AI media quality review

Review the actual downloaded file against its brief and source media. Run the bundled inspection script when FFmpeg and FFprobe are available:

python3 scripts/inspect_media.py /absolute/path/to/output.mp4 --output-dir /absolute/path/to/review

The script creates probe.json, review.json and a contact sheet or waveform. It reports technical signals; it does not decide whether the creative result is good. A diagnostic failure stops the script and removes an earlier review.json in that output directory. Treat a nonzero exit or missing report as unverified, never as a clean result.

Apply the acceptance criteria

Read the original request and list the few required outcomes and protected details. Inspect the generated artifacts at full size and the intended delivery size.

• Image: composition, crop, identity, anatomy, hands, geometry, product/character details, required text and logos, unwanted text and visible artifacts.
• Video: first and last frames, intended action and camera, identity/object continuity, flicker, warping, abrupt cuts, actual dimensions, frame rate, duration and audio streams.
• Speech/music: listen through the full export; check intelligibility, pronunciation, last word or phrase, clipping, silence, mouth timing, cut points and synchronization.

Inspect the actual last frame as well as the contact sheet. A gradual fade can darken a face without crossing the black-frame threshold; an empty black_segments list does not rule it out. Review contact-sheet flags in context. Intentional black or silent sections are not defects. A clean diagnostic report cannot prove smooth motion, accurate lip sync, good music or a compelling result. Full-speed viewing and listening remain required for those claims.