InsightFace Review

InsightFace is a research-grade face analysis project with training code, datasets and evaluation pipelines for face recognition (ArcFace, SubCenter ArcFace, Partial FC, VPL), detection (RetinaFace, SCRFD) and alignment. The insightface Python package (pip install -U insightface) wraps ready-made model packages: FaceAnalysis() downloads buffalo_l by default for detection, landmarks and embeddings, and newer raccoon_s and raccoon_l packages pick CoreML, CUDA or CPU automatically. Version 2.0 (September 2026) added PrivateFrame for local face blurring in video and an optional RGB liveness add-on; 2.1 (October 2026) added PersonAnalysis with person detection and body re-identification.

InsightFace Server is a separate Docker container (Linux x86_64, CPU or NVIDIA GPU) with a web UI, 31 REST operations, SQLite storage and exact 1:N search. The project describes it as a self-hosted alternative to AWS Rekognition, not an API-compatible one, and says one RTX 5090 held 58.9 million INT8 face vectors. Original uploads are not kept by default.

InsightFace has two licences, and the difference is the main thing to know before using it in a product.

The README routes commercial requests for open-source recognition models such as buffalo_l to recognition-oss-pack@insightface.ai, and inswapper face-swap models and the InspireFace SDK to contact@insightface.ai. The server's model installer asks you to --accept-license, and each model directory carries its own MODEL.LICENSE.

Running costs are your own hardware: the server needs Linux x86_64 with Docker, plus a supported NVIDIA GPU, driver and Container Toolkit for CUDA.

Prices checked on the vendor's pricing page on November 12, 2026. See the InsightFace licence section and model licensing page.

From the repository, Python package and server READMEs:

InsightFace suits researchers and ML engineers who want state-of-the-art recognition and detection models, training code and benchmarks, and companies prepared to buy a commercial model licence for a self-hosted, high-volume face search server.

If you need something free for commercial use today, the model licence rules InsightFace's pretrained packages out until you buy one. CompreFace (Apache 2.0, default FaceNet build) or DeepFace with a commercially licensed wrapped model are the usual paths, and Amazon Rekognition avoids model licensing by charging per call.

InsightFace is one of the most-starred face projects on GitHub (about 30,200 stars at the time of checking). The recurring developer complaint is licensing: in issue 2486 a maintainer confirmed that all open-source models from the repo are for non-commercial research only, and in issue 2587 several developers asked for buffalo_l pricing and were told to email from a business address. Downstream projects such as Facet now list InsightFace models separately because of this.