WebFace260M and WebFace42M: the dataset, its licence, and where access lives

If you arrived from a paper, a README or a model card, the short version is this: the dataset exists, it is still cited and used, and the way to request it is through its authors' current pages, not through this domain. The rest of this page explains what the dataset is, how the cleaned set was made, what the licence restricted, and why the request route changed. It sits in this site's face recognition benchmarks section, next to the page on the ICCV 2021 masked face recognition challenge that used the same data.

01What WebFace260M is

The paper, "WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition" (arXiv 2103.04098, CVPR 2021), describes three things under one name.

The first is the raw collection: a name list of about 4 million identities and about 260 million face images downloaded from the internet. The authors call this the noisy set, because web-scraped faces come with wrong labels, duplicates and non-face images.

The second is the cleaned training set, WebFace42M: about 2 million identities and 42 million faces. It was produced by a pipeline the authors named CAST, Cleaning Automatically utilizing Self-Training, which trains a model on the noisy data, uses it to detect mislabelled faces, retrains, and repeats. The paper's claim for the result is that WebFace42M was, at publication, the largest public face recognition training set, and that it was meant "to close the data gap between academia and industry": companies trained on far more faces than any public set offered, and academic results could not be compared with theirs.

The third is an evaluation protocol, FRUITS (Face Recognition Under Inference Time conStraint), with a test set the authors assembled and three time budgets, 100, 500 and 1000 milliseconds, so that a model's accuracy is read next to how long it takes to run. The paper reports baselines across MobileNet, EfficientNet, ResNet, SENet, ResNeXt and RegNet families under those budgets.

02WebFace42M, the cleaned training set

The headline results in the paper are about what the cleaned set does for a model. Trained on WebFace42M, the authors report a relative 40% reduction in failure rate on IJB-C, a standard verification benchmark, compared with their baseline, and a third place among 430 entries on NIST's FRVT at the time of writing. They also report that a 10% subset, WebFace4M, outperformed the public training sets then in common use.

Those are the paper's own numbers, stated in the paper's own terms. Two things are worth keeping next to them. The IJB-C figure is a relative improvement at the operating points IJB-C uses, not an absolute accuracy, and the FRVT ranking is a snapshot of a leaderboard that NIST updates continuously, so it describes 2021, not today. How to read either kind of figure is covered on the face verification metrics page.

03The test protocol

Most face recognition benchmarks report accuracy alone. FRUITS adds a constraint: the recognition pipeline must fit within a fixed inference time budget, so accuracy is always read next to speed. A model that is a little more accurate but twice as slow may lose under a 100 millisecond budget and win under 1000. The paper's baselines are reported in those three bands.

The idea travelled. The ICCV 2021 masked face recognition challenge that the same group organised used FRUITS as its protocol, with a new test set of celebrities and a real-world masked test set, and its public report describes the time limit as part of what made the challenge practical rather than purely academic.

04Licence and access

WebFace260M was released for research use under a licence agreement. The version of this page that lived here from 2021 to 2025 set out the terms and asked applicants to email a signed copy of the agreement to an address at this domain. That process no longer exists. The domain changed hands in 2026, nobody at this site can receive, store or forward a dataset application, and nothing sent to this domain reaches the dataset's authors.

Where the dataset lives now: the authors' own channels. The paper and its arXiv page carry the current contact route, and the project's public code and evaluation moved to Codalab at the end of 2021 according to the news item that stood on the old home page ("December-31-2021: The WebFace260M Benchmark has been re-opened and moved to the new Codalab website"). Start from the paper's arXiv entry and follow the links the authors maintain.

What this page will not do: host the dataset, host the licence PDF, mirror any face image, or act as a contact point for access requests. The old licence file's address on this domain now redirects here, and the old image paths return a "gone" status permanently. The about page explains that rule and the rest of this site's editorial standards.

05Consent and the data gap debate

WebFace260M belongs to a generation of face datasets built by scraping public web images of named people. The paper frames the motivation in terms of a data gap between academic and industrial training sets, and the cleaning pipeline is engineering in service of that goal. The consent question sits alongside it: the people in a scraped set did not agree to be in a training corpus, and several earlier datasets built the same way were withdrawn or restricted after public criticism.

This site covers that history, and what it means for how datasets are described, on the face datasets and consent page. The point for a reader of this page is narrower: the research licence that governed WebFace260M was the mechanism that tied use to academic purposes and to named applicants, which is exactly why an unmonitored copy of its application process should not exist on a domain the authors no longer control.

06The same question for deepfake datasets

The face datasets used to train and test deepfake detectors have the same shape: thousands of faces, a licence, an access form, and a question about where the faces came from. FaceForensics++ was built from YouTube videos, Celeb-DF from videos of public figures, and the Deepfake Detection Challenge set from paid actors who agreed to have their likeness altered. The public deepfake datasets compared page puts those side by side, with each set's counts, methods, licence and access route taken from its paper or official page.