InsightFace is a comprehensive open-source 2D and 3D face analysis toolbox primarily built on PyTorch and MXNet, maintained by Jia Guo and Jiankang Deng. The project implements state-of-the-art algorithms for face recognition, face detection, face alignment, and related facial analysis tasks, optimized for both training and deployment scenarios. The codebase is released under the MIT License, though training data and models trained with annotated datasets are restricted to non-commercial research purposes.
The repository encompasses multiple specialized modules covering the full spectrum of face analysis. The face recognition module includes implementations of ArcFace for both MXNet and PyTorch, SubCenter ArcFace, Partial FC, VPL, and framework-specific versions for OneFlow and PaddlePaddle. The project provides commonly used network backbones such as IResNet, MobilefaceNet, MobileNet, InceptionResNet_v2, and DenseNet. Training datasets include cleaned MS1M, VGG2, and CASIA-Webface datasets packaged in MXNet binary format, with standard evaluation pipelines for IJB and Megaface benchmarks. The face detection module is supported by implementations like SCRFD, which was accepted at ICLR-2022. Additional capabilities include face alignment and 3D face reconstruction, with the project achieving first place on the NIST-FRVT 1:1 VISA track and ranking first in the ECCV-2022 monocular 3D face reconstruction challenge.
Recent developments reflect a shift toward accessibility and practical deployment. InsightFace 1.0 introduced the InsightFace Evaluation Studio, a cross-platform GUI demo for Windows, macOS, and Linux that supports local face recognition, enterprise model evaluation with reporting, and face swap trials. The default Python package no longer requires C++/Cython compilation for the optional face3d extension, reducing local setup complexity. The project has also expanded into face swap technology, with single-line code implementation available in the Python package as of version 0.7, and commercial applications like the Picsi.ai macOS and iOS live face swap app launched in November 2025.
According to GitGenius activity tracking across 250 items, the repository shows a median issue and pull request response latency of 85.1 hours with a mean of 2412.2 hours, indicating variable response times. The primary maintainer nttstar has logged 166 tracked events, with anxiangsir contributing 35 events and tunmx contributing 14 events. The most active issue label tracked is Example with 3 occurrences. The repository shares overlapping contributors with paddlepaddle/paddleocr, sveltejs/svelte, and opencv/opencv, reflecting its integration within the broader computer vision and deep learning ecosystem.
The project maintains active development with regular updates documented in changelogs spanning from 2021 through 2026. Notable achievements include acceptance of Partial FC at CVPR-2022, SCRFD at ICLR-2022, and various papers on facial reflectance reconstruction and gaze estimation. The repository supports multiple challenges including the MFR-Ongoing challenge, ICCV21 Masked Face Recognition Challenge, and CVPR23 face anti-spoofing challenge. Web demos for face localization, recognition, and swapping have been deployed, and the project provides extensive third-party reimplementations across frameworks including TensorFlow, Caffe, TensorRT, ONNX Runtime, and mobile inference engines.
