Free · Private · Open Source · Runs entirely in your browser
🚀 Try it live · 📥 Install Guide · 💡 Why · 🛠 Deploy your own · 🤝 Contribute
WatermarkOut removes watermarks from AI-generated images (Gemini, DALL·E, Midjourney, Bing Image Creator, Adobe Firefly, Meta AI) and short videos — entirely client-side. The AI model runs in your browser via WebAssembly. Your files never touch any server. There is no server.
There are 3 ways to use WatermarkOut. Detailed steps for each in INSTALL.md.
👉 https://iamramgarhia.github.io/free-ai-watermark-remover/
- Open the link
- Drop in an image or video
- Pick a preset (Gemini ✨, DALL·E, Midjourney, etc.) or draw your own rectangle
- Click Remove watermark
- Download the clean file
That's it. The file never left your browser.
Visit the URL → your browser shows an "Install" icon in the address bar → click it → WatermarkOut now has a real desktop icon and opens in its own window. Works on Windows, macOS, Linux, iOS, Android. Zero downloads from any app store.
📥 Full install steps per OS in INSTALL.md →
git clone https://github.com/IamRamgarhia/free-ai-watermark-remover.git
cd free-ai-watermark-remover/static
python -m http.server 8000
# Open http://localhost:8000No build step. No npm install. No compilation. Just static files.
📥 Step-by-step in INSTALL.md →
Online watermark removers charge ₹800–₹2,500 / month ($10–$30), upload your private photos to unknown servers, and add their own watermark on the free tier. WatermarkOut runs the AI inside your browser tab — verifiable in DevTools (Network tab stays empty after first load).
- ✅ Free forever — costs $0 to host on GitHub Pages
- ✅ Open source under MIT — fork, audit, modify
- ✅ Works offline after first visit
- ✅ Zero accounts, zero telemetry
| 🖼 Image watermark removal | JPG, PNG, WEBP, BMP — up to 100 MB |
| 🎬 Video watermark removal | MP4, WEBM, MOV — audio preserved, real-time playback preview |
| 🤖 MI-GAN inpainting | Picsart Research ICCV 2023 model, 29 MB |
| ✨ Watermark presets | Gemini, DALL·E, Midjourney, Bing, Firefly, Meta AI |
| 🎨 Manual masking | Rectangle, brush, eraser, undo |
| 🔄 BEFORE/AFTER compare | Toggle to see original vs cleaned |
| 🔒 100% offline AI | After first load, zero outbound requests |
| 📱 Installable PWA | Desktop icon, standalone window, offline support |
- Any browser supporting WebAssembly + Service Workers:
- Chrome / Edge 90+
- Firefox 90+
- Safari 16.4+ (macOS / iOS)
- ~250 MB free storage (model + app cache)
- WebGPU is auto-detected and used when available (huge speedup); falls back to WASM multi-threaded otherwise
┌─────────────────────────────────────────────────────────────┐
│ Your Browser Tab │
│ │
│ [Image] → [Canvas mask] → [MI-GAN AI (WebAssembly)] │
│ │ │ │
│ │ ▼ │
│ └─────────────────→ [Composite result] ──→ ⬇ Download │
└─────────────────────────────────────────────────────────────┘
↑
No server. No upload. No exit point.
The model
- MI-GAN (Picsart Research, ICCV 2023). A GAN-based inpainting model designed specifically for mobile / browser deployment.
- Exported to ONNX (29 MB) and hosted on Hugging Face.
- Runs via ONNX Runtime Web with WebGPU → WebGL → WASM fallback chain.
- Cached in IndexedDB after first download — never re-downloads.
The pipeline (matches Picsart's official Python demo.py):
- Find user's mask bounding box; crop image with context.
- Resize crop to model's fixed 512×512.
- Normalize image to
[-1, 1], mask to1=keep / 0=inpaint(yes, the polarity is inverted from LaMa). - Build 4-channel input:
[mask-0.5, R*mask, G*mask, B*mask]. - Run model. Output is
[-1, 1]float32; map to[0, 255]via(v * 0.5 + 0.5) * 255. - Composite into original at full resolution — only masked pixels are replaced, rest stays pixel-identical.
Cross-Origin Isolation
A service worker injects COOP / COEP headers so the page is crossOriginIsolated. This unlocks WebAssembly multi-threading + SIMD + WebGPU. Without it, inference is single-threaded WASM (~5× slower).
- Fork or clone this repo.
- Push to your own GitHub.
- Settings → Pages → Source: Deploy from a branch →
main//static - Your site is live at
https://<username>.github.io/watermarkout/.
Connect the repo, set the publish directory to static/, deploy. All free for static sites.
- Create a
CNAMEfile instatic/containing your domain (e.g.watermarkout.dicecodes.com). - Add a DNS CNAME record pointing to
<username>.github.io. - Free HTTPS is auto-provisioned by GitHub Pages.
# Just serve the static/ folder over HTTP. ANY static server works.
cd static
python -m http.server 8000
# Then open http://localhost:8000There is no build step. No Node, no npm, no bundler. Edit a file, refresh, see the change. The Service Worker is registered at first visit; for code changes during dev:
Ctrl+R— pick up code changes (preserves the cached AI model)Ctrl+Shift+R— hard refresh (bypasses Service Worker; will redownload model)
| Key | Action |
|---|---|
| R | Rectangle tool |
| B | Brush tool |
| E | Eraser tool |
| Space | Toggle BEFORE / AFTER compare |
| Ctrl+Z | Undo brush stroke |
| Esc | Start over (with confirm) |
| D | Download result |
- No telemetry. No analytics, no error reporting, no "anonymous usage statistics."
- No accounts. No email, no signup, ever.
- No cloud. The model fetches once from Hugging Face. After that, nothing leaves your device.
- Auditable. Open DevTools → Network tab → use the app → see zero requests. Source is MIT and inspectable.
| Limit | Why | Workaround |
|---|---|---|
| Image cap ~100 MB | Browser tab memory ceiling | Resize before upload |
| Short videos only (~30 s) | Per-frame processing overhead + memory | Trim long clips first |
| Model is fixed 512×512 | MI-GAN ONNX export | Internally we crop + resize for higher effective resolution at the mask |
| Static-mask video mode | One AI pass + copy patch onto all frames | Best for corner watermarks (95% case); moving content behind the watermark may show a static patch |
| Complex backgrounds (faces, fine text, architecture) | LaMa/MI-GAN inpainting limitation, not a bug | Use a tighter mask; consider desktop tools like IOPaint for hard cases |
| WebGPU not in all browsers yet | Browser support catching up | Auto-falls back to WASM multi-threading |
watermarkout/
├── README.md
├── LICENSE
├── CONTRIBUTING.md
├── WATERMARKOUT_BUILD_SPEC.md
├── .gitignore
└── static/ ← deploy this folder
├── index.html main app
├── about.html about page (DiceCodes branding)
├── offline.html shown when offline & uncached
├── manifest.webmanifest PWA manifest
├── sw.js Service Worker (offline + COI headers)
├── css/
│ ├── app.css
│ └── about.css
├── js/
│ ├── app.js main controller
│ ├── coi-bootstrap.js enables Cross-Origin Isolation
│ ├── inpainter.js MI-GAN ONNX inference
│ ├── model-cache.js IndexedDB model caching
│ ├── mask.js canvas brush/rect/eraser
│ ├── upload.js drag-drop, validation
│ ├── video.js video pipeline (decode → mask → encode)
│ ├── fs-folder.js File System Access API integration
│ ├── debug.js diagnostic helpers (no UI)
│ ├── toast.js notifications
│ ├── updates.js Service Worker update flow
│ └── version.js
└── assets/
├── logo.svg
├── logo-icon.svg
├── prince-avatar.png (optional, falls back to GitHub avatar)
└── icons/ PWA icons (192/512/maskable)
Built solo by Prince Ramgarhia (DiceCodes) in Batala, Punjab 🇮🇳.
Solo-built. No VC. No growth team. Just one developer trying to make pro-grade creative tooling permanently free for everyone.
If WatermarkOut saved you money, time, or a privacy headache:
- UPI (India):
princeramgarhiaa-1@okaxis - GitHub Sponsors: github.com/sponsors/IamRamgarhia
- ⭐ Star this repo — free and helps others find it
- MI-GAN — the inpainting model (Picsart-AI-Research, ICCV 2023)
- ONNX Runtime Web — Microsoft's browser ML runtime
- lxfater/inpaint-web — reference implementation we learned from
- Sanster/IOPaint — gold-standard LaMa wrapper, source of our composite logic
- dinoBOLT/Gemini-Watermark-Remover — reference for browser-based watermark removal
MIT — fork, modify, use commercially. Made with ❤️ in India 🇮🇳.