Skip to content

Repository files navigation

✦ Gemini Watermark Remover Light

Remove supported visible Gemini / Nano Banana image watermarks directly in your browser, with an experimental local Video Light tool for Gemini/Veo diamond watermarks.

No media uploads. No processing backend. No account.

🌐 Open Image Tool · 🎬 Open Video Light · 🧩 Install Userscript

JavaScript Local processing Backend License

Ways to use it

🌐 Image web tool

For one-off images:

  1. Drop or choose a Gemini image.
  2. The visible watermark is processed locally in your browser.
  3. Compare the before/after result.
  4. Download the cleaned image as PNG, JPG, or WebP.

Your image never leaves your device. The site has no image-processing server or upload endpoint.

Open the image tool →

🎬 Video Light · experimental

Video Light is a separate browser-only MP4 tool for supported visible Gemini/Veo diamond watermarks.

Current pipeline:

  1. Read the local MP4 with Mediabunny and show the selected clip directly inside the upload workspace.
  2. Sample roughly 3 detection frames per second, bounded to 18–60 samples. These samples only locate and calibrate the watermark; export still processes every decoded frame.
  3. Score known layouts and, when needed, search the bottom-right area using multiple representative frames for relocated or smaller diamond variants.
  4. Apply an Allenk-inspired startup refinement around the detected anchor: a local ±4 px search tightens the position instead of trusting the catalog coordinate exactly.
  5. Compare the available calibrated 48, 96, and current 96-20260520 alpha profiles and keep an alternative only when it consistently improves the local restoration residual.
  6. Build a stable per-shot alpha seed, then use five feedback rounds on strong frames to match the restored watermark area to its local background. Per-frame gain is capped to ±0.05 around the shot consensus to avoid flicker or dark-hole over-removal.
  7. Weak-but-plausible watermark frames use the shot consensus instead of being skipped. Only frames with very little watermark evidence are left untouched.
  8. Apply reverse-alpha removal plus deterministic footprint repair. In the default High-quality cleanup mode, a small local FDnCNN denoiser then targets the remaining compression halo; WebGPU is preferred and WASM is the fallback.
  9. Rebuild the MP4 through Mediabunny's Conversion API and native WebCodecs H.264, preserving source timing and keeping the primary audio track when conversion supports it.

Current scope is deliberately conservative: low-confidence or ambiguous clips stop instead of blindly modifying the wrong region.

The video file is not uploaded. The page loads pinned browser dependencies and the cleanup model on demand; inference and video processing stay on-device. Current Chrome/Edge are the primary browser target because Video Light needs H.264 WebCodecs support.

The FDnCNN model is loaded only when High-quality cleanup is selected. If the model runtime is unavailable in a browser, Video Light automatically falls back to the deterministic cleanup path instead of failing the whole export.

The public allenk/VeoWatermarkRemover repository currently exposes the release/demo documentation rather than the native implementation and remastered mask assets themselves. Video Light therefore ports the documented position/gain behavior and uses the calibrated alpha maps available in the browser upstream instead of claiming to embed Allenk's private/native masks exactly.

Video bitrate

The default is 12 Mbps, which is the recommended balance for typical 720p/1080p clips.

  • 8 Mbps — smaller output, more additional compression.
  • 12 Mbps — balanced default.
  • 18 Mbps — larger output with less additional compression.

Bitrate affects only the newly encoded video quality and file size. It does not improve watermark removal.

Open Video Light →

🧩 Userscript

If you download Gemini images regularly, install the userscript once and keep using Gemini normally.

  1. Install Tampermonkey or another compatible userscript manager.
  2. Open the raw userscript.
  3. Confirm the installation.
  4. Download supported Gemini images as usual.

The userscript works on:

  • gemini.google.com
  • business.gemini.google

The userscript remains image-focused; Video Light lives on the web tool where full-file decoding and encoding belong.

Why Light?

  • Private by design — user media is processed locally.
  • Deterministic core — supported visible marks use calibrated reverse-alpha removal rather than generative fill.
  • Small default surface — the image tool and userscript stay compact; Video Light avoids ffmpeg.wasm and downloads the optional denoise model only when needed.
  • Fail-closed behavior — uncertain image/video detection is preferred over editing the wrong region.
  • Open source — the project is published under the MIT License.

What it removes

The image paths target the visible Gemini sparkle watermark on supported generated images.

Video Light targets supported visible Gemini/Veo diamond video watermarks. It is experimental and does not yet promise the same coverage or cleanup quality as dedicated native video tools on heavily compressed or unusual clips.

This project does not claim to remove SynthID or other invisible provenance signals.

Development

npm run check
npm test

Release notes and technical changes are kept in the changelog.

Credits

Based on the reverse-alpha approach and calibration work from GargantuaX/gemini-watermark-remover, with video behavior informed by the public work in allenk/VeoWatermarkRemover.

The on-demand FDnCNN cleanup path uses the public browser model exported by the upstream project from the allenk FDnCNN work; the model URL is pinned to a specific upstream commit for reproducibility.

This is an independent open-source project and is not affiliated with Google. Gemini and related names are trademarks of their respective owners.

Released under the MIT License.

Releases

Packages

Contributors

Languages