Forensic breakdowns
Step-by-step walkthroughs showing how ClipForensics analyzes videos — module by module, from raw data to verdict.
Face-swap deepfake breakdown
How ClipForensics detects a modern face-swap deepfake through multi-module analysis.
Metadata Analysis
File created with x264 encoder (not a camera). Creation timestamp recent. No GPS or camera model data.
Weak indicator — many legitimate videos are re-encoded with x264.
Compression History
Two encoding passes detected. First pass consistent with a screen recording tool. Second pass from social media re-encoding.
Moderate indicator — two passes suggest processing, but screen recording → upload is a common legitimate workflow.
Face Manipulation
Boundary artifacts detected at jawline in 73% of frames containing faces. Resolution mismatch between face region and background. Inconsistent noise patterns.
Strong indicator — these artifacts are characteristic of face-swap techniques.
Lip Sync
Phoneme-viseme misalignment exceeding baseline in 4 of 12 analyzed speech segments. Timing offsets of 80-120ms.
Moderate indicator — misalignment is present but could partially be explained by compression artifacts.
Temporal Consistency
Subtle flickering at face boundaries visible in optical flow analysis. Face region shows different temporal noise characteristics than background.
Strong supporting indicator — temporal inconsistency at face boundaries corroborates face manipulation finding.
Evidence Fusion
Face Manipulation (strong) + Lip Sync (moderate) + Temporal (strong support) = consensus across independent modules. No contradicting modules. Agreement bonus applied.
Verdict: High Manipulation Risk. Trust Score: 0.15. Confidence: High.
AI-generated video breakdown
How ClipForensics identifies a text-to-video generated clip.
Metadata Analysis
No camera metadata. Encoder identified as NVENC (GPU-based). No creation tool metadata.
Moderate indicator — absence of camera data is suspicious but not conclusive.
Spectral Analysis
Frequency domain shows characteristic patterns associated with diffusion-model upsampling. Periodic energy spikes at specific spatial frequencies.
Strong indicator — these spectral patterns are not produced by camera sensors or traditional editing.
Biological Motion
Human figure in video shows subtle gait anomalies. Eye blink rate (2.1/min) significantly below natural range (15-20/min). Hand movements show micro-jitter inconsistent with natural motor control.
Strong indicator — biological motion patterns differ significantly from natural human movement.
Optical Flow
Minor temporal morphing detected — objects gradually change shape over 10-15 frame windows. Background elements show subtle drift.
Moderate indicator — consistent with AI generation but could also indicate heavy post-processing.
Audio Synthesis
Speech shows spectral patterns consistent with neural TTS. Breathing absent between phrases. Environmental acoustics don't match visual environment.
Strong indicator — audio is synthetically generated.
Evidence Fusion
Spectral (strong) + Bio Motion (strong) + Audio Synthesis (strong) + Optical Flow (moderate) = broad consensus. Metadata supports. No contradictions.
Verdict: Likely AI-Generated. Trust Score: 0.11. Confidence: High.