// the take
A webcam gag tool that detects your facial expression or hand pose and overlays a matching meme, then republishes the composited feed as a virtual camera so Zoom/Meet/Teams see it instead of your real face. It's for people who want to prank-react during video calls, not for anyone needing serious computer vision infrastructure.
The calibration approach in v2 is the genuinely smart part: instead of hardcoding blendshape thresholds (which fail because resting facial expression varies wildly person to person), it records your own neutral face for seven seconds and scores everything as standard deviations from that baseline, which is a real fix for a real problem with MediaPipe's raw output. Geometry is normalized by face-width ratios rather than pixel distances, so pose detection stays correct regardless of how close you sit to the camera. The pose-authoring workflow (POSES list, decide() branches, ARM hold counts) is actually documented well enough that someone could add a new gesture without reading the whole file. Graceful degradation on missing assets (red placeholder instead of a crash) is a small but correct choice.
The dependency chain is a house of cards: MediaPipe pinned to one exact version because newer builds crash on macOS, which forces NumPy 1.x, which forces pinning OpenCV to avoid it dragging in NumPy 2 — any transitive update from pip can break the whole thing, and the README's own warning about this is a sign it's already bitten people. Virtual camera support isn't native; it depends on the user separately installing and running OBS (or v4l2loopback on Linux) as a backend, which is an extra point of failure outside the project's control. It's effectively a single-file application (its_giving_v2.py) with no tests, so correctness rests entirely on manual threshold-tuning against a HUD rather than anything automated. Running real-time face and hand tracking that auto-fires without your control during an actual work meeting is a legitimate privacy/professionalism risk the README itself flags but doesn't mitigate beyond 'try it on someone who likes you first.'
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →