How ClipForensics works
From upload to verdict in under 60 seconds. Here is everything that happens to your video.
Analysis Pipeline
Five stages, fifteen modules, one verdict
Every video goes through the exact same pipeline. No shortcuts, no black boxes.
Upload or URL submission
Drop a video file (MP4, MOV, AVI, MKV, WebM) or paste a URL from YouTube, TikTok, Twitter, or other platforms. We accept files up to 500 MB.
Frame & audio extraction
FFmpeg extracts up to 20 representative frames using scene-change detection, plus full audio tracks. No frames are cherry-picked — the sampling algorithm is deterministic.
15 forensic modules run independently
Each module analyzes a different dimension: metadata, compression signatures, facial manipulation, lip-sync, optical flow, spectral frequency, biological motion, lighting consistency, and more. Every module produces a confidence score and evidence list.
Evidence fusion with weighted scoring
Module results are combined using calibrated weights. Agreement bonuses reward cross-module consensus; contradiction penalties flag conflicting signals. The fusion engine outputs a single trust score with a confidence interval.
Verdict, evidence timeline & report
The final report includes a verdict category, trust score, per-module breakdowns, evidence timeline, compression history, and optionally a cross-video fingerprint. All results are downloadable as JSON or PDF.
Ingest
Extract
Analyze
Fuse
Report
Transparency
What makes ClipForensics different
We don't hide behind 'AI magic.' Every result is explainable.
Most deepfake detection tools give you a single score with no explanation. ClipForensics shows you why a video was flagged — which modules triggered, what evidence they found, and how confident each assessment is.
Our verdict categories range from "Verified Origin" (backed by content credentials) to "High Manipulation Risk" (multiple strong indicators). When evidence is insufficient, we say "Inconclusive" instead of guessing.
We also publish our known limitations and encourage users to treat results as one input among many.