Remove overlaid text and restore the pixels behind it

Why text is different from a logo

A logo is a compact shape with a mostly continuous interior. Text is the opposite: lots of thin strokes, lots of edge relative to area, and a huge proportion of partially-covered pixels along every curve. Anti-aliasing means the alpha value changes from nearly zero to nearly one across a single pixel at the edge of every stroke.

Region-based approaches — blur, fill, clone — cope badly with that geometry because they operate on a mask that has to enclose the whole caption. The mask is mostly background, so most of the pixels they alter did not need altering at all. That is where the ghost outline comes from.

A per-pixel inverse blend does not need a mask in that sense. It needs an alpha value for every pixel, and where that value is zero the pixel is left exactly as it was. Ninety per cent of the caption's bounding box is untouched, and the ten per cent that is corrected is corrected precisely.

Anti-aliasing, sub-pixel edges and why they matter

When a letterform is rasterised, edge pixels receive fractional coverage — a pixel half inside a stroke gets an alpha of roughly 0.5 times the overlay's base opacity. Treating that pixel as fully covered over-corrects it and leaves a bright fringe; treating it as clear under-corrects and leaves a dark one. Either way you get a visible outline.

The alpha maps used here store fractional coverage at every position, measured from reference renders of the actual overlay rather than approximated from a threshold. That is what allows glyph edges to disappear instead of turning into a halo.

The same precision helps with sub-pixel positioning. Overlays are not always placed on exact pixel boundaries, so the engine evaluates candidate offsets and keeps the alignment that best explains the observed pixel values before applying any correction.

Practical tips for text-heavy overlays

Text removal is more sensitive to source quality than logo removal, because compression artefacts cluster around exactly the high-contrast edges the correction depends on.

• Start from the original export; a re-saved JPEG will have blocky artefacts along every stroke.
• Do not resize before processing — resampling blends stroke edges with the background irreversibly.
• Avoid sharpening filters beforehand, which exaggerate the very artefacts that confuse detection.
• Zoom to 100% in the comparison view and check the thinnest strokes; that is where any residue would appear first.
• If detection fails, look for an earlier copy of the file rather than retrying the same compressed version.

What this tool does not do

It removes composited overlay text, not text that exists inside the photographed or generated scene. A street sign, a book cover, a slogan on a T-shirt — those are image content. There is no attenuated original behind them to recover, so removing them would mean generating replacement content, which this tool deliberately does not do.

It is also purpose-built for Gemini and Veo 3 style marks. Arbitrary captions added by unrelated software may not match any known template, in which case detection will report no supported overlay rather than attempting a guess.