Face Redaction

Face Redaction

Finds every face in a photo or video and hides it with pixelation, blur or a solid block while the rest of the image stays intact. For privacy teams handling CCTV, event or street footage.

Made by hamzaasyf on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Face Redaction finds faces in photos and videos and covers them, leaving the rest of the image as it was. It handles group photos, dense crowds and dark CCTV frames. When the detector loses a face in video, a tracker carries the cover over the gap, so the face does not reappear on blurry frames.

It comes as a web app and as a command-line tool.

Who it suits: People who must hide identities before footage or images are shared, stored or used for training.

Industry | Who uses it | How they use it | Example
Security | CCTV and body-camera operators | Cover faces before footage leaves the building | Redact a dark CCTV frame with the low-light option
Compliance | Privacy and legal teams | Answer data requests without exposing bystanders | Blur everyone except the person asking
Machine learning | Data teams | Anonymise image sets before training | Run a folder of photos through the command line tool
Media | Event and street videographers | Publish footage without showing faces | Pixelate faces in a crowd clip

You can skip it if you need to find a specific person, since it hides faces and does not identify them.

Where to use it:

• A web browser, on the Space's page at Hugging Face, with nothing to install
• Windows, macOS or Linux with Python, using the web interface
• A terminal, using the command-line tool

Use case | Who | What they get back
Redact a photo | Anyone sharing images | A copy with faces pixelated, blurred or blocked
Redact a video | Footage owners | A video with the same covering and the original audio
Check the result | Reviewers | Numbered boxes with scores, plus a clean copy without boxes
Count faces | Analysts | A JSON report of face boxes, or counts per track for video

How to set it up:

• Python on your computer
• The Python packages listed in requirements.txt
• The huggingface_hub package, which gives you the hf command

1. Install the huggingface_hub package.
2. Download the Space with the hf command.
3. Move into the downloaded folder.
4. Install the pinned Gradio version.
5. Install the packages from requirements.txt.
6. Start the app with Python.

pip install -U huggingface_hub
hf download hamzaasyf/face-redaction --repo-type space --local-dir face-redaction
cd face-redaction
pip install gradio==6.31.0
pip install -r requirements.txt
python app.py

How to use it: In the web interface you pick the photo or video tab, add a file and choose the style. From a terminal, you run the command-line tool with a file and an output name.

python cli.py photo.jpg -o redacted.png --style blur
python cli.py clip.mp4 -o redacted.mp4 --scales 1,2
python cli.py cctv.jpg -o redacted.png --scales 1,2,3 --low-light --threshold 0.5

Tool | What it does
app.py | Opens the web interface with photo and video tabs
cli.py | Redacts a photo or video from the command line
--style | Chooses how faces are covered, such as blur
--scales | Sets the zoom levels the detector scans
--low-light | Brightens dark frames for detection only
--threshold | Sets the detection cutoff, lower for hard footage

Example: You run the first command above on photo.jpg. It writes redacted.png with every detected face blurred and the rest of the picture unchanged.

For newcomers and experts: If you are new to this, open the Space and try one group photo with the default settings. Check that every face is covered, then try another style.

If you know the field, the detector is YuNet, run on overlapping tiles at several zoom levels so tiny faces are caught. Duplicate boxes are merged with non-maximum suppression. The tracker matches boxes across frames in two stages, like ByteTrack, and estimates each track's motion at constant velocity. Output video is encoded as H.264 and the sound is carried over.

Each frame is cut into tiles at 1x, 2x and 3x
YuNet detects faces on every tile
Duplicate boxes are merged
A tracker covers faces the detector misses
Faces are redacted and the output is saved

Good to know:

• License: MIT for the code, and the README says the same for the YuNet model
• Cost to run: none named
• Data it reaches: the photos and videos you give it, plus the files in the Space
• Not for: guaranteed redaction, since a face turned away, mostly hidden or tiny in the frame can still be missed, so review the output before you publish
• Limits: the online demo cuts each video to its first 20 seconds, and the README names a MAX_VIDEO_SECONDS setting for changing that

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: gradio
• License: mit

Links

• Space: https://huggingface.co/spaces/hamzaasyf/face-redaction
• Files: https://huggingface.co/spaces/hamzaasyf/face-redaction/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.