Calibrated identity scores
Per-character embeddings compared against an established reference, calibrated so a 0.82 means the same thing across every run.
Consistency analysis · v0.4 research preview
Find the consistency breaks in your AI-generated video
Character consistency, scene layout and object placement — measured frame by frame, scored on a calibrated scale, and written up as a timecoded findings report.
- Dimensions
- 3
- Max upload
- 2 GB
- License
- MIT
Drop a video here, or click to browse
MP4 · MOV · WEBM — up to 2 GB
What you get
Establishment ledger
Drift curves and object placement tracked from the first establishing shot forward, so late deviations are attributable.
Timecoded findings
Every violation lands on a frame range with a severity, a dimension and the evidence used to flag it.
How it works
- 01
Upload
Video goes to private storage; a job row is queued.
- 02
Analyze
An external worker claims the job and streams progress.
- 03
Report
Scores plus a timecoded HTML report, ready to share.
How scoring works
Identity consistency is calibrated per film, so 1.0 means the same subject re-rendered and 0.0 means a different subject entirely — not an arbitrary cosine distance.
The overall score multiplies consistency by persistence, so characters that drift can't hide by fragmenting into short-lived clusters.
Every finding is timecoded and derived from geometry and tracking, with an optional VLM judge pass for the qualitative breaks geometry can't see.