Inspiration

I use NotebookLM and Google Gemini constantly, but every export comes stamped with a watermark in the corner. The only "removers" I could find all wanted me to upload my files to their servers — not great for private notes and documents — and most of them just slapped their own watermark back on. I wanted something that was free, private, and actually did the job, so I built it.

## What it does

NotebookLM Watermark Remover cleans watermarks from:

  • NotebookLM videos (overview/explainer exports)
  • PDF slide decks and PPTX presentations
  • Infographics
  • Google Gemini generated images

It can also trim the "Made with Google" tail from podcast audio and strip image metadata. Everything runs 100% in the browser — no files are uploaded, nothing is stored, and no account is required.

## How I built it

The whole pipeline is client-side, served as a static site (Astro + Cloudflare Pages):

  • Images / infographics — a Canvas pass scans the fixed bottom-right region, runs connected-component analysis on the dark watermark pixels, then reconstructs the background with a vertical gradient fill.
  • Gemini images — the watermark is an alpha-blended overlay, so it can be reversed almost losslessly. For each pixel:

$$ \text{original} = \frac{\text{watermarked} - \alpha \times 255}{1 - \alpha} $$

  • PDF / PPTX — pdf.js and JSZip unpack the file, each page/image goes through the image engine, then pdf-lib / JSZip rebuild it.
  • Video — FFmpeg compiled to WebAssembly runs a delogo filter directly in the browser.

Heavy libraries (FFmpeg core/WASM) are self-hosted on Cloudflare R2 to keep loading fast and avoid third-party CDN failures.

## Challenges I ran into

  • Running FFmpeg WASM in-browser needs SharedArrayBuffer, which means getting the COOP/COEP cross-origin isolation headers exactly right, and self-hosting the FFmpeg ESM bundle to dodge cross-origin Worker errors.
  • Detecting watermarks without a fixed template — sizes and positions vary, so the connected-component + heuristics approach had to be tuned to avoid eating real content.
  • Reversing the Gemini alpha overlay cleanly for non-standard image sizes, with an inpainting fallback when the sparkle center can't be located.

## Accomplishments I'm proud of

A genuinely private tool: your files never leave your device, it's completely free, and the Gemini removal is mathematically near-lossless rather than a blurry patch.

## What I learned

A surprising amount of "server-side" media processing can now happen entirely client-side thanks to WebAssembly and the Canvas API — fast enough to feel instant, with perfect privacy as a side effect.

## What's next

More export formats, better detection for edge cases, and broader language coverage (currently available in 9 languages).

Built With

Share this project:

Updates