render-narrated-ugc-wardrobe-stitch
Assemble a narrated-UGC "stitch reply" ad from a config: a fast-cut vertical testimonial
where a single spoken VO carries a verbatim ~13-sentence testimonial over ONE creator
across ~5 wardrobe changes in ~3 micro-worlds, interspersed with product B-roll (e.g. product macro,
unboxing, a landing-page scroll), ~30 hard cuts on the VO cadence, closing on a brand end card.
This capability is the FREE, deterministic assembly — trim-to-EDL, hard-concat, the VO+music
mix, the word-by-word caption burn, the landing-page zoompan, and the end-card append.
scripts/config.example.json is one worked example (Bioma "Do NOT buy Bioma Probiotics", ~37s
1080×1920 9:16, ~30 body cuts + a ~2s end card) — its creator, voice, hook, worlds and music are
that demo's answers, not defaults; scripts/PIPELINE.md maps every config block to
its source step and scripts/README.md documents the free assembly.
Choices
The creative calls are made upstream by the user (the recipe's choices) and arrive in the config;
this assembly never picks them.
- creator — who is on camera →
character.descriptor / character.name. The demo used a
28-year-old blonde woman.
- voice — the narration voice →
vo.voice_id / vo.settings. The demo used an excited voice
matched to its creator.
- hook_angle — the testimonial's hook →
vo.hook_line / vo.script_md / vo.payoff_line. The
demo used a "Do not buy <brand>" reversal.
- worlds — the 3 places →
worlds.briefs. The demo used bedroom / kitchen / bathroom.
- music — the bed under the VO, or none →
audio_mix.music_brief. With "no music", skip the bed
and mix the VO alone.
Run
This is the FREE, deterministic assembly stage — it spends nothing. The paid inputs are
separate capabilities — the spoken VO (create-vo-elevenlabs) Whisper-aligned so the WORD
BOUNDARIES set the cut grid; one locked creator (create-image-gpt-image-fal anchor + ~5 wardrobe
edits chained off the anchor) + 3 world wides + per-cut start-frames (create-image-fal product
composites); and one Veo/Seedance i2v clip per cut (create-video-fal). Given the VO +
vo-final.words.json + edl.json + one clip per cut + a Playwright landing-page PNG + the brand
end-card PNG, render-narrated-ugc-wardrobe-stitch trims each clip to its EDL window, hard-concats
on the VO cadence, mixes the VO over the ducked bed, burns the word-by-word captions, appends the
end card → the master. Re-cuts reuse the existing VO / start-frames / clips and cost $0.
The assembly runs on a shared helper. montage.py lives in the shared stitch-videos-ffmpeg
atom, which this package lists in requires_skills, so it installs alongside. After
gooseworks fetch render-narrated-ugc-wardrobe-stitch it is at
/tmp/gooseworks-scripts/stitch-videos-ffmpeg/scripts/montage.py. Write a montage.json from the
config (field mapping in scripts/README.md), then:
M=/tmp/gooseworks-scripts/stitch-videos-ffmpeg/scripts/montage.py
python3 $M edl --spec montage.json --out edits/edl.json # check every clip + window first
python3 $M run --spec montage.json --out edits/master-final.mp4 --workdir edits/work
run builds the EDL from the VO word boundaries (word_range), trims and hard-cuts every clip
with the filter_complex concat filter, renders the landing-page scroll as a zoom/pan over the PNG,
holds the end-card PNG, burns word-by-word captions (one word at a time in one colour, with
respell; no active-word highlight), mixes the VO over the
ducked bed, and masters to -14 LUFS. It writes edits/work/manifest.json. This package ships no
scripts of its own.
Contract (the free assembly)
- The spoken VO carries the narrative — lock it FIRST. The VO IS the narration bed; the whole
ad is cut to it. Never plan the cut grid before the VO is locked and Whisper-aligned.
- Build the EDL from the VO's Whisper word boundaries. ~30 role-tagged cuts (
hook, feature,
reaction-insert, payoff-hold, b-roll-insert, landing-page); snap every cut window to the
word boundaries. The payoff line gets a HELD payoff-hold beat (~3× mean shot length).
- Hard cuts via
filter_complex concat, not the demuxer. Trim each clip to its EDL window and
hard-concat with filter_complex concat — the -f concat demuxer drops the audio when a
drawtext/scale step shaves a clip a few ms below its window. No dissolves.
- Karaoke-pop captions on every word, throughout. From the VO's
vo-final.words.json (VEED
Whisper preset, bold yellow), on every word; re-spell brand tokens Whisper mishears against the
locked script (Bioma demo: "synbiotic" over "symbiotic"; kept "I'ma" verbatim) — never edit the script to
match Whisper. Captions are suppressed over the end card. If VEED mis-captions a brand token,
hand-patch that sentence with local ASS karaoke.
- Product B-roll breaks up the talking head. Product macro (Bioma: capsules), unboxing, and a landing-page
scroll are interspersed with the creator cuts. The landing-page scroll is FFmpeg zoompan over
a Playwright-rendered PNG — not an i2v clip (i2v hallucinates the UI).
- VO over a ducked bed. Mix the optional instrumental bed sidechain-ducked UNDER the VO
(−20dB, 20:1) so the VO stays clearly on top; the bed can drop in on the payoff beat.
- End card via the brand's real PNG — never AI-render brand text. Append the brand's real
end-card PNG (~2s) on the tail, captions suppressed. A diffusion model garbles a wordmark.
- FFmpeg composite, deterministic, FREE. Trim-to-EDL,
filter_complex concat, VO+music mix,
caption burn, landing-page zoompan, end-card append, loudnorm I=-14 → a 1080×1920 h264+aac
master as long as the VO plus the end card (~37 s in the demo). No paid calls, no keys.