Blurrinator

Local Django web app for privacy redaction in YouTube videos. Originals never leave the workstation. Licensed MIT.

Current slice: project upload, SHA-256 + FFprobe metadata, model registry, allowlist scaffold, in-DB job queue, completed face analysis, manual draft review overrides, and final redaction export.

Video

What It Uses

Blurrinator currently uses OpenCV-based face analysis, not Ultralytics YOLO.

• Django for the web app and admin-facing workflows.
• Django-Q2 for the local database-backed job queue.
• FFmpeg / FFprobe for video metadata, transcoding and browser-compatible exports.
• OpenCV with CUDA for YuNet face detection, SFace face embeddings and optional GPU blur operations.
• OpenCV Zoo YuNet weights registered as MIT licensed model artifacts.
• OpenCV Zoo SFace weights registered as Apache-2.0 licensed model artifacts.
• Pillow and NumPy for supporting image and array operations.

The model registry stores license, provenance, file path and SHA-256 hash for each downloaded model file.

Setup

Requires Python 3.12+ and ffmpeg / ffprobe on PATH. GPU (NVIDIA + CUDA) is required for YuNet/SFace inference. cv2 must come from a CUDA-enabled OpenCV build prepared on the host; installing stock opencv-python is not sufficient.

python -m venv venv
# Make sure a CUDA-enabled OpenCV Python build is already installed in this venv.
venv/bin/pip install -e .
venv/bin/python manage.py migrate
venv/bin/python manage.py createsuperuser

Run

Two processes: web and worker. Open a terminal each:

venv/bin/python manage.py runserver 0.0.0.0:8000
venv/bin/python manage.py qcluster

Open http://localhost:8000/.

Configuration

Environment variables (all optional):

Variable	Default	Purpose
BLURRINATOR_SECRET_KEY	dev placeholder	Set before exposing publicly.
BLURRINATOR_DEBUG	1	Set to 0 for production-style serving.
BLURRINATOR_ALLOWED_HOSTS	*	Comma-separated host list.
BLURRINATOR_STORAGE_ROOT	./storage	Originals, exports and analysis artifacts.
BLURRINATOR_MAX_UPLOAD_BYTES	85899345920 (80 GiB)	Hard upload ceiling.
BLURRINATOR_FFPROBE_BINARY	ffprobe	Override if not on PATH.
BLURRINATOR_FFMPEG_BINARY	ffmpeg	Override if not on PATH.
BLURRINATOR_REQUIRE_GPU_PIPELINE	1	Refuse analysis/export unless runtime has CUDA-ready OpenCV for GPU inference.

GPU requirement

This project now fails fast when the runtime is not actually ready for GPU inference. Having an NVIDIA card alone is not enough:

• OpenCV must expose CUDA devices (cv2.cuda.getCudaEnabledDeviceCount() > 0)

FFmpeg hardware decode is optional in the current pipeline. Video decode can stay on CPU; the critical part is that YuNet/SFace inference runs through CUDA-enabled OpenCV.

Current limitation:

• Face analysis runs on OpenCV DNN CUDA when available.
• Final blur compositing uses GPU only when the OpenCV Python binding exposes cv2.cuda.resize and cv2.cuda.createGaussianFilter.
• The export job log now reports which blur backend is actually used.

Rebuild OpenCV For GPU Blur

The local issue was caused by an OpenCV BUILD_LIST that included dnn and cudev, but omitted the CUDA image modules needed for blur. Rebuild with:

./scripts/rebuild_opencv_cuda.sh

That script configures OpenCV from the checked-out sources under .build/ with:

• BUILD_LIST=core,imgproc,imgcodecs,videoio,dnn,objdetect,python3,cudev,cudaarithm,cudafilters,cudaimgproc,cudawarping
• WITH_CUDA=ON
• WITH_CUDNN=ON
• OPENCV_DNN_CUDA=ON
• WITH_GTK=OFF, WITH_GSTREAMER=OFF, WITH_QT=OFF, WITH_OPENGL=OFF to avoid GTK/GIO conflicts with Homebrew Python
• CUDA_ARCH_BIN=9.0 and CUDA_ARCH_PTX=9.0 by default, so CUDA 12.0 can JIT PTX on newer Blackwell GPUs even though sm_120 is not supported by that toolkit yet

The script also stages a repo-local cv2 package. This is intentional: the venv already contains a root-owned site-packages/cv2, so the project shadows it from the repo root instead of trying to overwrite it.

To validate the result without rebuilding:

./venv/bin/python scripts/validate_opencv_cuda.py

Layout

blurrinator/        Django project (settings, urls, context)
projects/           Project, SourceVideo, AnalysisJob, ExportJob + pipeline services
registry/           Model registry (license + hash bookkeeping)
allowlist/          AllowedPerson + EnrollmentImage
templates/          Hand-rolled UI templates
static/css/app.css  Single CSS file, no build step
storage/            Local artifact tree (gitignored)

Links

• Author: https://github.com/SmolinskiP
• Support: https://buymeacoffee.com/smolinskip
• Ashes Fireshow: https://ashes.pl - a fire arts collective creating precise, emotional shows built around light, movement and story. Ashes also runs workshops and performances for children and young people facing difficult circumstances.
• Daily Stoic PL: https://dailystoic.pl - a newsletter that sends Polish daily Stoic reflections inspired by Ryan Holiday's "The Daily Stoic".

License

Blurrinator is MIT licensed. The current codebase does not include Ultralytics YOLO or other AGPL-only runtime dependencies; third-party model artifacts keep their own licenses in the model registry.