| name | workflow-analysis-quality |
| description | Deep media inspection + automated QC — ffprobe stream details, MediaInfo diagnostics, VMAF/PSNR/SSIM quality metrics, PySceneDetect scene cuts, crop/silence/black-frame/interlacing detection, ffplay scope debugging, NAL/SEI bitstream forensics, metadata audits, loudness compliance against Spotify/Apple/ATSC/EBU specs, and automated CI QC gates. Use when the user says "QC this file", "run VMAF", "compare encoders", "detect scene cuts", "check loudness compliance", "validate delivery spec", "automated QC pipeline", or any deep inspection / quality gating. |
| argument-hint | [ref] [out] |
Workflow — Analysis + Quality
What: Inspect every byte and assess every rendition. The gate between "encoded" and "shipped".
Skills used
ffmpeg-probe, media-mediainfo, ffmpeg-quality, media-scenedetect, ffmpeg-detect, ffmpeg-playback, ffmpeg-bitstream, media-exiftool, ffmpeg-metadata, media-ocr-ai, cv-opencv, cv-mediapipe, media-tag.
Pipeline
Step 1 — Probe everything
moprobe + ffmpeg-probe — full format + streams + specific stream detail + packet-level bitstream dump when needed. moprobe --color for HDR side data.
Step 2 — MediaInfo deep diagnostics
media-mediainfo (--Output=JSON|XML|HTML). Shows encoded-library, CABAC/trellis/B-pyramid, source framerate variance, recording-device metadata, re-encode history. Use when ffprobe alone is insufficient.
Step 3 — Quality metrics (VMAF / PSNR / SSIM)
ffmpeg-quality or moqc:
| Model | Use for |
|---|
vmaf_v0.6.1 | PC monitor (1920×1080 default) |
vmaf_v0.6.1neg | No-enhancement (catches sharpening cheats) |
vmaf_4k_v0.6.1 | 4K TVs |
vmaf_b_v0.6.3 | Mobile / phone |
Interpretation: ≥ 95 transparent, ≥ 80 excellent, ≥ 70 good, ≥ 60 acceptable, < 60 visible loss.
Step 4 — Scene detection
media-scenedetect (content method, default threshold 27). Lower (15–20) for gradual transitions; higher (35–40) for abrupt cuts only. Inline scdet filter in ffmpeg if you don't need PySceneDetect's graph output.
Step 5 — Crop / silence / black / interlacing
ffmpeg-detect:
cropdetect — auto-detect letterbox/pillarbox, emits crop filter args.silencedetect — -35 dB default for dialogue, -50 dB for music.blackdetect — pic_th 0–1 luma threshold (0.98 = near-pure-black).idet — counts fields: high TFF/BFF = interlaced, high Progressive = progressive, mixed ~60/40 = telecined (use workflow-ai-enhancement or ffmpeg-ivtc to recover).
Step 6 — ffplay scopes
ffmpeg-playback — waveform, vectorscope, histogram live.
Step 7 — Bitstream forensics
ffmpeg-bitstream:
- NAL dump (HEVC SPS=33, H.264 SPS=7).
- SEI dump for HDR metadata / captions / DoVi / HDR10+.
Step 8 — Metadata audit
media-exiftool (image / video EXIF, XMP, IPTC). ffmpeg-metadata for chapter + MKV tags.
Step 9 — Content verification with CV/AI
media-ocr-ai for burned text; cv-mediapipe for face presence; media-tag CLIP/SigLIP for zero-shot content tags.
Step 10 — Loudness compliance
media-ffmpeg-normalize --measure-only (non-destructive). Compare to spec:
| Spec | Integrated | True peak |
|---|
| Spotify Master | −14 LUFS ± 2 | −2 dBTP |
| Apple Podcasts | −16 LUFS ± 1 | −1 dBTP |
| ATSC A/85 | −24 LUFS ± 2 | − |
| EBU R128 | −23 LUFS ± 0.5 | −1 dBTP |
Step 11 — Automated CI QC gates
moqc --ref source --out encoded --format json --vmaf-min 93 exits non-zero on failure. Wire into CI to fail builds that would ship broken encodes.
Variants
- Regression suite — new encoder vs golden corpus, VMAF per source.
- Streaming monitor — every 60 s, capture 10 s, VMAF against golden.
- Bulk spec validation —
media-batch parallel validate.
- Interactive debugging — ffplay with PTS overlay, vectorscope + waveform side-by-side.
- ABR ladder compare — VMAF per rung to tune quality.
- Broadcast spec QC — color legality (
signalstats), interlacing (idet), timecode continuity, caption presence, ATSC A/85 loudness.
Gotchas
- VMAF reference and distorted MUST match duration and frame rate. Mismatch = silent aligner failure → meaningless score.
- VMAF prefers matching resolution. If distorted is lower-res, it upscales silently. Downscale reference explicitly for HD-vs-4K comparison.
- VMAF default model (
vmaf_v0.6.1) is for PC monitors. Pick _4k, _mobile, _neg per target.
- PSNR-YUV vs PSNR-Y-only. ffmpeg reports overall; Y-only via
psnr=...y.
- Frame drops or duplications break VMAF sync. Need 1:1 correspondence.
- Different YUV ranges (limited vs full) change perceptual VMAF. Clamp to same range before compare.
- Scene-detect threshold 27 default: gradual transitions need 15–20, abrupt cuts tolerate 35–40.
cropdetect samples every N seconds. Variable content = inconsistent reads. Longer duration = reliable.silencedetect -35 dB is dialogue. Music is −50 dB; noise floor is −60 dB.idet gives counts, not decisions. TFF ≈ 2.5× Progressive = telecined (5 fields per 4 frames).blackdetect pic_th is 0–1 picture-luma, NOT dB. 0.98 = near-pure-black only.- MediaInfo "Encoded Library" is a heuristic. Re-encoded files may show wrong encoder.
ffprobe duration_ts is in stream timebase; duration is in seconds. Don't conflate.nb_frames=0 means unknown (stream not fully parsed). Use -count_frames for accurate.- ExifTool writes in place by default.
-overwrite_original strips backups.
- NAL types differ between H.264 and HEVC. Don't swap: H.264 SPS = 7, HEVC SPS = 33.
- SEI messages in HEVC use type+size+payload with UUID-prefixed T.35 for DoVi / HDR10+ / captions.
- Bitstream analysis on encrypted content fails. Widevine / FairPlay encrypt the essence below the container.
ffplay blocks. Use -autoexit to exit after duration or background with .
Example — CI gate on an encoded asset
moprobe --color source.mov → sanity-check tags. moqc --ref source.mov --out encoded.mp4 --vmaf-min 93 --format json → fail build non-zero. ffmpeg-detect silencedetect confirms no > 2 s dead air. media-ffmpeg-normalize --measure-only confirms −16 LUFS ± 1.
Related
- Every other workflow — this is the gate.
workflow-broadcast-delivery — broadcast-spec QC subset.workflow-streaming-distribution — ABR rung comparison.
