Remove a Logo Watermark From a Photo Without the Smudge

A logo is a much bigger problem than a line of text. It covers more pixels, its edges are more complex, and it is usually opaque rather than ghosted — which means the model has to rebuild a larger hole with less information about what was there. This is the workflow for marks like that, plus an honest map of where the technique stops.

The question that decides everything

If the logo belongs to the photographer, agency or brand that owns the picture, the answer is not a tool: it is a licence. Removing a mark to use somebody else’s image without paying is both an infringement of the photograph and, in the US, a separate violation under 17 U.S.C. §1202, which covers removing copyright management information. The full text of Title 17 is published if you want the exact wording.

Where this work is genuinely fine: your own product photography that a mock-up service marked; a client proof you are entitled to edit; a scan of a print you own and are archiving; a screenshot of your own app. The technique is neutral; the rights are the whole conversation.

Why a logo is harder than text

• Area. A 400-pixel-wide mark is a hole the model must fill from its surroundings; a 60-pixel caption is a scratch.
• Structure underneath. Logos are usually placed in a corner, but not always over something uniform. Sky and wall rebuild well; foliage, brick, hair and fabric do not.
• Two-tone marks. A logo with white text and a coloured icon contains more information than the model can infer, and leftovers show as pale ghosts.
• Hard edges. Straight sides and letter corners leave a visible seam if your mask is a pixel inside the mark. Give it room.

A two-pass strategy that beats one big mask

The single mistake that ruins most logo removals is painting one enormous blob and hoping. Do it in passes instead:

1. Open the photo in WatermarkRemover (PNG, JPG or WebP up to 25 MB; the first run downloads roughly a 200 MB model, then caches it).
2. Zoom to roughly 100% on the mark. Brush the outer ring of the logo first — outline it, including the anti-aliased fringe, and stay a few pixels clear of any subject edge nearby.
3. Run the erase. Now the interior of the mark sits inside already-rebuilt background, which is much better context than a single mask over a foreign shape.
4. Inspect at 100%, then paint a second, smaller mask over leftovers — the ghost of a letter, a coloured wedge. Second passes are cheap because the region is now small.
5. Where the mark overlapped structure, shrink the ambition: rebuild what you can, then hand-fix the remainder with a clone stamp in GIMP or Photopea.
6. Export PNG. Re-saving to JPG compresses exactly the area you just reconstructed and makes the seam visible again.

What the model can and cannot rebuild

This tool runs a LaMa-class inpainting network — Large Mask Inpainting, open source — whose distinguishing feature is a receptive field wide enough to reuse texture from across the image rather than only near the hole. That is why it can extend a wall, a patch of sky or a woven pattern convincingly, and why the tool can promise it rebuilds texture rather than smearing it.

What no general inpainting model does: invent a face, a hand, a legible sign, or a specific product shape that the mark covered. The tool says as much — logos, timestamps and small semi-transparent marks over fairly uniform backgrounds work best, and very large or high-contrast marks over detailed areas are hard for any inpainting model. When the mark sat on a person, the correct expectation is a soft, plausible blur, not a recovered detail.

Situations, ranked by likely outcome

Situation	Likely outcome	Notes
Corner mark over sky, blur or plain wall	Clean	Usually one pass
Mark over fabric or foliage	Patchy-smooth	Two passes; accept some loss of texture
Mark overlapping a person’s edge	Poor	Rebuild the background first, hand-fix the subject boundary
Mark covering a face or hand	Not realistic	Consider re-shooting or replacing the image
Repeated tiled logo across the frame	Partial	Every region is damaged; expect smoothing
Logo baked into a low-res screenshot	Mediocre	Little context to extrapolate from

Neighbouring problems this is not

If the goal is to make a background disappear rather than to fill a hole, that is matting, not inpainting — see making an image background transparent. If the goal is a date or handle rather than a graphic mark, the same tool applies with a much smaller mask: removing a text watermark. For objects — tourists, cables, bin bags — the model is identical but the masking strategy changes, covered in AI object removal. Our free image tools page lists the editors used for the hand-fix stage.

Frequently asked questions

Can an AI really remove a logo without leaving a patch?

Over uniform backgrounds, essentially yes, because there is little information to invent. Over textured or structured areas it reconstructs plausible texture rather than the original, and a careful viewer will find the seam. Large marks over faces or hands are outside what general inpainting does.

Is removing a company logo from a photo I took allowed?

Editing your own photograph is generally fine. What you do next matters: using the image to imply endorsement, or publishing a brand’s property in a commercial context, raises trademark and advertising-law questions that have nothing to do with the pixels.

Why do I still see a ghost of the mark after erasing?

Usually a mask that sat a pixel inside the anti-aliased edge. Paint a slightly wider mask, run again, and for two-tone marks do the outer ring first and the interior second — the second pass gets much better context.

Does the tool lower my image quality?

It keeps the original dimensions and changes only the painted region, so the untouched parts of your photo stay bit-for-bit identical. The quality risk is export: saving back to JPEG re-compresses the whole frame, which is why the output is PNG.

Do I have to upload the photo?

No — the model runs on your device, which is also why it can be free with no account and no daily cap. The trade-off is a one-time download of roughly 200 MB before the first result.