InsightFace: what the toolkit contains and which datasets its README lists

This page is part of the face recognition section and describes the toolkit and its data rather than how to run it. The pipeline the toolkit implements, detect, align, embed and compare, is explained on the how face recognition works page.

01What is InsightFace?

InsightFace is a GitHub project maintained by the deepinsight organisation that collects training code, pretrained models and evaluation scripts for face detection, recognition and alignment. Its README states that "the master branch works with PyTorch 1.6+ and/or MXNet=1.6-1.8, with Python 3.x" (github.com/deepinsight/insightface). The recognition section lists implementations of ArcFace in MXNet, PyTorch, OneFlow and Paddle, Sub-center ArcFace (ECCV 2020) and Partial FC (CVPR 2022); the detection section lists RetinaFace, which the README describes as "a practical single-stage face detector".

The project also runs challenges. Its news log records a Masked Face Recognition Challenge and Workshop launched for ICCV 2021 on 5 June 2021, a leaderboard released on 11 October 2021, and an ongoing extension, MFR-Ongoing, launched on 30 November 2021. The same log records first place on the VISA track of NIST's FRVT 1:1 on 29 October 2021 using Partial FC, and a wild face anti-spoofing challenge at a CVPR 2023 workshop.

02Which methods does the toolkit implement?

ArcFace is the loss function most of the toolkit's recognition models are trained with. The paper adds an angular margin penalty of 0.5 to the target logit after normalising features and weights, with a feature scale of 64, and produces a 512-dimensional embedding from a 112 by 112 face crop aligned on five landmarks. With a ResNet100 backbone trained on MS1MV2, it reported an LFW verification accuracy of 99.83 percent and a true accept rate of 95.6 percent at a false accept rate of one in ten thousand on IJB-C (arXiv 1801.07698).

RetinaFace supplies the five landmarks that ArcFace's crop depends on. It is a single-stage detector that predicts a box, five facial landmarks and a dense 3D shape for every face; it reported an average precision of 91.4 percent on the WIDER FACE hard test set, and the paper states that with ArcFace it raised IJB-C verification to a true accept rate of 89.59 percent at a false accept rate of one in a million (arXiv 1905.00641). Partial FC (ICCV 2021 workshop, CVPR 2022) is the training method behind the Glint360K release; its paper is titled "Partial FC: Training 10 Million Identities on a Single Machine" (arXiv 2010.05222).

03Which datasets does the README list?

The recognition datasets README is a single page that lists every set the toolkit's models were trained or tested on, with identity and image counts for most of them and a download link for some. The table below reproduces the counts as the README prints them, rounded as it rounds them.

Source for every row: the InsightFace recognition datasets README (raw.githubusercontent.com/deepinsight/insightface/master/recognition/datasets/README.md), read 9 November 2026. The README's own citation for LFW is "Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments, 2007"; the LFW dataset page covers why that set saturated.

04What do the licence terms say?

The README splits the licence in two. "The code of InsightFace is released under the MIT License. There is no limitation for both academic and commercial usage." The data and the models trained on it are different: "The training data containing the annotation (and the models trained with these data) are available for non-commercial research purposes only." The same restriction applies to the models the Python package downloads automatically, and the README directs commercial licensing requests for its open-sourced recognition packages to an email address at insightface.ai.

That split matters because most of the training sets in the table were assembled from web photographs of named people under research terms of their own. MS-Celeb-1M, the parent of the three MS1M sets, is no longer served by its original host; the datasets and consent page in this section follows that history. A model trained on those sets inherits their terms whatever licence the code carries.

05How does the toolkit connect to the WebFace260M challenge?

The dataset README lists WebFace260M with a link to the download page on this domain, and the InsightFace team organised the ICCV 2021 Masked Face Recognition Challenge whose WebFace260M track report is summarised on the ICCV 2021 challenge page. That track built a new test set of 2,478 celebrities and 60,926 faces plus a real-world masked set; in its first phase 69 teams submitted 833 solutions, and 49 teams beat the organisers' baseline (arXiv 2108.07189).

06Why does the dataset list matter for reading a score?

A verification score is a property of a model and its training set together. The same ArcFace loss trained on CASIA-WebFace's 10K identities and on Glint360K's 360K identities produces two different models, and a published LFW figure applies to the one the paper trained. The README makes that visible by listing the sets side by side with their counts, which is rarer than it should be: many papers name a training set without its size.

No single benchmark score predicts how a face matcher performs on a population, a camera or a gallery it was not trained or tested on. That rule holds for a matcher at an airport gate as much as for a research model, and it has a mirror image in detection, where the types of deepfakes a detector was trained on bound what its score means, and where re-encoding on social media removes the cues it learned.