Video: DINOv2 spectral HF + codec-response fingerprint + TDA/MMD ensemble. Audio: A04 + A02 ladder + AAC64 post. Drop video or use Audio tab.
Drop video here or choose a video
MP4, MOV — max 15MB
Drop audio here or click
WAV, MP3, M4A, OGG — max 10MB
Processing on Lennox — extracting frames and running the forensic checks…0s
VIDEO ORACLE SURFACES (frame spectral / residual analysis)
Forensic signal topology
live embedding of measured surfaces → ensemble decision
supports realanomalousensemble
Ensemble Decision
Fingerprint (primary)—
TDA/MMD (fallback)—
Final verdict—
Research embedding trajectories
Research model comparison
TDA / MMD diagnostics (research only)
Sub-1Hz drift Shake 2–10Hz HF noise ≥12Hz
Original vs after AAC-64k
Key A04 metrics — blue = original extract, amber = re-encode probe
Original Post AAC-64k
Silence PSD (log freq)
Quiet-frame spectrum before vs after compression — brown beds tilt low; white beds stay flat
Original Post AAC-64k
Compression ladder
Encoded sizes + residual RMSE — real mic noise resists cheap codecs
Signal: real cameras produce high-frequency temporal noise (>12Hz) from sensor readout and compression. AI generators produce only smooth low-frequency drift.