finds.dev

ageitgey/face_recognition

The world's simplest facial recognition api for Python and the command line

View on GitHub

★ 56,865PythonMITupdated Jun 2026

// the take

A thin, friendly Python wrapper around dlib's face detection and recognition models — three function calls (face_locations, face_encodings, compare_faces) get you from a raw image to a face match, plus CLI tools that do the same over a folder of photos with no code at all. Good fit for someone who wants face detection/recognition working in an afternoon without reading a paper on embeddings first; not the right starting point for anyone building a production identification system in 2026.

The API surface is genuinely minimal and well chosen — detect, landmark, encode, compare covers the whole pipeline and matches how you'd actually use it in a script. CLI tools (face_recognition, face_detection) work directly on a directory of images with comma-separated output, useful for quick batch jobs without writing Python. Batch and multiprocessing examples (webcam, video file, KNN/SVM classifiers) are included and cover the common real-world usage patterns, not just toy cases. Docker images and a docker-compose setup are provided specifically because the dlib build is painful, which is a practical acknowledgment of the actual friction point.

The core model is dlib's ResNet-based face encoder from around 2017 — the quoted 99.38% LFW accuracy is a stale, saturated benchmark, not a meaningful comparison to current ArcFace/InsightFace-style embeddings; this library isn't getting more accurate over time, it's wrapping a frozen model. The project's own README documents real fairness problems: it explicitly says accuracy is lower on children and varies by ethnic group, which is a serious liability for anyone using this to identify real people rather than toy demos. Installation requires compiling dlib with C++ bindings via cmake; Windows has no official support, only community-posted workarounds, which undercuts the 'simplest API' pitch for a lot of potential users. The CNN detector (the more accurate option) needs CUDA and a GPU to be usable at all — on CPU you're stuck with the weaker HOG-based detector, so the accuracy/hardware tradeoff isn't optional for serious use.

We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →

// at a glance

Stars
56,865
Forks
13,669
Language
Python
License
MIT
Last commit
25 Jun 2026

// nearby

If face_recognition isn't quite it.

All repos →