather say plainly: only remove text you have the right to remove. Your own photo, your own screenshot, your own document, a watermark on an image you've licensed — fine. Someone else's watermark or credit line, to pass their work off as unmarked, is not something we want this used for.

Common questions