PureFrame

Watch any movie with your family. Without cutting a single second.

PureFrame applies smart, localized blurs over explicit visuals — no cuts, no audio edits, no streaming, no subscription.

Synthetic demo — left: original, right: PureFrame output. The flagged region is Gaussian-blurred and smoothly tracked as the figure crosses the frame; everything else is untouched. Regenerate with scripts/make_demo.py.

PureFrame is a local AI tool that finds explicit visuals in common video files — nudity, sexual activity, intense kissing — and applies a localized, smoothly-tracked blur over the flagged regions. No scene skipping. No audio cuts. No streaming, no cloud, no subscription. The full movie plays normally; you just don't see the parts you'd rather not.

Downloads

Pre-built binaries. Grab the latest release.

Desktop GUI (Tauri)

Native installer with the graphical UI:

• Windows — PureFrame_x64-setup.exe or .msi
• macOS (Apple Silicon) — PureFrame_aarch64.dmg
• macOS (Intel) — PureFrame_x64.dmg
• Linux — .AppImage, .deb, or .rpm

CLI / standalone (PyInstaller)

No Python needed:

• Windows — pureframe-windows-x86_64.zip
• macOS (Apple Silicon) — pureframe-macos-arm64.tar.gz
• Linux (x86_64) — pureframe-linux-x86_64.tar.gz

Intel mac users: no standalone PyInstaller build (GitHub-hosted macos-13 runners are EOL'd and perpetually backlogged). Use pip install pureframe or the Tauri PureFrame_x64.dmg from the latest release.

Extract anywhere, then run pureframe --help (Windows: pureframe.exe --help).

• Windows zip is fully self-contained — bundled ffmpeg.exe + ffprobe.exe, no install required.
• macOS / Linux tarballs require ffmpeg on PATH (brew install ffmpeg / apt install ffmpeg).

Code signing

Binaries are not code-signed (no paid Apple/Microsoft certs yet).

• Windows: SmartScreen warns "Windows protected your PC" → click More info → Run anyway.
• macOS: Gatekeeper blocks "from an unidentified developer" → right-click the .app / binary → Open → confirm. Or run xattr -dr com.apple.quarantine /path/to/PureFrame.app.
• Linux: no signing required.

Verify downloads against the SHA256SUMS.txt asset attached to each release:

# Linux / macOS
curl -LO https://github.com/MayonaiseLover/PureFrame/releases/latest/download/SHA256SUMS.txt
sha256sum -c --ignore-missing SHA256SUMS.txt

# Windows PowerShell
$expected = (Select-String -Path SHA256SUMS.txt -Pattern 'pureframe-windows-x86_64\.zip').Line.Split(' ')[0]
$actual   = (Get-FileHash pureframe-windows-x86_64.zip -Algorithm SHA256).Hash.ToLower()
if ($expected -eq $actual) { 'OK' } else { 'MISMATCH' }

Install

# From PyPI
pip install pureframe

# From source (development)
git clone https://github.com/MayonaiseLover/PureFrame.git
cd PureFrame
pip install -e ".[dev]"

Requirements: Python 3.11+, FFmpeg installed and on PATH. GPU recommended but not required. See Installation Guide for platform-specific instructions, GPU setup, and troubleshooting.

Quick Start

# One-shot: detect and blur in a single pass
pureframe process movie.mp4 --output movie_clean.mp4

# Or split it: generate a plan, review it, then apply
pureframe plan movie.mp4                              # → movie.censorplan.json
pureframe plan-edit movie.censorplan.json              # Review in your editor
pureframe plan-whitelist movie.censorplan.json 3       # Whitelist false positive
pureframe apply movie.mp4 movie.censorplan.json        # Render final output

# Preview flagged shots without watching the whole video
pureframe preview movie.censorplan.json                # → HTML contact sheet

The plan file is plain JSON — open it, review every flagged shot, whitelist anything you disagree with, then apply. Nothing renders until you say so.

Content-Type Profiles

Different content needs different detection settings:

# Live-action movies and TV (default)
pureframe process movie.mp4 --content-type live-action

# Animated content (higher thresholds to reduce false positives)
pureframe process cartoon.mp4 --content-type animation

# Anime (tuned for anime art styles)
pureframe process anime.mkv --content-type anime

# Dark/low-light scenes (increased sensitivity)
pureframe process horror.mp4 --content-type low-light

Strictness Levels

Control how aggressively PureFrame flags content:

pureframe process movie.mp4 --strictness low     # Minimal censoring
pureframe process movie.mp4 --strictness medium   # Balanced (default)
pureframe process movie.mp4 --strictness high     # Aggressive
pureframe process movie.mp4 --threshold 0.35      # Custom threshold

See Evaluation Report for threshold calibration guide.

Why PureFrame?

No scene skipping. Most "family-friendly" tools just fast-forward through flagged scenes. You lose dialog, plot, pacing. PureFrame applies a localized Gaussian blur tracked to bounding boxes — the scene plays normally, you just can't see what's behind the blur.

No cloud, no subscription. Everything runs on your machine. Your videos never leave your disk. Once the AI models download on first run (~400–500MB), PureFrame works fully offline. Zero telemetry.

Works on any local video file. VidAngel and ClearPlay only support a curated list of popular titles. PureFrame uses computer vision — it works on any MP4, MKV, AVI, or WebM you throw at it. Foreign films, indie movies, decades-old DVDs.

Audio-aware detection — only when it matters. An audio classifier runs alongside the visual pipeline to disambiguate ambiguous scenes. A verdict-preserving gate means it only runs on shots where the scene signal could actually change the outcome — most shots skip it entirely, which is a large part of why detection got an order of magnitude faster in the September 2026 speed offensive (details).

Review before rendering. The plan command generates a JSON file with every detection, bounding box, confidence score, and reasoning. Inspect it, whitelist false positives, or adjust thresholds before committing to the render.

How It Works

flowchart TD
    A[Input Video] --> B[Scene Detection]
    A --> C[Audio Extraction]
    B --> D[NudeNet Frame Analysis]
    B --> D2[CLIP Scene Classification]
    C --> E[PANNs Audio Classifier]
    D --> F[Confidence Fusion]
    D2 --> F
    E --> F
    F --> G[Censor Plan JSON]
    G --> H[Review & Edit]
    H --> I[FFmpeg Renderer]
    I --> J[Clean Output Video]

1. Scene detection splits the video into shots using adaptive threshold detection (PySceneDetect).
2. NudeNet analyzes sampled frames for nudity with localized bounding boxes.
3. CLIP provides scene-level semantic classification for sexual activity detection.
4. PANNs classifies audio events (moaning detection) for context disambiguation — behind a verdict-preserving gate, so it only runs on shots where the scene signal makes audio relevant.
5. A confidence fusion engine combines all signals with configurable per-category thresholds.
6. Results are written to a censor plan (.censorplan.json) — fully editable before rendering.
7. The renderer applies tracked bounding-box blurs frame-by-frame and re-encodes with FFmpeg.

Comparison

Performance

Measured with pureframe bench --duration 30 --reps 3 (medians) on the author's machine — i5-10400F (12 threads), RTX 3060, Pop!_OS — after the September 2026 speed offensive: seek-based frame extraction, resident ONNX sessions, lazy audio classification, pipelined decode/inference, int8 CPU quantization (on by default, --no-quant to disable), and per-profile encoder presets.

The offensive targets ~10–20 minutes for a 90-minute movie on CPU-only hardware. Honest caveats: this is a 30 s synthetic clip with 1–2 shots, not a movie — real films amortize per-shot work differently, and the CPU run's zero detections under-represents real content. Full tables, per-phase breakdowns, the "before" numbers, and the engineering narrative: docs/performance.md · BENCHMARKS.md.

Measure your own machine:

pureframe bench --duration 30 --reps 3 -o bench-report.json

Desktop App (Experimental)

PureFrame includes an experimental Tauri desktop GUI — dark theme, live job progress, and shot-level plan review:

Job queue	Plan editor

• ✅ File drag-and-drop queue
• ✅ Plan editor with color-coded timeline
• ✅ Shot-level thumbnail preview
• ✅ One-click whitelist/blacklist
• ✅ Hardware profile settings
• ✅ Detection sensitivity slider
• 🔜 Timeline scrubbing
• 🔜 Before/after preview

cd gui && npm install && npm run tauri dev

Known Limitations

PureFrame is honest about what it can and can't do. See KNOWN_LIMITATIONS.md for a full breakdown of false positive/negative categories, audio detection gaps, and rendering limitations.

Briefly:

• False positives happen on swimwear, skin-tone backgrounds, and stylized animation.
• Dark scenes reduce detection confidence. Use --content-type low-light.
• Not perfect. Some explicit content may slip through. PureFrame is a tool — not a replacement for parental judgment.

FAQ

Documentation

Acknowledgments

PureFrame builds on excellent open-source work: NudeNet for nudity detection, PySceneDetect for shot boundary detection, CLIP for scene understanding, PANNs for audio classification, FFmpeg for video I/O, and Tauri for the desktop GUI.

Contributing

Contributions welcome! See CONTRIBUTING.md. Look for issues labeled good first issue.

License

MIT