Meta AI · Muse Image

Some of these
are fake.

Real people and Meta Muse Image, mixed together and unlabelled. Your detector was built before this generator existed. Download the set and find out if it can still tell the difference.

How to test Download the 16 Questions

16 faces · Google Drive · no signup

The test

Three minutes.

01

Feed them in

Send all sixteen through the same pipeline you use in production. No warning, no special handling.

16 faces in

02

Watch what slips

Count the fakes your system waved through as real. One is enough to matter. In production, that one is an account.

some slip past

03

Close the gap

Caught them all? We would like to know how. Missed a few? Book a call and we will walk through what fixes it.

close the gap

When you find a gap

Sixteen faces found it.
Now close it.

The sample set shows you whether Muse Image gets through. The Deepfake Dataset Service keeps your detection current as every new generator ships, so the gap does not open again next quarter.

Fresh samples from new generators as they launch
Labelled, balanced, ready for training and evaluation
A short call first, to see if it fits what you are building

Book a first call

Tell us where to reach you. We will set up a 20-minute call to walk through your results and the service.

Before you run it

Questions, answered.

Sixteen faces. Some are real people, some are Meta Muse Image, and we are not telling you which. Standard image files, nothing to unpack or convert. It is a spot check, not a full benchmark. When you want data that keeps pace with every new generator, that is the Deepfake Dataset Service. Book a call to see it.

Nothing. The set runs entirely on your side. It does not phone home, report scores, or tell us how you did. Whatever you find stays with you until you decide to share it.

Straight from Meta Muse Image, the way an attacker would use it. We picked realistic outputs across different faces, ages, and settings. We did not tune them to beat any specific detector, and we did not pick easy ones to make a point. They are meant to look like real submissions, because that is what you will face.

Correct, and we are not pretending otherwise. Sixteen is enough to show you whether a gap exists, not to measure its exact size. If even one fake passes as real, you have learned the thing that matters. Sizing the gap properly is what the Deepfake Dataset Service is for, and a short call is the way in.

Yes. Send it to whoever runs your detection and verification work. It is meant to be tested by the people who own the pipeline.

Reply to the email or message that sent you here. We will get you in straight away.

Some of theseare fake.