# Icop – open-source local AI NSFW filter for VLC (no cloud, no telemetry) > Source: > Published: 2026-07-17 04:19:16+00:00 An open-source, privacy-first VLC plugin for local AI sensitive-content detection and frame blocking with ONNX Runtime. Important icop is a best-effort content filter. Detection models can produce false positives and false negatives, so test your configuration before relying on it for child-safety or accessibility needs. icop is an open-source AI content filtering plugin for VLC media player. It uses local ONNX Runtime inference to classify buffered video frames before display, then shows the original frame or applies a black, blur, or warning treatment. Processing stays on your device; video frames are not uploaded to a service. The plugin targets Windows and Linux today, with experimental macOS build and installation support. It is designed for privacy-conscious playback, offline video filtering, and developers exploring ONNX-based computer vision in VLC. [Watch the full ICOP VLC walkthrough](/asayed18/icop/blob/main/demo.mp4) Plugin setup, blocking styles, muted playback, and the detection debug overlay. The public Git repository is source-only. Models, ONNX Runtime binaries, build trees, portable VLC copies, and release archives are downloaded or generated locally and are not committed. | Platform | Status | GPU inference | |---|---|---| | Windows x86_64 | Tested | CUDA (NVIDIA) / D3D12 (GPU fallback) | | Linux x86_64 | Tested | CUDA (NVIDIA) / MIGraphX (AMD) | | Linux ARM64 | Tested | CPU only | | macOS x86_64 | Tested | CPU / CoreML | | macOS ARM64 | Tested | CPU / CoreML | GPU runtimes are detected automatically at plugin load. Contributions that improve GPU support or validate the macOS path are welcome. - Buffers frames so playback does not intentionally outrun classification - Runs inference locally with ONNX Runtime - Supports multiple model profiles and automatic model-sized preprocessing - Provides black, blur, and warning block styles - Can mute audio while blocked output is shown - Supports precomputed decision maps for scan-ahead playback - Includes a D3D11 processing path for supported Windows hardware decoding - On Linux opaque VAAPI/VDPAU-style inputs, analyzes via VLC image conversion and drops blocked frames fail-closed when in-place masking is not available - Produces versioned, platform-specific packages with checksums and metadata - Includes unit, integration, and benchmark targets - Local and offline sensitive-content filtering in VLC media player - Best-effort parental controls and safer shared-screen playback - Privacy-first video moderation without cloud uploads - ONNX Runtime, CUDA, and D3D11 video-filter development - Cross-platform computer-vision research and prototyping - VLC decodes a video frame. - icop holds the frame before presentation. - The detector converts and resizes a sample for the selected model. - ONNX Runtime classifies the sample and applies the configured block window. - The plugin releases the original or masked frame to VLC. The central safety invariant is simple: a frame that requires analysis should not be shown before its decision is available. Ask an AI assistant to install icop using the [INSTALL.md](/asayed18/icop/blob/main/INSTALL.md) guide. Give it the file and it will handle downloading, copying, enabling the filter, and applying best-practice settings for your platform. Requirements: - CMake 3.16 or newer - Ninja or another supported CMake generator - A C compiler and a C++17 compiler - VLC 3.0 development files for the target platform Configure, build, and package: ``` make build make test make release ``` On Windows, use `mingw32-make` instead of `make` when that is the installed GNU Make command. Direct CMake commands remain available in [CONTRIBUTING.md](/asayed18/icop/blob/main/CONTRIBUTING.md). The package is written to `releases/v//` . See [INSTALL.md](/asayed18/icop/blob/main/INSTALL.md) for end-user installation steps. When upgrading from VLC iClean, remove `libnsfw_filter_plugin` and `nsfw_filter_core` files before regenerating the plugin cache. The runtime files are named `libicop_plugin` and `icop_core` . Build the host release, detect the installed VLC, copy the matching x86/x64/ARM64 payload, and regenerate VLC's plugin cache: ``` make install_plugin ``` Use `mingw32-make install_plugin` on Windows when GNU Make is installed under that name. The target selects the PowerShell installer on Windows and the POSIX installer on Linux/macOS. Both validate release checksums, remove legacy VLC iClean files, roll back failed copies, and regenerate the plugin cache. Windows requests administrator access when needed; Linux and macOS use `sudo` only for protected plugin directories. Preview the operation or select a specific installation directly: ``` powershell -ExecutionPolicy Bypass -File .\tools\install_icop_plugin.ps1 -WhatIf powershell -ExecutionPolicy Bypass -File .\tools\install_icop_plugin.ps1 -VlcRoot "C:\Program Files\VideoLAN\VLC" sh tools/install_icop_plugin.sh --dry-run sh tools/install_icop_plugin.sh --vlc-root /usr ``` Close VLC before installation. To stop it explicitly, pass `INSTALL_ARGS=-StopVlc` on Windows or `INSTALL_ARGS=--stop-vlc` on Linux/macOS. See [CONTRIBUTING.md](/asayed18/icop/blob/main/CONTRIBUTING.md) for lightweight Windows, Linux, and WSL build and test commands. See [docs/releasing.md](/asayed18/icop/blob/main/docs/releasing.md) for package versioning, checksums, and release verification. | Profile | Input | Upstream project | Upstream license | |---|---|---|---| `marqo` | 384x384 | | `adamcodd` [AdamCodd/vit-base-nsfw-detector](https://huggingface.co/AdamCodd/vit-base-nsfw-detector)`falconsai` [Falconsai/nsfw_image_detection](https://huggingface.co/Falconsai/nsfw_image_detection)`legacy` [iola1999/nsfw-detect-onnx](https://github.com/iola1999/nsfw-detect-onnx)Model and runtime files retain their upstream licenses. Review [THIRD_PARTY_NOTICES.md](/asayed18/icop/blob/main/THIRD_PARTY_NOTICES.md) before redistributing a binary package. The VLC module settings cover the most common choices: | Setting | Purpose | |---|---| | Model profile and path | Select a bundled profile or custom ONNX model | | ONNX provider | Choose CPU, CUDA, or automatic provider selection | | Processing backend | Choose portable CPU or supported Windows D3D11 processing | | Detection threshold | Control the model score that triggers blocking | | Analysis stride and buffer | Balance coverage, latency, and throughput | | Block style and padding | Choose the mask and extend detections around unsafe frames | | Audio muting | Mute playback while blocked frames are presented | | Decision map | Reuse precomputed blocked time ranges | See [INSTALL.md Step 4](/asayed18/icop/blob/main/INSTALL.md#step-4--apply-best-practice-settings) for best-practice values. The `marqo` profile at threshold `0.17` prioritises sensitivity and can produce more false positives than the default. Adjust threshold upward if false alarms are too frequent. No single profile is right for every video — validate with representative, legally shareable media. Build the plugin, detector core, tests, and benchmark: ``` cmake --build build-ninja --target icop_plugin icop_core icop_test icop_benchmark -j 8 ctest --test-dir build-ninja --output-on-failure ``` CI also supports a lightweight source build with model downloads disabled. The full integration suite requires locally downloaded ONNX models. Runtime testing requires a VLC installation or portable VLC tree compatible with the package. - Classification runs locally and does not require uploading video frames. - Optional debug dumps can write frame data to disk and should stay disabled for private media. - Logs may contain local paths and should be reviewed before sharing. - Detection quality depends on the model, threshold, source material, and processing configuration. - Public issues and tests must use synthetic or redistributable media, never private or explicit personal content. icop accepts sponsorship only through [GitHub Sponsors](https://github.com/sponsors/asayed18). Sponsorship does not affect issue priority, security handling, or project licensing. icop is licensed under [GPL-2.0-or-later](https://www.gnu.org/licenses/old-licenses/gpl-2.0.html). See [LICENSE](/asayed18/icop/blob/main/LICENSE) for the full license text. VLC and VideoLAN are trademarks of VideoLAN. This independent project is not affiliated with or endorsed by VideoLAN.