Privacy-first photo & video anonymization — local-first, zero third-party data collection.
A free, open-source tool by Web3Privacy Now for anonymizing faces and sensitive data in images, videos, audio, live camera/mic, and documents. Detection, rendering, and export run entirely in your browser — faces via YuNet, optional objects via YOLO, and sensitive text via on-device OCR (Tesseract.js), all on ONNX Runtime Web / WebAssembly. No uploads, no servers, no tracking.
Try it online · Source on GitHub · Roadmap · Custom image libraries · Report a bug
- Anonymization effects — blur, zoom blur, pixelate, blackout, emoji, custom image, ASCII art, glitch, Color Ball, noise, and source-derived Prism refraction
- Modular privacy detection — faces (YuNet, on by default) plus optional YOLO targets: people, license plates, screens, documents, signs, tattoos
- Sensitive text on photos — opt-in on-device OCR (Tesseract.js, EN + CS) finds emails, phone numbers, payment cards, IBANs, national IDs, crypto addresses, secrets/keys, and more, then boxes and redacts them
- Purposeful sensitivity defaults — faces 25%, people 25%, plates 50%, documents/IDs and sensitive text 100%, vehicles and extra classes 10%
- Progressive detection — face boxes appear immediately while YOLO/OCR scans continue in the background
- All YOLO classes — an optional sheet exposes every raw class the object model supports for power users, beyond the featured targets
- Tiny detection labels — each detected region gets a small caption of its type (faces stay label-free), toggleable in settings
- One effect per region — cross-type de-duplication guarantees a single chosen effect is applied to any overlapping area
- Auto detection — runs locally in your browser (ONNX Runtime WebAssembly); cloud APIs are never used
- Zone editing — draw rectangles or paint with a brush over any region
- Brush tool — variable-size brush with real-time preview
- Color adjustments — brightness, contrast, saturation, shadows, highlights + presets
- Transform effects — halftone, glitch, pixel shift (wave/zoom/shear/ripple/mirror), color shift
- Snapshot system — save intermediate versions as new images in the explorer
- Selectable analysis — check every frame or choose 0.25–120 checks/s (up to output FPS), with additional analysis passes and uncertain-interval markers for review
- Local video export — masks apply to every output frame; WebCodecs and a MediaRecorder fallback use private temporary storage and bounded queues
- Supported formats — MP4, WebM, MOV, AVI, MKV, M4V, OGV
- Video audio — keep original track, remove audio on export, or apply voice-distortion presets (preview; mux export best-effort)
- Manual frame fixes — capture the current timeline frame, retouch it as an image, then bake it back into the next video render
- 100% in-browser — video detection, masking, and encoding never leave your device
- Audio files — open WAV/MP3/OGG/M4A in a dedicated audio mode with a full-width waveform/sound graph
- A/B comparison — toggle between the original and the anonymized signal while scrubbing
- Presets — Maximum Mask, Heavy Scramble, Broken Timing (voice-modulation focused; no noise bed)
- Advanced controls — pitch, formant shift, ring modulation, bitcrush, tremolo/timing wobble, filters, and intensity
- Export — WAV (offline render) plus WebM/OGG/MP4 where the browser's
MediaRecordersupports them - Limitations — DSP-based disguise, not a guaranteed forensic defeat; no speech-to-text or speaker identification
- Live camera — real-time face/zone anonymization from the webcam, photo + video capture to library
- Voice Mask — live microphone de-identification via AudioWorklets; presets mirror audio mode (Maximum Mask, Heavy Scramble, Broken Timing, Off for raw recording)
- Capture preview — last photo/video thumbnail in the corner for quick open-in-editor
- Local only — capture, DSP, and recording never leave the device
- Formats — PDF, TXT, and MD (DOCX import is on the roadmap)
- Local PII detection — regex + checksum recognizers (Luhn cards, IBAN mod-97, Czech rodné číslo mod-11, emails, phones, IPs, crypto, secrets/keys)
- Review UI — colored highlights over text and PDF page renders; toggle individual detections, draw manual redaction boxes
- Redaction styles — PDFs open with blackout selected; TXT/MD open directly in the anonymized blackout preview; blur/pixelate remain available for PDFs and token replacement for text
- Safe export — flattened PDF, ZIP of page images, or token-replaced TXT; Copy/Save for text documents
- 6 image formats — JPEG, PNG, WebP, BMP, GIF, TIFF
- SVG vectorization — convert images to SVG using imagetracer.js with 8 presets and custom parameters (live preview)
- Batch processing — resize, crop, format conversion, color grading, transforms, auto-anonymize across hundreds of photos
- ZIP export — download all processed photos at once
- 100% local — images, videos, audio, mic, and documents never leave your device
- No analytics, no cookies, no tracking — zero third-party network requests at runtime
- Self-hosted fonts and models — Material Symbols, YuNet ONNX, optional YOLO ONNX, ORT WASM, and the Tesseract OCR engine + language data are all served from the same origin (no CDN)
- CPU timing proof — shows processing time to verify local execution
- CSP + cross-origin isolation — Content-Security-Policy and COOP/COEP/CORP headers in production
- Lean first load — reach the home screen before detector initialization; YuNet loads when photo detection or live mode needs it
- On-demand privacy models — heavier YOLO/OCR assets start loading only after the user enables a target that needs them; one progress surface reports the complete model load and respects Save-Data / slow connections
- Hypnotic home screen — a high-detail, dotted multi-layer GPU spiral illusion (WebGL fragment shader, pixel-by-pixel, high FPS) that reacts to pointer/touch and morphs from the logo, with a
prefers-reduced-motion/ no-WebGL SVG fallback
- Electron support is kept in the codebase for future desktop releases
- Public desktop downloads are temporarily hidden until the installers are polished
Visit anonymizer.web3privacy.info — everything runs in your browser.
# Clone and start
git clone https://github.com/web3privacy/w3pn-anonymizer.git
cd w3pn-anonymizer
npm install
npm run dev
# → http://localhost:5173Desktop packaging is currently kept in the repository but hidden from public download links while the installers are polished. For now, use the web app or browse the source on GitHub.
# macOS
npm run electron:build
# Windows
npm run electron:build:win
# Linux
npm run electron:build:linux
# All platforms
npm run electron:build:allOutput goes to the release/ directory. Release artifacts are grouped by platform and architecture, for example:
release/
├── linux/arm64/ # AppImage + deb
├── linux/x64/ # AppImage + deb
├── macos/arm64/ # dmg + zip
├── macos/x64/ # dmg + zip
├── windows/arm64/ # installer exe + portable exe
└── windows/x64/ # installer exe + portable exe
Local macOS builds created without an Apple Developer ID are unsigned and not notarized. Sign and notarize the .dmg / .zip artifacts before publishing them as official macOS downloads.
w3pn-anonymizer/
├── src/
│ ├── App.tsx # Main React application
│ ├── App.css / index.css # Core styles + CSS variables
│ ├── button-system.css # Shared button typography and geometry
│ ├── main.tsx # React entry point
│ ├── types.ts # Shared TypeScript types
│ ├── components/ # UI (incl. document/ viewers, tool-panels/, batch/)
│ ├── desktop/ · mobile/ # Desktop and mobile shells / layouts
│ ├── hooks/ # React hooks (detector, voice, library, prefetch…)
│ └── lib/
│ ├── detector.ts # YuNet face detection orchestrator (WASM)
│ ├── detectors/ # YuNet + YOLO + OCR-PII detector implementations
│ ├── detections/ # Detection config, adapters, availability, image run
│ ├── effects.ts # Image effects engine (blur, pixelate, glitch…)
│ ├── video.ts # Video frame-by-frame processing
│ ├── audio/ # Audio pipeline, presets, and live voice-mask DSP
│ ├── document/ # PDF/TXT parsing, PII recognizers, redaction/export
│ ├── gl/ # WebGL renderers (hypno spiral)
│ ├── asset-prefetch.ts # Background model/OCR cache warming
│ └── vectorize.ts # SVG vectorization (imagetracer.js)
├── server/ # Optional same-origin anonymous feedback service
├── android/ · ios/ # Capacitor wrappers; native device tests are in progress
├── electron/ # Electron main process
├── public/
│ ├── models/ # YuNet + optional privacy YOLO ONNX (see public/models/privacy/README.md)
│ ├── onnx/ # ONNX Runtime WebAssembly assets
│ ├── tesseract/ # Self-hosted OCR engine (LSTM core) + eng/ces language data
│ ├── worklets/ # AudioWorklet processors (level meter, noise gate, voice mask)
│ ├── fonts/ # Self-hosted Material Symbols
│ ├── vendor/ # Browser image compression + imagetracer libs
│ └── demo/ # Demo images, audio, video, and a sample document
├── scripts/ # Build/release tooling + Playwright e2e smokes
├── index.html · package.json · vite.config.ts · tsconfig.json
└── package-lock.json # Locked JavaScript dependencies
npm run build
# Output → dist/The dist/ folder is a static SPA deployable to nginx, Caddy, Apache, or another static host. Production currently serves it from nginx on the W3PN VPS.
The optional feedback service is documented in server/README.md. Media processing does not use a Python detection server.
- Node.js ≥ 20.19 — nodejs.org
- npm ≥ 10 (bundled with current Node.js)
See Dependencies and integrated runtimes for the complete package, browser API, model, worker, desktop, native iOS/Android, and optional server inventory.
- Python,
fastapi, anduvicorn; see server/README.md. - The static application and all media processing work without this service.
See native setup and device signing and USB instructions. Android debug builds are available for testing; iOS uses automatic Xcode signing. Store/F-Droid releases still require real-device acceptance tests and completion of the third-party asset license inventory.
- Faces (YuNet) — the browser loads
public/models/face_detection_yunet_2023mar.onnxand runs it via ONNX Runtime Web (WebAssembly). Large images are scanned full-frame plus 640 px tiles; video samples are downscaled to 1280 px on the long edge first. - Objects (YOLO, optional) — when a non-face target or a raw class is enabled, the matching YOLO ONNX model in
public/models/privacy/lazy-loads through ORT. Outputs are thresholded using the per-target defaults, NMS'd, and cross-type de-duplicated so each region yields a single effect. - Sensitive text (OCR) — when the Sensitive text target is on, Tesseract.js (self-hosted under
public/tesseract/, EN + CS) recognizes words with bounding boxes; the document PII recognizers then locate matches and box them for redaction. Runs on still images only. - No image, video, audio, or document pixels/text leave the browser.
- The app loads the source file into browser memory as a
Blob. - A preview
ObjectURLis created for the session UI. - YuNet runs in-browser via ONNX Runtime WebAssembly.
- The app stores only normalized face boxes and user-edited zones in React state.
- Anonymization effects are rendered onto canvases in the browser.
- Output is written to disk only if the user explicitly exports, downloads, or overwrites the original.
- The source video stays in browser memory as a
Blob. - The app samples detection frames across the timeline.
- Sampled frames are detected in-browser with YuNet.
- Timeline interpolation, masking, frame overrides, and final encoding stay in-browser.
- Audio is preserved from the source stream when the browser runtime supports it.
- The app stores loaded media, original backups, snapshots, zone masks, and temporary video overrides only in memory for the active session.
- Preview
ObjectURLs are revoked when media is replaced or removed, and remaining previews are revoked when the app unloads. - The app persists only UI preferences in
localStorage:anonymizer-theme— theme preferenceanonymizer-enable-optical-mode— home logo / hypno animation toggleanonymizer-privacy-settings— privacy target toggles, thresholds, label toggle, enabled raw classes, audio effect prefs (no media)anonymizer-voice-mask— live voice-mask preset/strength prefs (no audio)
sessionStoragemay hold lightweight live-capture metadata (anonymizer-live-meta) — never image blobs.
- Browser queue: up to 2,000 media items per session.
- Images: up to 50 MB per file in the browser queue.
- Videos: up to 500 MB per file in the browser queue.
- Audio: up to 100 MB per file in the browser queue.
- OCR languages: English + Czech (self-hosted
tessdata_fast, ~8 MB total). - Video detection: sampled frames are analyzed at up to 1280 px on the long edge.
- Video export: 6 Mbps video bitrate + 128 kbps audio bitrate.
- FPS handling: defaults to 30 fps when unavailable and normalizes detected rates into the 10-60 fps range.
- Batch resize controls: width and height inputs are clamped to 25,000 px.
- SVG preview: vectorization preview is capped to 1,200 px on the long edge for responsiveness.
- Loaded media, zone masks, snapshots, original backups, and temporary video overrides live in memory as
Blob/ObjectURLstate for the current session. - Preview
ObjectURLs are revoked when media is replaced or deleted, and remaining previews are revoked when the app unloads. - The app persists only UI preferences in
localStorage(theme, optical animation toggle). - No image or video content is written to disk unless you explicitly export, download, or overwrite originals through the File System Access API.
- ZIP export re-encodes images through canvas, which strips EXIF, GPS, ICC, and other embedded metadata.
- No data leaves the device. All detection, rendering, and export run locally in the browser.
- No sessions, cookies, or tracking. Pure SPA with no analytics and no third-party API calls at runtime.
- Self-hosted assets. Fonts, ONNX models, WASM binaries, and the Tesseract OCR engine + language data are served from the same origin.
- CSP + isolation headers. Content-Security-Policy and COOP/COEP/CORP configured for production (see
vercel.json). - Processing proof. The app displays CPU timing after each detection to verify local execution.
npm run build
npm run release:audit
# Deploy dist/ to static hosting (e.g. Vercel)Requirements:
- Production domain and metadata must remain
https://anonymizer.web3privacy.info - Serve
.mjsasapplication/javascript - Ensure
public/models/,public/onnx/,public/tesseract/, andpublic/worklets/assets are included in the deployment - Headers (configured in
vercel.json):Cross-Origin-Opener-Policy: same-originCross-Origin-Embedder-Policy: require-corpCross-Origin-Resource-Policy: same-originContent-Security-Policy(seevercel.json)
- Apply the same headers to model/WASM/OCR assets under
/onnx/,/models/, and/tesseract/ - Treat tracked
dist/files as release artifacts, not source of truth. Ifnpm run release:auditwarns about ignored untrackeddist/assets/*references, either include the matching build artifacts intentionally or deploy from the fresh source build output.
See docs/RUNTIME_AND_PRIVACY.md for a fuller runtime, privacy, and deployment walkthrough, and docs/RELEASE_CLEANUP_PLAN.md for the current cleanup checklist.
| Library | License | Purpose |
|---|---|---|
| React 18 | MIT | UI framework |
| Vite 5 | MIT | Build tool |
| TypeScript | Apache 2.0 | Type safety |
| ONNX Runtime Web | MIT | Browser-side YuNet + YOLO inference (WebAssembly) |
| Tesseract.js | Apache 2.0 | On-device OCR for sensitive-text detection (self-hosted) |
pdf.js (pdfjs-dist) |
Apache 2.0 | PDF parsing + page rendering for document mode |
| jsPDF | MIT | Flattened (redacted) PDF export |
| OpenCV YuNet | Apache 2.0 | Server face detection (optional backend) |
| imagetracer.js | MIT | Raster → SVG vectorization |
| nodeca/pica | MIT | High-quality image resizing |
| smartcrop.js | MIT | Content-aware crop |
| img-halftone | MIT | Halftone canvas effect |
| mp4-muxer / webm-muxer | MIT | In-browser video (re)muxing on export |
| JSZip | MIT/GPL | ZIP archive creation |
| FileSaver.js | MIT | File download trigger |
| Electron | MIT | Desktop app shell |
Pull requests welcome! Please open an issue first for larger changes.
This project is part of Web3Privacy Now — building privacy tools that anyone can use and verify.
Project source code: MIT. Bundled third-party libraries, model weights, fonts and media retain their own licenses; see asset licensing status. The model/media provenance audit for store and F-Droid distribution is still in progress.