Is this face real?

Check a photo or short video for signs of face swapping. You get a clear verdict, a score, and a view of what the model examined.

Best with a single, clear, front-facing face. Files are deleted automatically within an hour.

Your result will appear here.

How it works

1Detect the face

The largest face is located and cropped, matching how the models were trained.

2Run two checks

A seam detector looks for the boundary left when one face is blended onto another. A whole-face model, built on a large vision transformer, assesses the face as a whole.

3Combine the evidence

A cross-attention model weighs both checks for each image, focusing on the regions that matter most.

Tested performance

97%on FaceForensics++, the benchmark it was trained on
93%on Celeb-DF, celebrity deepfake videos it had never seen
95%on DFD, another set of unseen deepfake videos
80%on DFDC, the hardest unseen benchmark

Scores are AUC: the probability that a randomly chosen fake is rated more suspicious than a randomly chosen real sample, where 50% is chance. Three of the four benchmarks were not used in training. On DFDC, the combined model did not improve on the whole-face check alone.

Limitations

  • Fully AI-generated faces. The models were trained on face swaps and facial reenactment, not faces generated from scratch, so results on these are unreliable.
  • Multiple or small faces. Only the largest face in the frame is analysed.
  • Heavy compression or filters. These can mask genuine traces or introduce misleading ones.
  • Not for decisions about individuals. This is a research tool, not an identity verification service.