Tell real footage from Sora 2 and Veo 3.

Authini samples 15 frames from a clip, degrades them the way social media does, and tells you whether they look like real camera footage, Sora 2 or Veo 3.

Check a video

Frame 015 of 144 sampledFrame 143

84.4%Veo 3

Real: 7.6%
Sora 2: 8.0%
Veo 3: 84.4%

This clip most closely matches Veo 3.

The file was made with Veo 3. The model got it right.

Pick a file and see what the model said

Each one is a real run of the model. No sign-in.

What happens to your video

The same four steps run on every check, in this order. The pictures below are from a real run.

1.Sample 15 frames
A video is read as 15 frames, evenly spaced from the first to the last. A still image is read as one frame.
Frame 015 of 150 sampledFrame 149
2.Degrade them like social media
Every frame is cut to 360 px high, blurred, saved as a JPEG at quality 55 and resized to 224 × 224. A clean upload and a re-encoded TikTok are judged the same way. The slider stretches the 224 × 224 input back to the frame's shape so you can compare them.
As uploadedAs the model sees it
Move the divider between the original frame and the model's input
3.Turn each frame into numbers
EfficientNet-B0, trained on ImageNet with its last layer removed, turns each frame into 1280 numbers. The 15 sets are averaged into one.
EfficientNet-B0
ImageNet weights, classifier head removed
1280 numbers per frame
averaged over all 15 frames
SVM
probabilities for Real, Sora 2, Veo 3
4.Classify
A support-vector machine gives a probability for Real, Sora 2 and Veo 3. The highest one is the result, and its probability is the confidence. Under 60%, we call it inconclusive.
Real
57.6%
Sora 2
33.9%
Veo 3
8.6%

Every check ends in a report like this one

Frame 015 of 62 sampledFrame 61

fishermen-on-a-lake.mp4

Example report. Made with Sora 2, so the right answer is Sora 2.

This clip most closely matches Sora 2.

74.7%

Confidence: the probability the model gives Sora 2.

Real: 1.0%
Sora 2: 74.7%
Veo 3: 24.4%

Compared against: Real, Sora 2 and Veo 3
Frames analyzed: 15 of 62, evenly spaced

Technical details

The frame as uploaded — Frame 30, as uploadedAs the model sees it, squeezed to 224 × 224

The same frame as the model receives it: 360 px high, blurred, JPEG quality 55, 224 by 224 — Frame 30, as uploadedAs the model sees it, squeezed to 224 × 224

Returned class: Sora2
p(class): 0.7467
Features: EfficientNet-B0, ImageNet weights, 1280 values per frame
Classifier: SVM with probability outputs over Real, Sora 2, Veo 3
Preprocessing: 360 px height, 3×3 Gaussian blur, JPEG quality 55, 224 × 224
Inconclusive: below 0.60
What the file is: Made with Sora 2

The file comes first
You see exactly what was checked, with the 15 sampled frames under it.
One sentence
Which of the three classes the clip matches most closely, in plain words.
Confidence
The probability the model gives that class. Under 60% the report says it can't call it.
Technical details
The raw class, the exact probability, the preprocessing, and the frame as the model saw it.

One request, one answer

Send a file or a public TikTok link to the endpoint the web app uses. Every account, including free ones, can call it with its sign-in token. Each call uses one check.

Read the API docs

curl -X POST https://ajnasiou-authini-api.hf.space/detect/ \
  -H "Authorization: Bearer $AUTHINI_TOKEN" \
  -F "file=@clip.mp4"

Response for latte-art-in-a-cafe.mp4

{
  "id": 4127,
  "type": "file",
  "result": "Veo3",
  "confidence": 0.8714,
  "probabilities": {
    "Real": 0.0946,
    "Sora2": 0.0340,
    "Veo3": 0.8714
  },
  "frames_analyzed": 15,
  "timestamp": "2026-10-24T09:41:07",
  "user_id": 312
}