Safi — POC

This is a proof-of-concept built with AI assistance (under SDE supervision and guidance, without exhaustive manual review of every output). The goal is to validate that the idea is technically feasible and see what the cleaned result looks like — not to ship production-ready code.

Process a YouTube video to make it suitable for a Muslim audience:

• Music removed — Demucs extracts the vocals-only stem; instrumentals are discarded.
• Women blurred — InsightFace detects female faces; every flagged second gets a whole-frame blur.
• Sexual content blurred — NudeNet v3 catches any exposed or suggestively-covered private area.

Note on blur scope: In this POC the entire frame is blurred (fully opaque) whenever a flag is triggered. A future real product would blur only the specific detected region.

---

Prerequisites

• Python 3.10+
• ffmpeg on your PATH (used for audio extraction, frame extraction, and encoding)
  • Windows: winget install ffmpeg or download from https://ffmpeg.org
  • Linux/macOS: sudo apt install ffmpeg / brew install ffmpeg

---

Local setup

git clone <this-repo>
cd safi-poc

# Create and activate a virtual environment (recommended)
python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate

# Install Python dependencies
pip install -r requirements.txt

# Run
python safi.py "https://youtube.com/watch?v=VIDEO_ID"

Note: InsightFace downloads ~500 MB of model weights to ~/.insightface/ on the first run. NudeNet also downloads its own weights automatically. Both are cached for subsequent runs.

---

Google Colab setup

Paste this into a Colab cell to install everything:

# System dependency
!apt-get install -y ffmpeg

# Python packages
!pip install yt-dlp demucs librosa numpy nudenet insightface onnxruntime opencv-python

# Clone the repo (or upload safi.py + steps/ manually)
!git clone <this-repo>
%cd safi-poc

Then run in a second cell:

!python safi.py "https://youtube.com/watch?v=VIDEO_ID"

---

Performance expectations

Your CPU may run at 100% throughout the entire pipeline.

Technical limitation: This POC runs on CPU only. GPU acceleration is not wired up in the current implementation. Demucs, InsightFace, and NudeNet all support CUDA — GPU support is a straightforward next step for a real product.

Measured on a 1-minute video with a high-end CPU:

Step	Time	Share
Download	3.6 s	1%
Music removal (Demucs)	21 s	8%
Frame extraction	0.7 s	—
Model loading	2.1 s	1%
Frame analysis (InsightFace + NudeNet)	1 m 20 s	30%
Rendering (ffmpeg)	2 m 39 s	59%
Total	4 m 28 s

On another video with 1:40-min:

Step	Time	Share
Download	3.1 s	1%
Music removal (Demucs)	35 s	10%
Frame extraction	2.0 s	1%
Model loading	2.2 s	1%
Frame analysis (InsightFace + NudeNet)	2 m 37 s	44%
Rendering (ffmpeg)	2 m 36 s	44%
Total	5 m 56 s

Rendering dominates because ffmpeg re-encodes the full video with the partial-opacity blur filter. Longer videos scale roughly linearly.

---

Output files

File	Contents
output/safi_<id>_clean.mp4	Music removed + women blurred + sexual content blurred
output/safi_<id>_sexual_only.mp4	Music removed + only sexual content blurred (for evaluating NudeNet in isolation)
tmp/<id>/annotations.json	Full JSON record of every flagged second and its detections
tmp/<id>/frames/	One JPEG per second of the original video

tmp/ is intentionally preserved for inspection — delete it manually when done.

---

How it works

YouTube URL
    │
    ▼
[1] yt-dlp          → best-quality MP4
    │
    ▼
[2] Demucs          → htdemucs --two-stems vocals → vocals.mp3 (music gone)
    │
    ▼
[3] ffmpeg          → 1 JPEG per second
    │
    ▼
[4] For each frame:
    │
    ├─ Stage 1: InsightFace buffalo_l
    │       female face found + blur_women ON?  → flag second, skip Stage 2
    │
    └─ Stage 2: NudeNet v3  (only if Stage 1 didn't flag)
            explicit class ≥ 0.45?  → flag second
    │
    ▼
[5] annotations.json  (flagged seconds + bounding boxes)
    │
    ▼
[6] ffmpeg render ×2:
    clean.mp4         original video + vocals.mp3 + blur on (women ∪ sexual) seconds
    sexual_only.mp4   original video + vocals.mp3 + blur on sexual seconds only

Music removal

Music separation is based on Demucs (htdemucs model, --two-stems vocals). The approach was informed by studying the NoMusic project, which tackles the same problem and provided useful implementation reference.

---

Configuration

All tunable values live at the top of their respective modules:

Constant	File	Default	Purpose
BLUR_WOMEN	safi.py	True	Blur frames with female faces
BLUR_MEN	safi.py	False	Blur frames with male faces
DEMUCS_MODEL	steps/audio.py	htdemucs	Demucs model name
FPS	steps/frames.py	1	Frames extracted per second
CONFIDENCE_THRESHOLD	steps/detect_gender.py	0.5	InsightFace minimum face score
SCORE_THRESHOLD	steps/detect_sexual.py	0.45	NudeNet minimum detection score
BLUR_CLASSES	steps/detect_sexual.py	(11 classes)	NudeNet classes that trigger blur
BLUR_LR / BLUR_LP	steps/encoder.py	10 / 10	ffmpeg boxblur radius / iterations