The model behind Authini

Two stages. A frozen feature extractor turns each frame into numbers, and a small classifier decides which source those numbers look like.

Feature extractor

EfficientNet-B0 with ImageNet weights, classification head removed

Features

1280 values per frame

Video

15 frames, evenly spaced from first to last; features averaged into one vector

Images

One frame

Classifier

Support-vector machine with probability outputs

Before analysis every frame is degraded to look like a video that has been through a social platform, so that a clean upload and a re-encoded TikTok are judged the same way.

1. Resize to 360 px high, keeping the aspect ratio.
2. Apply a 3 × 3 Gaussian blur.
3. Encode as JPEG at quality 55 and decode again.
4. Resize to 224 × 224. The frame is squeezed, not cropped.

Classes

Real, Sora 2, Veo 3

Result

The class with the highest probability

Confidence

That class's probability, from 0 to 1

API response

Result, confidence, the probabilities for all three classes and the number of frames analyzed

Inconclusive

The web app shows any confidence under 0.6 as inconclusive

• The model only knows Sora 2 and Veo 3. Video from other generators can come out as Real, or as whichever of the two it is closer to.
• Real footage that looks unlike the real video it was trained on, such as polished stock footage, is often classed as generated.
• Fifteen frames are averaged into one prediction, so a clip that mixes real and generated shots gets a single, blended answer.
• Confidence is the classifier's probability, not a measured accuracy. A high number means the clip is far from the other classes, not that the answer is certain.
• Squeezing every frame to a square changes the proportions of tall and wide video.

Before Authini, the author studied a neighbouring forensics problem: telling which camera took a photo.