Tracy Harris Co · video pipeline · 5 September 2026

Same clip, three renders

Tracy's test clip (IMG_7647, 99.8 s, untouched footage, captions only) rendered by the reel-captions command we already have, and by HyperFrames. Same transcript, same caption chunks, same emphasis words, same fonts. Only the renderer differs.

A. reel-captions (what we have)

ffmpeg + ASS subtitles

Download A (720)

Captions	Burned in by ffmpeg from an ASS file. Whole chunk appears at once.
Emphasis word	Editors Note italic, same white as the body
Grade	Warm lift applied
Loudness	-14 LUFS master
QA gate	PASS

B. HyperFrames

HTML + CSS + GSAP, headless Chrome + ffmpeg

Download B (720)

Captions	Chunk fades up, each word brightens as it is spoken
Emphasis word	Editors Note italic in cream with a gold underline
Grade	None, footage untouched
Loudness	-14.60 LUFS, true peak -0.22 dBTP (mastered after render with ffmpeg)
Render	2m 44.3s for 107 s (3,216 frames) on the box, 720 copy 13MB
QA gate	PASS, same five checks

C. HyperFrames, full brand build

Karl, 5 Sept: "a version that matches the brand but uses all the functions available." Same footage, untouched. Everything else is a layer or a card.

cards + overlays + audio chain + punch-ins

Download C (720)

Hook card	Cream card, THC symbol, kinetic Editors Note line, gold rule, wipes up off the footage
Lower third	Name, role, handle at 8 s, sub-composition with variables
Captions	Same rail, word lights on the beat, no underline (Karl)
Hero line	"no method to the madness" lands word by word, replaces the rail for one beat
Stat callout	1,500 counts up with a progress track, from her own words
B-roll	Five Seedance cutaways cut to her exact words (the outfit reel, no mic, the music up, the comments, doing the right things), mixed ages and backgrounds, real rooms
Punch-ins	16, one per new thought, eased, on a wrapper (footage untouched)
Voice chain	Rumble cut, mud cut, compressor, clarity lift, limiter, all inside the composition
Music bed	Under the hook, ducked as she starts, out by 12 s, back under the outro (volume automation)
Outro	Logo, "So press record.", handle, FFB pill, slides up over the last beat
Progress	Gold hairline across the whole clip
Render	2m 44.3s for 107 s (3,216 frames) on the box, 720 copy 13MB
Loudness	-14.32 LUFS, true peak -0.94 dBTP (post-mastered with ffmpeg to the house target)
Check	PASS, 0 errors, 13/13 contrast; QA gate PASS

What to look at

Caption motion. A cuts each chunk in hard. B fades it up and lights each word on the beat.
The emphasis words (worried, talking, people, and the rest). A keeps them white italic. B sets them in cream with the gold hairline, the 2026 system.
Colour. A carries the warm grade from the August pipeline. B ships the footage as shot, which is what the HyperFrames caption rules ask for.
Both use the same transcript timings, so if a word lands late in one it lands late in both.

How B was made

One HTML file, generated from our own plan.json and words.json by build.py, then npx hyperframes check (0 errors, 11/11 contrast checks) and npx hyperframes render. The composition lives at /home/claude-bot/hyperframes-studio/tracy-7647-captions/. Change the CSS, re-render. No vendor holds the edit.

What this does not test

Cuts, punch-ins, hook and CTA cards, and the ad preset. Those all still come from the reel-captions manifest. B is captions on the raw clip only, which matches the stage one default.