CineQC
A broadcast-grade quality-control pass for AI-generated video. CineQC decodes your clip on your own machine, flags the artifacts machines can catch, then walks you through the judgment calls only an eye can make — and hands you a report you can defend.
Your video never leaves this device. Standard decoding and scoring run in the browser; optional SyncGuard sends the clip only to the local DVP engine on this machine — no cloud upload or account.
How a QC pass runs
Four steps, one continuous flow. Drop a clip and you leave with a scored verdict and an exportable report.
Drop the clip
MP4, WebM, or MOV. It decodes locally, frame by frame, downscaled for fast analysis.
Automated scan
Seven metrics run per frame and render as timeline heat lanes you can click to seek.
Score the rubric
Seven anchored categories, 1–5, blended by a weighting preset into a composite verdict.
Export the report
A machine-readable JSON or a one-page PDF scorecard with the worst frames pulled out.
What it measures
Module A is the machine pass: automated signals across picture and sound, most with a 0–100 sub-score and a heat lane. It is deliberately honest about what automation can and cannot catch.
Temporal Flicker
High-frequency luma oscillation — the brightness pumping and exposure instability generators leave behind.
Stutter / Fake FPS
Duplicate-frame detection. A "24fps" clip that is really 8fps content shows a collapsed unique framerate.
Color Drift
Histogram distance across a sliding window catches palette and white-balance wander mid-shot.
Sharpness
Per-frame focus variance flags sudden detail loss — the melting and resolution collapse motion exposes.
Motion Coherence
Coarse block-flow disagreement between frames — a proxy for morphing, warping, and chaotic motion.
Scene Cuts
Histogram spikes segment the clip so every metric resets per shot and cuts never read as drift.
Structural
Resolution, aspect, duration, and measured framerate — the delivery-spec basics, read straight off the clip.
Loudness (EBU R128)
Integrated loudness in LUFS to the ITU-R BS.1770 standard, read from the clip's audio, so a mix that is too hot or too quiet is caught alongside the picture.
Out of scope for the standard browser scan, and handled by the human pass instead: anatomy errors, text and logo integrity, and semantic continuity. Lip sync can be checked separately with the optional local SyncGuard gate.
Optional AV and lip-sync gate
Choose Run SyncGuard when a clip contains a visible speaking face and lip sync matters to delivery. The local DVP engine measures signed audio/picture offset across speech windows, detects progressive drift, and returns inspectable intervals that seek to the exact source frame.
Explicit and local
Nothing runs automatically on an ordinary CineQC upload. Run, Cancel, and Retry control a CUDA analysis at 127.0.0.1; there is no provider key or cloud fallback.
Evidence, not guesswork
The panel shows confidence, coverage, global offset, maximum drift, and a keyboard-operable list of flagged intervals. Negative offset means audio leads; positive means it lags.
Conservative verdict gate
CineQC preserves the calculated verdict. A conclusive SyncGuard CONDITIONAL can cap a calculated PASS, and FAIL forces the final verdict to FAIL. Offline, cancelled, busy, or low-confidence runs never alter it.
SyncGuard measures audiovisual correspondence; it does not promise phoneme-perfect recognition. Clips without audio, usable speech, or one trackable speaking face return a non-conclusive result for manual review.
The scoring pass
Module B is where your trained eye does the work automation can't. Seven categories, each scored 1–5 against anchored definitions so scores stay repeatable across clips and reviewers.
- Temporal Stability — pre-filled from the flicker and motion scores; override if the numbers miss the eye's read.
- Anatomical / Object Integrity — hands, faces, limbs, object permanence.
- Physics Plausibility — gravity, collisions, cloth and fluid behavior.
- Text & Logo Integrity — signage, jerseys, UI elements.
- Continuity — lighting direction, shadows, props across the shot.
- Composition & Framing — horizon, safe areas, headroom.
- Prompt Adherence — optional; paste the generation prompt and judge the match.
A weighted blend of the categories you scored, tuned by preset — Broadcast Delivery, Social / Web, or Research Eval — resolves to a 0–100 score, a letter grade, and a PASS, CONDITIONAL, or FAIL call. Pin timestamped notes to any category as you scrub.
The report you leave with
Two exports, both built for defensibility. Nothing here is marketing gloss — it's the record of the pass.
JSON
The full data: every metric array, flag, rubric score, note, weight, and file detail. Machine-readable for batch comparison across clips and models.
PDF scorecard
A one-page artifact: verdict banner, a radar of the categories, a flagged-timestamp table, a strip of the worst frames, and the evaluator and date.
Beyond a single pass
One clip is the start. CineQC also scales to many, compares takes, explains what it found, and hands a failure straight to the rest of the suite.
Fast mode
Analyze at higher playback speed for a quicker, coarser read when you just need a gut check.
Precision mode
Frame-accurate seek stepping: every source frame captured, deterministic scores, immune to dropped frames. Slower — and the most defensible pass.
Batch triage
Drop several clips at once and get a table of verdicts and top failures, so a whole render queue is one pass.
A/B compare
Put two takes side by side and read every metric delta, to prove which version actually holds up.
AI Explain
Send the worst flagged frames to Gemini, under your own key, for a plain-English read of what broke and how to fix it.
Shareable report
Copy a link that reopens the full read-only report in any browser, with no account needed.
Team sign-off
Stamp a clip Approved, Conditional, or Rework, so the verdict carries a decision, not just a score.
Model Arena
Rank models against each other over many clips instead of judging one render at a time, so a routing decision rests on a record rather than an impression.
RLHF preference pairs
Export an A/B compare or a batch run as JSONL {chosen, rejected} training pairs, each carrying the five metric scores, a confidence grade from the margin, and the metric that decided it. This is QC output turned into preference-tuning data.
Pin timecode
Mark the exact frame where something breaks so the report points at it directly, instead of asking a colleague to go hunting for the artifact you saw.
Verify provenance
Hand the clip to OriginTrace to read its C2PA credentials. A clip can score well and still be misdeclared about how it was made.
The QC loop
Send a failed clip to Shot Doctor to diagnose and repair, or Save to Vault to remember the filename, composite verdict and score, automated metrics, human rubric, loudness, and team sign-off. The clip and captured frames are not saved. Allow the popup and wait for Saved to Vault ✓. See the full QC to Repair to Remember loop.
Trust, but verify.
Bring the clip you're not sure about. Leave with a score you can put your name on.
Open CineQC