Face print · 512 numbers
Photos deleted
0.079
0.986 · you
0.12
0.05
0.21
0.985 · you
Blurred
2 copies of one image

Is it fake?

DetectorDINOv3 · face + whole image
0.81 ≥ 0.742
ReconstructionStable Diffusion autoencoder
rebuilds almost unchanged
C2PAsigned AI label in the file
no label found

Is it harmful?

IntimateNSFW score, the law's scope
0.93 ≥ 0.5
Contexthashtags, links, the poster
#deepfake, sales link, 6-day-old account
CopiesPDQ fingerprint
2 posts, 8 bits apart
Example case
Example case · 2 copiesHighest priority
Blurred · Reveal
Face match
0.986
Detector
0.81
Intimate
0.93
Context
flagged
Copies
2
This is meNot consensual
Confirmed · both copies
TAKE IT DOWN noticee-signed
Drafted by Factory Droid · test mode
48 h
Checks X again
Removed: resolved
Still up: FTC report
Six Eyes: the post is still up after 48 hours. Your FTC report is ready.
Text via Photon
1 · Enroll

Enroll once

3–5 selfies become a 512-number face print. Each selfie has to match the others, and the photos are deleted.

0 photos kept
InsightFaceArcFace
2 · Search X

Search X

X can't search by image, so we search the words around it: the user's name and handles, plus 12 deepfake terms. Then we face-match every image.

3 queries · $0.005 a post
X API v2only new posts
3 · Identify

Is it you?

ArcFace compares every face to the face print. Below 0.3 we drop it, so the heavy models only run on matches.

0.986 you · 0.079 a stranger
InsightFace buffalo_l
4 · Verify

Is it fake? Is it harmful?

Three checks for fake: our DINOv3 detector, reconstruction through Stable Diffusion's autoencoder, and C2PA labels. Three for harm: intimate content, the post's context, and copies. None decides alone.

cutoff 0.742 · 95% precision
DINOv3SD VAEC2PAPDQ
5 · You decide

You decide

Matches stay blurred until you choose to look. Nothing is filed until you confirm it's you and it isn't consensual.

1 confirm covers every copy
Next.js on Vercel
6 · Take down

Take it down

A case worker drafts the TAKE IT DOWN notice and checks X again at 48 hours. If the post is still up, it prepares the FTC report and texts you.

48 hours, by law
Factory DroidPhotontest mode
HackHarvard 2026

Six Eyes

Finds deepfakes of you on X, checks them, and takes them down when you say so.

Since May 2026, the TAKE IT DOWN Act gives platforms 48 hours to remove a reported intimate deepfake. Finding it and filing the notice is still up to the victim.

Behind the detector

The work behind the detector

23,000

training and validation images

3,000

fakes we generated by passing real photos through 3 autoencoders

100 + 50

consented team photos and hand-made Grok Imagine edits; one teammate held out entirely

5

detectors benchmarked on a teammate they never saw

~120

papers reviewed to pick the methods

0 of 49

of the held-out teammate's real photos falsely flagged by the model we shipped

Model choice

We shipped the model that generalized

Fine-tuning wins the benchmark (validation AUC) 0.95 0.90 0.85 frozen DINOv3 probe 0.887 0.909 0.926 0.928 epoch 1epoch 2epoch 3 …and loses on a person it never saw (Grok AUC) chance 0.5 00.8 0.662 0.686 0.695 0.514 probeprobe + Grok LoRALoRA + Grok ships 92% real flagged

Given 25 Grok images, the 2.36M-weight LoRA memorized "a teammate's phone photo" and flagged 92% of the held-out teammate's real photos. The 2,049-weight probe improved and flagged none, so it ships, and the user's confirmation makes the final call.

Six Eyes

Finds deepfakes of you on X, checks them, and takes them down when you say so.

23,000 training images · 150 team photos · 5 detectors tested · 0 of 49 real photos falsely flagged

FastAPI on ModalNext.js on VercelX API v2InsightFaceDINOv3Factory DroidPhotonGrok Imagine
0:30 / 0:30 Space plays · ← → step · Recording view hides these controls; Esc exits
1 / 10 → or Space next · ← back · N notes · F full screen · click a step to jump

Turn your phone sideways or use full screen for the best view.