01The source
rank03_1984790972155504.mp4 · 360×640 (delivered) · 30fps · 14.7s · AAC music bed, no voiceover
This is Sweat's (Kayla Itsines' app) #3 ad by rank in the page's top 15. One locked-off bird's-eye camera looks straight down at a model on a hot-magenta mat over whitewashed oak. She flows through a continuous pilates choreography while a kinetic ladder of white italic type stacks up over her. At 11s she breaks the fourth wall — looks up, smiles, raises a phone running the app to the lens — and the ad hard-cuts to an end card where the phone physically drops onto a magenta ground: SWEAT · 7-DAY FREE TRIAL, screen playing a real in-app workout.
Audio was transcribed (ElevenLabs) and came back empty — the ad carries no spoken words at all. Every message is on-screen type; the music only supplies tempo. Scene detection finds three genuine cuts (7.77s, 9.4s, 12.03s); everything before 7.77s plays as one unbroken take.
04The 50Queen version
The format transfers almost untouched — top-down mat, monochrome brand world, kinetic type ladder, fourth-wall phone reveal, physical end card. What must change is who is on the mat, what she does, and what the words promise. Sweat sells athletic aspiration to 25-year-olds; 50Queen sells "I could actually do that, and it would fix my mornings" to women 50+.
Sweat
20s pilates athlete, blush bra top + shorts
50Queen
Roxie (canonical charsheet), sage-green athletic unitard — the
identity_source wardrobe. Modest, high contrast on a mulberry mat, and it keeps glute-centric framing (and Higgsfield moderation) out of play.
Sweat
Hot magenta mat on whitewashed oak
50Queen
Mulberry mat (brand mark color) on the same pale oak. The monochrome trick carries over 1:1 — our brand color happens to live in the same family.
Sweat
Power-pilates choreography (arabesques, bicycles, deep folds)
50Queen
Mobility-sweep movements from our own library that read top-down and stay believable: butterfly knee openers, hip rocks, glute bridge, cat-cow, mermaid side reach, dead bug. Smooth and flowing — aspirational for
this audience, not intimidating.
Sweat
NEW AT-HOME PILATES → STRONGER CORE. → BETTER MOVEMENT. → REAL RESULTS. → PRECISION PILATES
50Queen
NEW AT-HOME MOBILITY → LOOSER HIPS. → EASIER MORNINGS. → REAL RESULTS. → FIFTYSTRONG (program name slot — open question below). First two benefits are in the audience's own vocabulary from quiz research; "Real results." is universal and stays.
Sweat
End card: SWEAT · 7-DAY FREE TRIAL, phone playing app demo
50Queen
Mulberry ground, 50Queen mark, offer line, phone playing a real screen-recording of a session. Offer copy is an open decision: free trial vs founding price.
Recommendation on scope: don't fully replicate the athletic choreography with Roxie as a shipping ad — a 50-something spokesperson doing power-pilates arabesques reads as uncanny, not aspirational. Run the faithful replication as a one-window academic spike (it answers "does top-down motion transfer hold Roxie's identity?"), then spend the real budget on the adapted version driven by our own movement clips.
05Production plan
Pipeline inherits the locked verdict from the rank06 ChillFit pilot (2026-08-26): MiniMax H3 motion-transfer (S1) wins over start-frame (S2). All generation waits for an explicit per-batch go-ahead.
Phase 1 — Academic spike: one faithful window
~34 cr
Cut W1 (0.0–2.0s, spine twist → fold — face mostly hidden, kindest to identity). Generate 4 Nano Banana Pro top-down start-frame candidates (refs: bilateral charsheet + source frame + mulberry-mat art direction) to validate the overhead look, then one MiniMax H3 run with --video-references = W1 segment and --image-references = charsheet + picked still. Deliverable: 3-up compare vs source.
Approval gate: verdict before Phase 2
Phase 2 — Adapted choreography windows
~150–200 cr
Six windows mirroring the source beat sheet, but motion references come from our movement-library clips (or newly picked reference footage of the six mobility moves), not the Sweat athlete. Same H3 motion-transfer recipe per window; 1–2 takes each. Top-down framing held constant by sharing one approved plate/still across windows.
Approval gate: per-batch generation
Phase 3 — Post: type, edit, end card
0 cr
All type is done in post, never generated — heavy italic condensed caps, white, centered, cumulative stack, ~2s dwell per message (ffmpeg/ASS, same toolchain as the rank06 overlay pass). Assemble windows to the music grid with speed-conform. End card is built natively: mulberry ground, 50Queen mark, offer line, a phone plate with a real app screen-recording composited in (never generate UI — it won't be legible or truthful).
Phase 4 — Music + finish
~0–10 cr
Upbeat, beat-forward track (licensed library first; Seed Audio fallback). Master 9:16 1080×1920, ~15s, and cut a 4:5 crop for feed. QA pass: identity check against charsheet, text legibility at thumbnail size, silent-autoplay read-through.
Cost basis from the rank06 pilot: 2 cr per still, 24 cr per 6s H3 clip. Full plan lands around 190–250 credits; the Phase 1 spike alone is ~34.