How OriginAI checks a video
It never judges a single picture. It follows the face through the whole video, looks at how it appears and how it moves, and only then gives you one answer.
What happens after you press the button
The same five steps run for every video. The technical name for each step is shown in grey, in case you need it.
Find the face
It finds the person's face in every frame and follows the same person through the whole video, so other faces in the shot don't get mixed in.
MTCNN face and landmark detection, identity tracking
Line it up
Each face is centred and resized to the same size, and very dark frames are brightened. This way, any differences come from the face itself, not the camera.
Landmark alignment to 224 × 224 pixels, low-light normalization
Cut it into short clips
The video is split into many short clips of 16 frames each, covering the whole video from start to finish.
16-frame clips, spaced to match the training data
Look at it two ways
One part studies how the face looks in each frame. Another studies how it moves from frame to frame. Fakes often look fine but move wrongly.
Swin Transformer (appearance) and Video Swin Transformer (motion)
Combine and decide
The two parts compare what they found, and the results from all the clips are combined into one score. If the score is above the detector's threshold, the video is called likely deepfake.
TLBG-HCA fusion, clip scores averaged into one likelihood
What it learned from
The detectors learned by studying thousands of real and fake videos from three well-known research collections.
FaceForensics++
Several kinds of face swaps and face edits, saved in both good and poor quality. Used for Veni HQ and Veni LQ.
Celeb-DF v2
Very convincing face swaps of many different people, with few obvious flaws. Used for Vidi.
DeeperForensics-1.0
Face swaps with real-world problems added, like blur, noise and odd colours. Used for Vici.
How we made sure the tests are fair
It's easy to look accurate by only testing on easy videos. These rules keep the hard ones in.
Equal real and fake
Tests use the same number of real and fake videos, so the detector can't score well by always guessing the same answer.
Unseen test videos
Videos are split 70% for learning, 15% for tuning and 15% for testing. No test video was ever seen while learning.
New kinds of fakes
Each detector is also tested on collections it never learned from, to see whether its skills carry over.
What we report
| Question | Test | Measures |
|---|---|---|
| How well does it do on familiar videos? | Same collection | ROC-AUC, F1, false alarms and missed fakes |
| Does poor quality hurt it? | Good and poor quality, reported separately | Same as above |
| Does it work on new kinds of fakes? | Other collections | ROC-AUC, false alarms and missed fakes |
See it on your own video
It takes about a minute, and your video is deleted right after. Pick the video type that matches yours and OriginAI does the rest.