Blur Project

Prerequisites Commands

To set up the environment for the Blur project, follow these detailed steps:

1. Install Dependencies

Run the following commands to install the required dependencies:

cd blur/
uv sync

You also need to install ffmpeg :

• Mac : Install Homebrew :

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Then install ffmpeg:

brew install ffmpeg

• Ubuntu :

sudo apt update
sudo apt install ffmpeg

3. Accepted formats of files

Here are the accepted formats (others have not been tested) :

• photos : .jpg, .jpeg, .png
• videos : .mp4, .mov, .m4v

Execution

To blur the faces on your pictures or videos, follow these simple steps :

• Add your files (or folder) to the same repository of the main.py file.
• Execute the following line in your command line :

uv run python main.py name_of_file_or_folder.extension

• Wait for the process (the progress bar indicates how long it will take - usually really fast for photos but longer for the first one, a bit longer for videos)

If you give a name of folder, all the files in the folder will be processed. The files will be saved at the same place of the original ones.

Images

Images will be named the same as the original ones, with _blurred at the end of their name (and the same extension).

Videos

Once the execution is finished, you end up with 4 different files :

• filename_blurred.extension : this is the original video, with all faces blurred
• filename_blurred_anonymized.extension : this is the video with all faces blurred and voices anonymized
• filename_subtitled.extension : this is the video with all faces blurred and voices anonymized, with subtitles (generated automatically). If there is any mistake, you can change them by modifying the following file.
• filename_blurred.srt : this is the subtitles file. They are added by default to the _anonymized video. To rectify a subtitle mistake, you can open it in a text editor, rectify the mistake, then execute the following command in your command line :

ffmpeg -i filename_blurred_anonymized.extension -vf subtitles=filename_blurred.srt -c:a copy filename_subtitled.extension

Real time video blur

You can also blur your face in real time with your camera :

uv run python realtime_blur.py

You can switch between basic and pixelated blur by clicking on b or p.

For both pre-recorded or realtime blur, you can change the power of the pixelated blur by changing the pixel_size variable in the config file. For pixel_size, the higher it is, the better we will guess who is behind the blur. For blur_size, the higher it is, the less we will guess. Note that for blur_size, the number has to be odd.

Optional : Change the voice anonymization

To anonymize the voice, we created a non-inversible method, using fast fourier transformations. However, depending on the speaker's voice, anonymization may be more or less satisfactory. To modify the result, you can change the value of the shift_amount variable in the config file.

Optional : Other type of voice blur

Another type of voice blur is implemented in the script. It uses 2 AI libraries, first to proceed to Speech-To-Text (STT) with Whisper from OpenAI, then use this text for Text-To-Speech (TTS) with XTTS-v2 from Coqui. We will have a completely synthetized voice, which makes it completely unrecognizable. However, nowadays models are still not fully convincing, because you lose a lot of speech details and emotions. To use this format, just add --tts to the end of your command line :

uv run python main.py name_of_file_or_folder.extension --tts

Optional : Different blur types

The basic blur method is to pixelize the faces, but you can add the following blur argument to your command line to have a gaussian blur :

uv run python main.py name_of_file_or_folder.extension --blur

Optional : Upgrade performances

The script uses the YOLO model, which has different versions, and impacts the performances and the precision of the detection. Here, the yolov8n model is used, which is the smallest one, but you can download and use some different models by choosing these different models trained by lindevs :

Name	Model Size (MB)
YOLOv8n-Face	11.7
YOLOv8s-Face	42.7
YOLOv8m-Face	98.8
YOLOv8l-Face	166.6
YOLOv8x-Face	260.1

Make sure to add it to the utils folder, and to change the name in the variable model located in the config file.

Acknowledgements

This project uses the YOLOv8-Face model provided by lindevs/yolov8-face, which offers pre-trained YOLOv8 models specifically optimized for face detection.

The original models were trained on the WIDERFace dataset and are released under the AGPL-3.0 license.

The STT and TTS version uses the Whisper from OpenAI, and XTTS-v2 from Coqui libraries, actual open source SOTA models for STT and TTS performances.

LICENSE

This project is released under the GNU Affero General Public License v3.0 (AGPL-3.0), which requires that:

• If you redistribute this project or use it in a public-facing service (e.g., a web API or SaaS platform), you must also publish your full source code, including any modifications.
• If you wish to use this project in a closed-source commercial product, you must obtain a commercial license from Ultralytics

Commercial Use

You are free to use this project for commercial purposes as long as it remains open source and complies with the AGPL-3.0 requirements for YOLOv8. If you plan to keep your code private or integrate it into proprietary software, you must either replace YOLOv8 with an alternative model under a more permissive license (e.g., MIT or Apache 2.0), or purchase a commercial license from Ultralytics.

Please ensure you understand and comply with all relevant licenses before deploying or distributing this project.