Test results

What we measured on video face swap, and what it does not prove

These are our own numbers for Video FaceSwap Online 1.1.0, written by the people who make it. They are a first look, not an independent benchmark, and we have not compared the product with any other tool. The method is below so you can check it, and so you know how much weight to give it.

Updated . The live examples are on the product page.

The numbers

99.6%

Frames swapped

Average share of the chosen person's frames that were swapped, across 74 public clips. The rest are frames where the head turned too far and the swap eased back to the original on purpose.

0.65

Match to the reference

Mean similarity between the result's face and the reference photo, 74 clips. Before the swap the same clips scored 0.08. A score above 0.36 is the usual threshold for "same person".

0

Blank frames

Blank frames in the 74 swapped clips. Frame count, frame rate and size matched the input in every clip.

0.99

Person left alone

In a two-person test, the person who was not chosen stayed at 0.99 similarity to their own original.

~1 in 20

No usable face

About 1 in 20 public test clips had no usable face. The studio reported that instead of producing a bad result.

Five hard cases

Separate from the averages, we rendered five clips we chose because other face swaps tend to fail on them, in Fast mode. You can watch each one before and after on the product page.

  • A head turning from the front to a full profile
  • A hand covering part of the face, with glasses
  • Hands in the hair while the head tips back
  • A low-light shot with the head turning toward the camera
  • An arm and a phone crossing the face

What this test does not cover

  • Crowds and sports footage
  • Masks
  • Night footage beyond the clips shown
  • A head-to-head comparison with any other tool. We have not run one, so we do not claim to beat any of them.

Questions

How was the match to the reference measured?

With a face-recognition model, as the cosine similarity between the face in the result and the reference photo. 0.36 is the usual threshold for "same person". Front-facing reference photos give the strongest match.

What footage was used?

The figures were measured on 80 clips of a public talking-head dataset; 74 swapped clips are behind the averages. The five hard cases on the product page use free stock footage from Mixkit, rendered in Fast mode, and the faces swapped in belong to stock-photo models used only to test the product.

Is this a benchmark?

No. It is a first look at our own product on footage we chose or could get. It shows the hard cases can work; it does not show how it ranks against other tools, and results depend on your footage and reference photo.

Can I check it myself?

Yes. A new account gets free credits, and a short Fast-mode clip costs about 20 credits at 1080p, so you can run your own footage. Use a front-facing reference photo, and score the result with any face-recognition model you trust. If you test it against other tools and publish the result, tell us through the contact form.

Test it on your own footage

The honest check is a clip of yours that is hard. A free account is enough to try one. Only swap a face you have permission to use.

Open Video FaceSwap Online