English · العربية
Block Haram and NSFW images
A browser extension that blurs pictures of women, men or both, and blocks nudity and other NSFW images, as pages load. Everything runs on your own device.
- Choose who to blur: women, men or both, on every website.
- Blocks NSFW images: nudity and other explicit images are blurred, whoever is in them. This is on by default.
- Private by design: the AI models run inside your browser. No image, URL or browsing history is ever sent anywhere. There are no accounts, analytics or ads.
- Nothing flashes: images stay blurred while they load and are checked, then only the matching ones stay blurred.
- Finds images however a site shows them:
- plain images,
srcsetand<picture> - lazy-loaded images, carousels and CSS backgrounds
- shadow DOM, iframes and
blob:images - thumbnails down to 24 px
- plain images,
- You stay in control:
- three sensitivity levels
- a per-site on/off switch
- right-click Show this image or Blur this image to fix a single mistake
- English and Arabic, with a right-to-left layout in Arabic.
- Works in Chrome, Edge, Brave and other Chromium browsers, and in Firefox on desktop and Android.
Install directly from the Chrome Web Store (works in Chrome, Edge, Brave and other Chromium browsers).
- Go to Releases and download
noharam-firefox-<version>.zipfrom the latest release's Assets. - Open
about:debugging#/runtime/this-firefox, click Load Temporary Add-on…, and select the downloadednoharam-firefox-<version>.zip(or extract it and selectmanifest.json).- Permanent install (Firefox Developer Edition / Nightly): set
xpinstall.signatures.requiredtofalseinabout:config, then openabout:addons→ gear icon → Install Add-on From File… and selectnoharam-firefox-<version>.zip.
- Permanent install (Firefox Developer Edition / Nightly): set
- Click the NoHaram toolbar icon — if the popup says NoHaram needs access to websites, click Allow on all sites.
You need Node.js 22 or newer.
git clone https://github.com/mohamedsadiqaguenaou/NoHaram.git
cd NoHaram
npm install
npm run buildThen load the built folder. Use dist/, not src/:
| Browser | How |
|---|---|
| Chrome / Edge / Brave | Open chrome://extensions, turn on Developer mode, click Load unpacked and choose dist/chrome |
| Firefox (desktop) | Open about:debugging#/runtime/this-firefox, click Load Temporary Add-on and choose dist/firefox/manifest.json |
| Firefox for Android | npx web-ext run -t firefox-android -s dist/firefox with a USB-connected phone (guide) |
- Ready-made zips: prebuilt packages are attached to each release.
npm run zipbuilds them yourself, intodist/noharam-<browser>-<version>.zip. - Firefox: if the popup says NoHaram needs access to websites, click Allow on all sites.
- Skipping the test browser:
npm installalso downloads a test copy of Chrome, about 150 MB. If you only want to build, usePUPPETEER_SKIP_DOWNLOAD=1 npm install.
Click the toolbar icon:
- Site card: the current website, with a switch to turn filtering off for that site only.
- Live panel: how many images on the page were blurred and checked.
- Blur images of: Women, Men or both.
- Sensitivity: Relaxed, Balanced (the default) or Strict. See Accuracy.
- Turn on / Turn off: pauses or resumes filtering everywhere.
- Settings (the gear):
- Blur nudity and explicit images (on by default)
- Blur images until they are scanned (on by default)
- Blur people whose gender is unclear (on by default)
- Keep images blurred if they cannot be scanned (on by default)
- Language
Right-click any image and choose NoHaram: show this image or NoHaram: blur this image to override a mistake. The badge on the toolbar icon counts the blurred images on the current tab.
NoHaram has no server. The extension makes no network requests of its own except downloading the images the page is already showing, from the same addresses, so it can check them.
On your device, it keeps:
- your settings
- a cache of results, so an image already seen isn't checked twice. The cache stores a SHA-256 hash of each image address with its scores, in the extension's own storage.
Nothing is collected, logged or shared. See PRIVACY.md.
| Permission | Why |
|---|---|
| Access to all websites | To find the images on every page and download them for checking |
storage |
Settings, the result cache, and the badge counts |
contextMenus |
The right-click show / blur this image items |
scripting |
To take over tabs that were already open after the extension updates or restarts |
offscreen (Chrome only) |
To run the AI engine in a background page, since MV3 service workers can't start web workers |
flowchart LR
A["Web page<br/>content script finds images<br/>and keeps them blurred"] -- image address --> B["Background<br/>result cache"]
B -- not seen before --> C["Stage 1 · TinyCLIP-8M<br/>is anyone in the picture?"]
C -- nobody --> D["Clean<br/>~0.1 s"]
C -- maybe a person --> E["Stage 2 · TinyCLIP-39M<br/>woman, man or neither?<br/>+ NSFW check · nudity?"]
E --> F["Faces · YuNet<br/>+ face gender check"]
F --> G["Verdict"]
D --> B
G --> B
B -- blur or show --> A
- Two stages. A fast model first rejects pictures without people: logos, charts, products, scenery. That's most of the web. Only pictures that may contain a person get the slower, accurate check, which looks at the whole image and at up to two faces. A small NSFW classifier checks the same pictures for nudity.
- Zero-shot. The models are CLIP-style image-text models: they compare a picture with plain sentences such as "a photo of a woman", "a photo of a man" or "a logo". The sentences are in
scripts/models/prompts-stage2.json. Nothing was trained for this extension. - The page stays smooth:
- The page itself only tracks which images exist and applies the blur.
- Downloading, decoding and model inference all happen in the extension's own background workers, on WebAssembly (ONNX Runtime Web, SIMD, 1 thread per worker).
- Images on screen are checked first, then the next 300 px.
- While a finger is scrolling, only one image is checked at a time.
- Fail-safe: an image that can't be downloaded or checked stays blurred, unless you switch that setting off.
All models are existing, published models, converted to run in the browser. NoHaram did not train or fine-tune any of them.
| Role | Model | Size | License |
|---|---|---|---|
| Stage 1: anyone in the picture? | TinyCLIP ViT-8M/16, Microsoft | 9.4 MB | MIT |
| Stage 2: woman, man or neither | TinyCLIP ViT-39M/16, Microsoft | 40.6 MB | MIT |
| Face detection | YuNet, OpenCV Zoo | 0.23 MB | MIT |
| Face gender (second opinion) | age-gender-recognition-retail-0013, Intel | 1.4 MB | Apache-2.0 |
| Nudity / explicit images | nsfw-image-detection-384, Marqo (ViT-Tiny, full precision) | 22.5 MB | Apache-2.0 |
The built model files are committed in src/models/, so building the extension doesn't need Python. To rebuild them from the original downloads, run npm run models:prepare. That needs Python with torch, transformers, timm, onnx, onnxruntime and onnxsim. Changes and license texts are listed in THIRD_PARTY_NOTICES.md.
No image filter is 100% accurate; people disagree on blurry or partial images too. These numbers were measured in Chrome on 600 images the thresholds were never tuned on:
- 150 women-only, 150 men-only and 50 mixed photos, plus 100 photos without people, from Open Images V7
- 150 people-free web images from Wikimedia Commons
Blur women (default):
| Sensitivity | Women caught | Mixed photos caught | Men-only photos blurred by mistake | People-free photos blurred | People-free web images blurred |
|---|---|---|---|---|---|
| Relaxed | 93% | 64% | 7% | 1% | 0% |
| Balanced | 94% | 74% | 12% | 3% | 0% |
| Strict | 97% | 92% | 29% | 21% | 5% |
Blur men:
| Sensitivity | Men caught | Mixed photos caught | Women-only photos blurred by mistake | People-free photos blurred | People-free web images blurred |
|---|---|---|---|---|---|
| Relaxed | 95% | 66% | 8% | 2% | 0% |
| Balanced | 96% | 80% | 15% | 10% | 0% |
Typical mistakes involve long-haired men, children, masks and helmets, very small or background people, drawings and crowded groups.
Nudity: Marqo reports 98.6% accuracy for its NSFW classifier on its own test set of 20,000 images (photos, drawings, memes and AI images). On NoHaram's 1,488 normal labelled photos and graphics it flagged 1.7%: none of the photos without people, 0.2% of the graphics, 1.1% of the men-only photos and 7.5% of the women-only photos. About half of those flagged photos of women show swimwear or underwear. Its catch rate on real explicit content was not re-measured here.
Measured on an Intel i7-9750H laptop in Chrome, on the CPU:
| Time | |
|---|---|
| Image without people (most of the web) | ~0.1–0.2 s |
| Image with people (accurate check + faces + nudity check) | ~0.9–1.2 s per worker |
| Image seen before (cache) | ~1 ms |
| Engine start, first page after the browser starts | ~5–8 s |
- Firefox is about 1.3× slower.
- While images wait, they stay blurred, so a page full of photos of people takes a few seconds to settle.
- Memory: each engine worker uses about 160 MB. NoHaram runs 1 worker on phones, 2 on most desktops, and 3 on machines with 8+ cores and 8 GB+ of memory.
- Videos,
<canvas>drawings and SVG artwork are not filtered. - Images smaller than 24 px (icons, emoji) are skipped.
- The nudity check only looks at pictures where the first stage sees a person, so for example a close-up without an obvious person can be missed.
- Inference runs on the CPU only. A GPU (WebGPU) path is possible but not built yet.
- Firefox for Android and Edge Mobile have not been tested on real devices.
- The extension is about 89 MB unpacked: 75 MB of models and 14 MB for the runtime.
npm run watch # rebuild dist/chrome on every change
npm run zip # build every browser and write store-ready zips
node test/e2e.mjs --images <dir> # end-to-end test (add --browser firefox, --select male|both)
node test/lifecycle.mjs --images <dir> # extension reloads, worker restarts, frozen tabs, back/forward cache
node test/flash.mjs --images <dir> # no picture may ever appear unblurred before its check
node test/audit-sites.mjs --out <dir> <url>... # scroll real websites and save every image's verdict<
6B8D
/span>The tests need a folder of photos with a labels.json that maps each file name to "female", "male", "both" or "none". No photos are included in this repository; npm run models:evalset downloads a suitable set.
After changing a model or the prompts, recalibrate the thresholds in src/models/config.json:
npm run models:evalset # evaluation photos and graphics, about 90 MB
npm run build:chrome
npm run models:scores
npm run models:calibrate -- --writeSee CONTRIBUTING.md before opening a pull request. Bugs, wrong blurs and ideas go to the issue tracker.
NoHaram is released under the MIT License. The bundled models, ONNX Runtime and one icon keep their own licenses (MIT, Apache-2.0 and ISC), listed in THIRD_PARTY_NOTICES.md.
- TinyCLIP and ONNX Runtime by Microsoft
- YuNet by Shiqi Yu and the OpenCV team
- OpenVINO Open Model Zoo by Intel
- Lucide for the settings icon
- Open Images and Wikimedia Commons for the evaluation images