About Face Benchmark

01What this site is

A reference on how faces are measured by machines. Each dataset page says how the data was collected, under what licence, and how to get it from its maintainers. Each result on the results page is copied from its source with the conditions it was measured under. The metric pages explain what the numbers mean and where they stop meaning anything. The site has two sections, deepfake detection and face recognition, and a home page that shows how a single result is read field by field.

The site exists because the numbers in this field travel badly. A detector's AUC gets quoted without its test set, a dataset's size gets quoted in the wrong unit, and a leaderboard snapshot gets quoted years after the leaderboard moved. Putting the conditions back beside each figure is the whole editorial job.

02This domain's history

From 2021 to 2025 this domain was the WebFace260M benchmark site, home of a large face recognition dataset and the masked face recognition challenge held at ICCV 2021. The domain changed hands in 2026.

The two legacy addresses, /challenge.html and /download.html, still answer because they are cited in papers, package pages and READMEs. They now describe the challenge and the dataset from their public papers and say plainly that nothing is downloadable here and that this site cannot receive dataset applications. Other addresses from the old site redirect to the nearest current page or return a permanent "gone" status, so a stale link never lands on an unrelated page.

03Who writes here

Pages are written and checked by The Face Benchmark editors, a small editorial group that reads the primary papers and official repositories directly. The byline is the group's, not an individual's, because every page passes through more than one pair of hands before it is published and after every correction.

• E1Every number is matched against its source before it is published, and again on the date shown beside it.
• E2We explain a result before we judge it, and we say when sources disagree.
• E3We do not accept payment or data from dataset owners, tool vendors or detector developers, and no page recommends a product.

04How we report numbers

Every figure on the site carries four fields. A number missing any of them is not printed.

The how to read a detection paper page applies the same four fields to papers that leave one out, and the metrics page explains why a value without them cannot be compared with anything.

05Corrections and contact

If a figure is wrong, send the page, the value you believe is correct, and where it appears in the source. Corrections that change a figure are dated on the page they change, and the home page lists the most recent checks. The contact page says what else we can and cannot help with.

06What we do not publish

• N1Photographs or renderings of faces, real or synthetic. Figures are diagrams drawn from sourced numbers.
• N2Verdicts on whether a particular person's video is real. The checking a face video page explains what a person can check instead.
• N3Results we cannot trace to a paper or an official leaderboard, and rankings of detectors or vendors.
• N4Dataset files, licence documents or mirrors of any kind.