Video breakdown: reference.mp4

Content and structure

• No whole-video summary (needs the VLM or an agent watching the full video); subtitle text is under “Packaging layer”

Cast, scenes and props

• No cast extracted (needs the VLM whole-video pass, or an agent filling top-level cast in review_answers.json)

Sound

• loudness -49.6 dBFS, peak -45.5 dBFS; sound runs across cuts, laid in post as one bed (background changes 0.5dB at cuts vs 0.6dB within shots)
• Sources: Video type: music bed only (no voice) → music: stock library music

Voice

• None / very little

Music

• Tempo: 143.6 BPM (or 71.8/287.1); key A minor (confidence 0.653); 83% of the duration (0.0–5.0s); ending: ends_early
• Cuts vs beat: 0% of cuts land on a beat (chance: about 30%)
• Source: stock library music (existing_library 60%, ai_generated 20%, popular_song 20%)
• The music is a bed under the picture: in a regular video the bed usually comes from a stock library (prior)
• How to confirm: Check the music tag on the original Douyin / Xiaohongshu / TikTok post: a song title → existing track; "original sound - @creator" → uploaded by the creator (for music videos, usually an AI song)
• How to confirm: Run it through Shazam / ACRCloud or another song-ID service: a hit means an existing song
• Remake: The original track leans stock library music: search a library by BPM / mood / key for a similar track, or generate a royalty-free track in a similar style with AI

Sound effects

• Nothing prominent at the signal level (quiet effects under voice and music need the VLM, see --vlm)

• Separated stems: stems/original.wav (original), stems/vocals.wav (dry voice; transient sound effects land here too), stems/music.wav (music + ambience)

Camera, lighting and editing

• Editing: 20.0 cuts per minute, 2.0s average shot; no live-action shots

• Note: lighting ratio, soft / hard, colour temperature and shot size are measured on the subject's pixels (the face when there is one, otherwise the in-focus region, which is an estimate); platform transcodes crush the shadows, so hardness and fill read conservative; camera angle, light source, catchlight and rim light follow the image readers. Where measurement and readers disagree, both are listed.

Packaging layer (subtitles / styled text / stickers / sound-effect hits)

• No overlaid text or stickers detected
• Exports: packaging/timeline.json (SRT is the subtitle timeline; ASS carries colour, outline, position and entrance animation, swap in the original font; timeline.json is the full timeline)

Remake: burn subtitles and styled text in with packaging/overlays.ass (ffmpeg -vf ass=overlays.ass), or rebuild them from timeline.json in CapCut / Remotion; stickers and GIFs need matching assets (meme / GIF libraries), crops are under packaging/.

Remake plan

still-image move / static frame 3 shots → rebuilt locally

Local rebuild

• #1 static hold 2.0s; #2 static hold 2.0s; #3 static hold 2.0s: one image + the measured zoom / pan parameters, rendered straight from ffmpeg

Shot list

Open questions

• Shot #1 (0.0–2.0s): photo / image: real material or generated, to confirm
• Shot #2 (2.0–4.0s): photo / image: real material or generated, to confirm
• Shot #3 (4.0–6.0s): static hold at only 46% confidence, runner-up code_render
• Shot #3 (4.0–6.0s): photo / image: real material or generated, to confirm

Appendix

Shot by shot

Evidence, probabilities, text tracks and prompts per shot. Folded by default.

#1 00:00.000–00:02.000 · static hold

• Probabilities: static hold 71%, code / template render 22%, screen recording / screenshot 3%（deterministic）
• Still source: unknown
• Material: to confirm; real or not, to confirm: photo / image
• Picture is almost entirely still: 100% of adjacent frames are pixel-identical, first-to-last frame difference 0.00
• Photographic texture: noise σ=0.37, 75 colours, 9% flat blocks
• Remake: Static frame: one held image is enough

#2 00:02.000–00:04.000 · static hold

• Probabilities: static hold 68%, code / template render 25%, screen recording / screenshot 3%（deterministic）
• Still source: unknown
• Material: to confirm; real or not, to confirm: photo / image
• Picture is almost entirely still: 100% of adjacent frames are pixel-identical, first-to-last frame difference 0.00
• Photographic texture: noise σ=0.20, 67 colours, 11% flat blocks
• Remake: Static frame: one held image is enough

#3 00:04.000–00:06.000 · static hold

• Probabilities: static hold 46%, code / template render 45%, screen recording / screenshot 5%（deterministic）
• Still source: unknown
• Material: to confirm; real or not, to confirm: photo / image
• Picture is almost entirely still: 100% of adjacent frames are pixel-identical, first-to-last frame difference 0.00
• Remake: Static frame: one held image is enough

Metadata clues

• Encoder: Lavf62.12.100

Notes

• Deterministic forensics is reliable on motion, whole-frame pan/zoom, rendered texture and 24→30 fps conversion; telling live action from AI from 3D relies on the VLM and human review.
• Platform transcoding usually strips metadata; missing metadata proves nothing.
• The full data is in breakdown.json; agents read the compact view with rea-video summary.