Face datasets and consent: MS-Celeb-1M, MegaFace, WebFace260M and the Clearview cases

This page is part of the face recognition section. This domain's own history is part of its subject: the WebFace260M dataset was distributed from here under a research licence from 2021 to 2025, and the account below describes that model alongside the others. This site is not the WebFace260M project and is not affiliated with its authors, Tsinghua University or XForwardAI.

01Where the faces in face datasets come from

Deep face recognition needs millions of labelled faces, and the only place millions of labelled faces exist is the web, where photographs of named people sit beside their names. The standard recipe, from 2014 onward, was to take a list of names, search for images of each, download them, clean the results with a model, and release the result for research. The people in the photographs were not asked, because asking a million people is not possible and because the images were public. The licences that accompanied the sets allowed research use and forbade commercial use, which is a statement about the downloader, not about the subjects. Everything that followed, the withdrawals, the regulatory cases and the consent clauses in newer dataset papers, is a reckoning with that recipe.

02The withdrawn datasets

MS-Celeb-1M was released by Microsoft in 2016 as a benchmark to "recognize one million celebrities from their face images", with a training set "which contains 10M images in version 1" described by its authors as the largest publicly available (arXiv 1607.08221). The dataset's website, msceleb.org, no longer serves it; its removal in 2019 followed reporting by the Financial Times and others on who was in it and who was using it, as documented by the Exposing.ai project (exposing.ai/msceleb). Copies and derivatives, including the cleaned MS1M variants listed in the InsightFace dataset README, continued to circulate and to train models for years afterwards.

MegaFace was released by the University of Washington in 2016 as an identification benchmark with a gallery of "1 Million Photos, 690,572 Unique Users" as distractors and a training set of "4.7 Million Photos, 672,057 Unique Identities" (megaface.cs.washington.edu). Its page now states that "the MegaFace challenge has concluded, reaching a benchmark performance of over 99%" and that "MegaFace is being decommissioned and MegaFace data are no longer being distributed". The stated reason is that its goals were met.

The pattern in both cases is the same: a dataset assembled from public photographs without consent, distributed under a research licence, used far beyond the academic groups it was meant for, and withdrawn by the institution once the people in it became the story. Withdrawal does not recall the copies.

03WebFace260M's licence model

WebFace260M (CVPR 2021) collected a name list of about 4 million identities and about 260 million face images from the internet, cleaned 42 million of them into WebFace42M, and released the result to researchers under a licence agreement that applicants signed and emailed to the project (arXiv 2103.04098). That is the same model as its predecessors, with the application step as the control: a named applicant, a stated purpose, a document. From 2021 to 2025 this domain was where that process ran. It no longer does; the dataset page explains where the project lives now and why no copy of the application process should exist on a domain the authors do not control. The licence governs the downloader. It says nothing about the four million people in the name list, who were not asked.

04The Clearview AI cases

The same recipe at commercial scale produced the clearest regulatory answers. Clearview AI built a face search product on a database its regulators describe as holding more than three billion images scraped from the web. On 3 November 2021 the Australian Information Commissioner determined that the company had breached Australians' privacy by scraping their biometric information from the web and ordered it to stop collecting images of people in Australia and to destroy those it held (OAIC). On 26 May 2022 the UK Information Commissioner issued a monetary penalty notice of 7.5 million pounds and ordered UK residents' data deleted (ICO). On 17 October 2022 the French CNIL imposed an administrative fine of 20 million euros (CNIL deliberation SAN-2022-019). The UK penalty was later contested before a tribunal; this page does not state the outcome because the decision could not be read from a primary source at the time of checking.

The cases are about a company, not about research datasets, and the law applied differs by country. What they establish is that collecting faces from the web without the subjects' knowledge is not, in several jurisdictions, made lawful by the images being public.

05What a dataset card should state

A face dataset's documentation can be judged on four points, whatever its size. The licence: who may use it and for what. The consent basis: whether the people shown agreed, and if not, where the images came from. The takedown route: how a person can ask to be removed, and whether removal propagates to copies already distributed. And the current status: whether the set is still distributed and by whom. The datasets table on this site records the second and fourth for every deepfake set it lists, in a column of their own, because the question does not go away when the faces are manipulated rather than recognised. The demographic differentials page is the other reason a card should say where the faces came from: a set gathered by search is balanced however search balanced it.

06Deepfake datasets have the same question

The detection datasets inherited the recipe and, in places, improved on it. FaceForensics++ took its 1,000 source videos from YouTube and Celeb-DF its 590 from videos of public figures on the same platform, neither with the subjects' consent; both offer takedown on request. The Deepfake Detection Challenge set went the other way: over 100,000 clips from 3,426 paid actors, and "all recorded subjects agreed to participate in and have their likenesses modified" (arXiv 2006.07397). DeeperForensics-1.0 and the Google and Jigsaw Deep Fake Detection set also used consenting actors. The in-the-wild sets, built from fakes that circulated, cannot have consent at all, since the people shown did not agree even to the fake. The checking a face video page is the reader's side of that last fact.

The rules that apply to face data in one large jurisdiction are read from their official texts on the face recognition in China page.