Event recap video

Tested end to end on a 780-item event library. Requires ffmpeg/ffprobe on PATH and Python 3.10+ with pip install pillow pillow-heif numpy.

Scripts live next to this file. Copy into the project:

• scripts/analyze.py, build.py, reel.py, motion.py, fonts.py → project root
• scripts/work/*.py → <project>/work/
• examples/*.json → project root, then edit them

Project layout: raw/ (untouched originals), music/, work/ (cache), output/, edit.json (the whole edit). Fonts are found automatically on Windows/macOS/Linux; override with FONT_SERIF, FONT_SERIF_BOLD, FONT_SANS_BOLD, FONT_BLACK, FONT_EMOJI, FONT_DEVANAGARI. Nepali/Hindi titles, labels and reel text switch to a Devanagari font automatically and ffmpeg shapes them correctly, but it shapes a whole line in the script of its first letter: start a mixed line with the Nepali word ("दिन १ · Jun 12", not "Jun 12 · दिन १"). Emoji only work in motion.py stickers; build.py and reel.py drop them because ffmpeg draws them as boxes. motion.py stickers and the shortlist are drawn with Pillow; their Devanagari is shaped by ffmpeg word by word when Pillow has no raqm.

1. Get media

• Put each export/zip in its own folder under raw/. Never modify originals.
• Google Photos has no API for an existing library: drive the user's own logged-in browser over CDP (see the brave-cdp skill in this repo). Grid items are a[href*="/photo/"]; day headers are [role=checkbox][aria-label^="Select all photos from"].
• Select day headers with REAL mouse events (Input.dispatchMouseEvent), verify the "N selected" text, press Shift+D. The tab must be in front (Page.bringToFront) or the grid stays grey.
• Google clears the selection after a download and header aria-checked can go stale; reload before reselecting.
• Browser.setDownloadBehavior with allowAndName saves the zip under its GUID with no extension; watch Browser.downloadProgress for completed.

2. Analyze and pick

• python analyze.py → work/media.json (capture time from EXIF / video creation_time in local tz, blur, brightness, duplicates, loudness) + per-day contact sheets in work/sheets/.
• python work/vsheets.py → 3-frame strips per video. LOOK at every sheet. Skip clips that rotate mid-recording, pocket shots and dark night clips.
• python work/loudest.py ID:seconds ... finds the loudest window (singing, speech, cheering). Verify with a 1 fps frame strip before trimming.
• Shared/forwarded files have no capture date (falls back to file mtime) — place them by eye.

3. Edit (what audiences liked)

• Lead with VIDEO (~90% clips); photos only for strong group shots. Photo-heavy cuts feel slow.
• Cold open hook: 8–10 clips of 1.2–1.8 s, hook_xf 0.1 near-hard cuts, then the title card. Then days in order with a small "label": "Day N · date" on each day's first clip, 3–4 s cuts, and a few 9–12 s moments flagged "speech": true (music ducks under them).
• Music makes or breaks the hook. Chain 2 tracks of similar tempo ("music": [a, b], 3 s crossfade) instead of looping one. python work/bpm.py file.mp3 estimates tempo.
• edit.json keys: title, subtitle, closing, closing_sub, title_bg / closing_bg (+ _bg_at for video ids), music / vertical_music, music_volume, duck_volume, ambient_volume, photo_seconds, hook_xf, closing_seconds, horizontal_hook, horizontal, vertical_hook, vertical. Segment: {"id", "at" (0–1 centre) | "start", "dur", "speech", "label", "xf"}.
• python work/shortlist.py → output/shortlist.jpg (a frame at every cut). Show it to the user before rendering.

4. Render and QC

• python build.py [horizontal|vertical|vertical-full] — per-segment cache in work/seg, xfade chain, loudnorm −14 LUFS, x264.
• EDIT=other.json OUT_SUFFIX=-v2 python build.py renders an alternative edit without touching the main one.
• QC every render: nb_frames == duration*30 (xfade chains can drop frames), blackdetect count 0, ebur128 ≈ −14 LUFS, and a tile sheet of frames (fps=1/5,tile=8x6) to eyeball rotation and cut-off titles.
• ffmpeg 7+ prefers -/filter_complex file over the deprecated -filter_complex_script.

5. Beat-synced reel (IG / TikTok / FB) — reel.py + reel.json

• python reel.py → output/reel.mp4 (~45 s), built on build.py's segment renderer.
• Cuts land exactly on the beat grid (bpm hint, auto phase, frame-rounded cut times). Double-check tempo: naive estimators often return half/double (85 vs 154).
• First second = the most surprising moment, with a "POV: …" line. 4-beat cuts, 6–8 beats on hero moments, slow push-in on every clip, white flash on section starts.
• Text: heavy sans, white with thick black outline, ~24% from the top (clear of TikTok/IG buttons), slide-up fade. 5–7 short lines total; the last one is the event name.

6. Meme / motion cut — motion.py + motion.json (only when the user asks for memes/stickers)

• Many people find this style too much for a heartfelt event. Default to sections 3–5.
• One long, steady track (~3 min) instead of chaining; beat grid via comb-filter search.
• Per clip: beats, speed (0.5 slow-mo / 2 fast-forward), freeze (beats; tape-stop + silence + desaturate), cap sticker (rounded box, tilt, elastic pop, colour emoji), day slam titles + whoosh, title, verse, shake, punch, flash, zoom.
• If a render is killed, delete clip files that fail ffprobe before re-running (the cache would reuse them).

7. Vertical and social delivery

• When someone approves the 16:9 edit and asks for 9:16, give them the SAME edit vertical (python build.py vertical-full). Don't restyle an approved edit; "make it more attractive" means polish only.
• Platforms (especially Facebook) mute posts with unlicensed music. Offer an original-sound version (an edit with "music": null) and/or a silent copy (ffmpeg -i in.mp4 -an -c copy out.mp4) so music can be added from the platform's licensed library.
• Final polish: eq=contrast=1.07:saturation=1.2:gamma=0.97,colorbalance=rm=0.03:bm=-0.03:rh=0.02:bh=-0.02,unsharp=5:5:0.45,vignette=angle=PI/5.5 + audio highpass=80,dynaudnorm,acompressor,loudnorm=I=-14.