Phenomainon
Authenticate · /analyze

Run a photo or video through the lab.

Phenomainon runs four independent pipelines on your upload: YOLOv8 object detection, an 83-class aircraft classifier we trained ourselves, motion-event tracking, and forensic authenticity checks against metadata, frame duplication, scene cuts, and audio splices. Nothing is stored — the file is deleted as soon as we read it.

Heuristic provenance check · not deepfake / GAN detection
Upload

Drag a photo or video here, or pick one from your device. Up to 500 MB.

How it works
  • Detection. YOLOv8m runs on the image, or every 15th frame of a video, and reports what everyday objects appear: airplane, car, kite, person.
  • Classification. A ResNet50 fine-tuned on 83 specific aircraft (F-22, MiG-31, SR-71, Tu-95…) casts a confidence-weighted vote per frame. Low confidence is the “not a known aircraft” signal.
  • Motion. MOG2 background subtraction tracks moving regions over time and groups them into events with start/end timestamps.
  • Authenticity. ExifTool + OpenCV + librosa look for editing-software tags, missing timestamps and GPS, duplicate-frame runs, abrupt scene cuts, and audio splices.
Limits & caveats
  • Up to 500 MB. Files under 4 MB stream straight to the proxy. Larger files upload to blob storage and run through an async job queue with progress polling — same result, just longer wait.
  • Videos take time. Inference runs on CPU. A 30–90 second video can take tens of seconds.
  • Authenticity is not deepfake detection. It flags editing and handling artifacts — signs of cuts, re-encoding, missing provenance — not AI-generated content.
  • The classifier has no “unknown” class. Genuinely anomalous objects get force-matched to the nearest of 83 known aircraft with low confidence.