gemini-logo-remover

Remove the Gemini watermark from images — locally, no server, no upload.

pip install gemini-logo-remover

from gemini_logo_remover import remove_logo

remove_logo("gemini-image.png", "clean.png")

How it works

Gemini stamps a translucent white star in the bottom-right corner: out = (1 − α)·original + α·white. This package reverses that exact blend, so the pixels come back instead of being guessed:

1. Locate the mark at the known, size-derived slot positions.
2. Fit the overlay's strength, its position (down to a fraction of a pixel, so no outline ring is left) and colour so one scaled alpha shape cancels it. Text and texture cannot be cancelled this way, so they are never mistaken for the mark.
3. Reverse the blend — lossless; text under the star comes back.
4. Level the edge: measure the small tone mismatch between inside and outside the star's outline (robustly, so real edges crossing it are outvoted) and remove it as a smooth offset, then clean compression dots along the edge - only isolated outliers in smooth areas, never real detail.
5. Check what is left, and inpaint only a leftover or a mark that no alpha overlay explains.

Everything outside the star's 96×96 box (plus a 2–4 px blend margin at its tips) is returned bit-for-bit as it was. It runs on your machine with numpy and Pillow — about 0.1 s per image.

The calibrated alpha maps for the current star come from gemini-watermark-remover (MIT, © 2025 Jad, © 2024 AllenK / Kwyshell).

Usage

One image

from gemini_logo_remover import remove_logo

result = remove_logo("in.png", "out.png")        # saves out.png
print(result.method, result.clean, result.warnings)

# no file written, work with the Pillow image instead
result = remove_logo("in.png")
result.image.show()
result.save("later.jpg")                          # .jpg -> JPEG, otherwise PNG

image can be a path, bytes, an open binary file, or a PIL.Image.

Result fields: image, method (alpha = lossless, inpaint, alpha+inpaint, none), detected, slot (v1 old mark / v2 current star / v3), confidence, clean, warnings, steps (what ran), output_path.

Many images, in parallel

from gemini_logo_remover import remove_logo_many

items = remove_logo_many(["a.png", "b.png", "c.jpg"], "cleaned/", workers=4)
for item in items:
    print(item.source, "ok" if item.ok else f"failed: {item.error}")

One bad file never stops the others; failures are returned in the list.

Command line

gemini-logo-remover photo.png                     # writes photo_clean.png next to it
gemini-logo-remover *.png -o cleaned/ --workers 4
gemini-logo-remover photo.png --mode alpha        # lossless only
python -m gemini_logo_remover photo.png           # same thing

Options

Errors

• InvalidImageError — the input is not a readable image.
• ApiError — only for backend="api".
• GeminiLogoRemoverError — base class of both.

Optional extras

pip install "gemini-logo-remover[api]"       # backend="api" (needs requests)
pip install "gemini-logo-remover[inpaint]"   # OpenCV, only used if you opt in (see below)

The rare marks that no alpha overlay explains are filled by a built-in numpy routine: smooth and predictable, it never draws streaks. OpenCV's Telea inpainting is off by default because on textured areas (carpets, fabric) it draws cross-shaped smears, and OpenCV is preinstalled in many environments (for example Colab). To use it anyway: install the extra and set GEMINI_LOGO_REMOVER_INPAINT=opencv. Most current Gemini images never need inpainting at all.

Hosted API instead of local

remove_logo("in.png", "out.png", backend="api",
            api_url="https://your-server.example/api/v1/remove", api_key="...")

Only the bottom-right corner (≤ 480×480 px) is uploaded and the cleaned corner is pasted back locally. Defaults can be set with GEMINI_LOGO_REMOVER_API_URL and GEMINI_LOGO_REMOVER_API_KEY.

Limits

• It targets original Gemini exports. Screenshots, resized or heavily re-compressed copies move the mark off its exact position and change its pixels, so the reversal may not apply; you then get method == "none" and a warning, never a silently damaged image.
• A very faint (1–3 levels) trace can remain on a few backgrounds, mostly dark panels and the tip over flat bands.
• It does not remove invisible watermarks (such as SynthID).

Development and publishing

cd python
pip install -e ".[dev]"
pytest

python -m build
twine check dist/*
twine upload dist/*

Bump version in pyproject.toml and __version__ for every release — PyPI never accepts the same version twice. tools/gen_alpha_data.py regenerates the embedded alpha maps from the web project.

Testing a change

python python/tools/check_testlogo.py                  # run the test folder, compare with the baseline
python python/tools/check_testlogo.py --save-baseline  # accept the current results as the new baseline

Writes cleaned images and before/now/baseline sheets to analysis/testlogo_run/. Exit code 1 means a regression.