Gemsaver Gemini Watermark

Highlights

🧠 Local AI inpainting — ships with LaMa FP32, runs entirely offline in the browser via ONNX Runtime Web.
⚡️ Native button takeover — click Gemini’s “Download full-size image” and GemSaver auto-cleans plus saves.
🛡 Triple fallback — direct fetch → page context → debugger capture keeps working through 403/CORS blocks.
📥 Clear feedback — subtle status bubble shows “Generating… / Unlocking download… / Done ✅”.
🔒 Open-source & transparent — inspect or fork every part of the workflow.
How to Use

Visit https://gemini.google.com and hover any generated image.
Click the original “Download full-size image” control.
GemSaver removes the watermark locally and triggers the download.
If Google still blocks the request, drag the image into GemSaver’s overlay (optional fallback).
Privacy

All pixels and AI inference stay on-device. No uploads or telemetry.
Permissions: downloads (save file), scripting/offscreen (inject script + run ONNX), debugger (only for last-resort capture, detached immediately).
FAQ

Why does Chrome show “this page is being debugged”? When fallback #3 runs, Chrome must display that banner; nothing else is inspected.
Missing model file? Place lama_fp32.onnx inside src/assets/ following INSTALLATION.md.
Supported platforms? Desktop Chromium browsers only (Chrome, Edge, Brave…).
Highlights (bullet list)

One-click Gemini downloads, watermark-free
LaMa AI runs locally — total privacy
Automatic fallbacks overcome 403/CORS
Optional drag-and-drop workflow
Fully open source and free
Permission rationale

downloads: trigger automatic saving of processed images.
scripting, offscreen: inject page logic and run ONNX model in offscreen document.
debugger: capture image responses when Google blocks normal requests.
Host permissions for gemini.google.com and Google image CDNs.
Privacy policy snippet
GemSaver processes all images locally inside the user’s browser. No data is collected, stored, or transmitted to external servers.

Assets

Icons: 16/48/128/512 px.
Screenshots (1280×800 recommended): Gemini UI with status bubble, before/after comparison, debugger fallback tip, drag-and-drop overlay.

- Reduced package size to 12 MB by hosting the LaMa model on Cloudflare R2.
- Added host permission for https://s2.litsource.net/* and load the model at runtime with caching.
- No functional change: the “Download clean image” button still processes Gemini images fully on-device after the model is fetched once.