REEL/REAL
AI Forensics

DEEPFAKE AI DETECTION Your Truth Amplified

Detect synthetic video and face manipulation. Clips are sampled frame by frame, and the report shows which moments were flagged.

reel.mp4synthetic sample
REEL generated
VS
real.mp4authentic sample
REAL original

Our system flags the moments that look manipulated, and shows its working.

Check a video

MP4 · MOV · WEBM — up to 200 MB

Drop a clip here, or choose from device

Your video is sent to the analysis server and deleted as soon as the result is returned. You get a verdict and a timeline of which moments were flagged.

✓ Face visible in most frames ✓ No re-encode needed ✓ Deleted after analysis
—
—
Extracting frames…

Analysis Report

—

—

—

Likely synthetic

Measured, and nothing was flagged.

Likely AI-manipulated0%
Confidence—
Suspicious footage—
Preview window

Timeline

Built from frames sampled across the clip, not from every frame. Click any bar to jump to that moment.

0:00—0:24
Below threshold Near threshold Above threshold — flagged

What we measured

Every row below is a number this analysis produced. The detector samples frames across the clip, finds the face in each, and scores it. The last row shows which part of the face the model paid most attention to — that describes how the model read the clip, not proof of editing.

Face visible in—
Frames flagged—
Longest run—
Flagged section—
Frame score—
Where it looked—

Evidence

These sentences are written directly from the measurements above — each one is produced only when the pipeline has a number to report.

    Reading the score: This detector was tested on videos it had never seen, and got 92% of them right. It is noticeably less reliable on clips that have been re-uploaded or compressed, as most social media video has been, and on kinds of manipulation it was not trained on. Treat the result as a reason to look closer, not as proof. The timeline below tells you more than the single number.

    Services & Methodology

    Multi-Signal Detection

    1. Your clip arrives

      It goes to the analysis server, and is deleted the moment the verdict comes back. Nothing is kept.

    2. Snapshots, not the whole film

      Thirty moments are taken from across the clip, start to end. Checking every single frame would be slower without being wiser.

    3. Find the face

      Each snapshot is searched for a face and cropped to it. The face is where manipulation happens, so it is the only thing examined.

    4. Score each one

      Every cropped face is compared against thousands of real and generated faces the system has studied, and given its own score.

    5. Look for a pattern

      Scattered odd frames are usually noise. A long unbroken run of them is not, so the two are weighed differently.

    6. Answer, or say it cannot

      If too few snapshots held a face, you get not enough to judge rather than a guess dressed up as an answer.

    Known Limits

    Full Transparency

    Novel Generative Models

    Unseen architecture types may lower detection confidence. Unrecognized patterns are treated as unresolved rather than clean.

    Heavy Re-compression

    Social platform re-uploads erase fine-grained camera noise, which makes a clip harder to judge. This is not currently detected, so a heavily re-compressed video is scored as if it were untouched.

    Micro-Edits

    A short edit inside a long video can be averaged away by the clip score. Check the timeline bars rather than relying on the single number alone.