Replicating ChillFit's #6 ad, frame for frame — with Roxie in the chair

Overview — one variable, everything else pinned

A perfect replication is only meaningful if we change one thing. The controlled variable is the performer; every other element is either rebuilt to match the original exactly or reused verbatim from the source file.

Source anatomy — the full timeline

Measured with scene detection (threshold sweep at 0.10/0.25) plus frame sampling at the boundaries, and an ElevenLabs transcription of the audio track (language: nld).

Voiceover (full transcript, ends 20.8 s; music continues to 31.4 s)

"Ik ben 65. Ik doe deze work-out elke dag, gewoon 7 minuten per dag en na één week zul je een verschil opmerken. Na twee weken zullen anderen het opmerken en na drie weken zullen al je vrienden vragen hoe je het hebt gedaan. Het geheim? Download deze app."

Before assembly I'll re-run the transcription with --full to get word-level timestamps, and snap every caption in/out to the spoken word — the original's captions track the VO within ~0.1 s.

00.00 – 04.20Act A — the 3×3 grid hook

Nine simultaneous portrait panels, all the same set and framing, each a different chair move. Caption band sits between panel rows 1 and 2. This is the highest-effort shot to fake and the strongest test of character consistency: nine Roxies on screen at once.

So the grid costs only four dedicated extra clips (G1–G4); five panels are centre-crops of the Act B masters at staggered start offsets (and one mirror) so no two panels read as the same footage. Each panel is a 1:1.78 portrait crop scaled to 120×213 in the 360×640 master — small enough to be forgiving of minor artifacts, which is why the grid clips are the right place to absorb the weaker takes.

04.20 – 20.73Act B — the workout montage, five locked shots

One camera setup, five jump cuts. Every clip below is generated at 6 s and trimmed to its slot; generating a hair long gives trim room to pick the cleanest motion cycle. Shared prompt preamble for all five (and G1–G4):

Locked-off tripod shot, vertical 9:16, no camera movement. A fit woman in her
fifties with long wavy blonde hair, wearing a dark sage-green seamless workout
unitard and white crew socks, sits centred on a light wooden spindle-back chair
in front of white paneled closet doors with a black lever handle, light grey
carpet. Soft even daylight. She smiles naturally while exercising. Amateur
iPhone footage look, slight compression, true-to-life motion, no slow motion.

20.73 – 31.44Acts C–D — the app-UI tail

Act C · 28-day challenge cards (20.73 → ~28.3)

Parchment-textured page, header "28-DAAGSE STOELYOGA UITDAGING" in a heavy rounded serif (Cooper Black class), stats bar LENGTE 168CM · GEWICHT 100KG · DOEL 70KG (values red, goal green). Below, a 4×7 sketch calendar of seated exercises. A hand with red nail polish taps day 1 → the DAG-1 card zooms in: a live-action video of a heavier woman in a plum workout set doing seated knee lifts, three Dutch exercise lines, a cyan progress border animating around the card, four sketch thumbnails at the bottom. Back to the calendar, tap → DAG-11 card: the same woman visibly slimmer. The before/after arc lives inside these two card videos.

If rebuilt: the page is an HTML/CSS animation captured at 30 fps (parchment, calendar, zooms, cyan border), the tapping hand is a cut-out PNG animated in CSS, and the two inner card videos are generated — a softened, heavier "day-1 Roxie" variant and today's Roxie for day 11. That heavier variant is an identity edit of the charsheet (Nano Banana) used as the start frame.

Act D · End card (~28.3 → 31.44)

Magenta-to-pink gradient, three fanned iPhone mockups of the ChillFit app (plan list, workout player at 00:08, calories screen), ChillFit icon + script wordmark, "Planner voor thuistraining", App Store / Google Play badges. Contains no human.

Graphics kit — rebuilding every overlay

Every overlay is rebuilt at 720×1280 (2× the source) so the master looks clean, then the delivery encode downsamples to the source's exact 360×640 for honest A/B comparison.

Audio — English VO in Roxie's voice over the source music bed

Revised 2026-08-28. The original scope reused the Dutch track verbatim, but the direction has shifted toward a usable 50Queen asset: captions are now English and the ChillFit watermark is dropped, so a Dutch VO no longer fits. The new audio recipe:

1. VO — Roxie's own voice. Her character sheet defines a Higgsfield TTS preset (qwen_audio_tts, preset "Roxie", f6448975…, English, "cheerful upbeat morning coach"). Script = the English translation of the Dutch VO, one line per caption: "I'm 65. I do this workout every day — just 7 minutes a day, and after one week you'll notice a difference. After two weeks others will notice, and after three weeks all your friends will ask how you did it. The secret? Download this app." Generated per-line so each line can be nudged to fit its shot's slot.
2. Music bed. The source track has VO and music mixed; run source separation (demucs, local, free) on the 31.44 s track to recover the instrumental, and lay the new VO over it. Fallback if separation artifacts: any upbeat library track at the same BPM — the music is generic.
3. Caption timing. Captions word-time to the new English VO (ElevenLabs transcription of our own track), replacing the plan's earlier snap-to-Dutch rule.

Purist fallback, still one command away: mux the untouched Dutch track for a strict A/B-comparison cut. Both cuts can share all video.

Generation plan — models, strategies, pilot

Model & constraints

• Video — MiniMax H3 (minimax_h3, 2K, ~4 cr/s: 24 cr per 6 s clip). Chosen per prior research: Seedance 2.0 reads too polished for UGC-style ads, and H3 was queued as the next candidate. Its param sheet exposes start_image/end_image, image_references and video_references — with one hard constraint: start-frames cannot be combined with reference media. That constraint forces the strategy split below.
• Stills — Nano Banana 2 (2 cr/img) for all start frames and identity edits, anchored on the Roxie charsheet.

The locked recipe, per clip

higgsfield generate create minimax_h3 \
  --prompt "Replicate the reference video exactly: same locked camera framing, same
  rhythm, same choreography of <shot motion, with cadence> — performed by the woman
  from the reference images: early fifties, long wavy blonde hair, dark sage-green
  seamless workout unitard, white crew socks, smiling warmly throughout. Same room:
  white paneled closet doors, black lever handle, light wooden spindle-back chair,
  grey carpet. Amateur iPhone look. No text, no captions, no watermarks." \
  --video-references <source segment upload> \
  --image-references <charsheet> --image-references <cand2 still> \
  --duration 6 --aspect-ratio 9:16

Full-run clip manifest (motion transfer throughout)

Grid-clip caveat: G1–G4's video references are 120×213 panel crops upscaled 4× — much weaker motion signal than the full-frame W refs. If transfer degrades, the fallback is prompted motion (S2-style) for those four: the panels render at 120×213 in the final grid, small enough to forgive tamer motion. Decide per clip at review, not up front.

Assembly pipeline

1. Normalize — every generated clip → 720×1280, 30 fps, trimmed to its slot (best motion cycle chosen by eye).
2. Grid — 9 panel crops → xstack layout=3x3 on white, 4.20 s.
3. Body concat — A + W1…W5 cut at exactly 4.20 / 7.17 / 10.73 / 13.73 / 17.60 / 20.73 s.
4. Overlays — watermark PNG → muscle-strip PNGs (per-shot enable windows) → icon bug at 19.0 s → ASS captions burned last so outlines sit above everything.
5. Tail — splice source 20.73–31.44 s (re-encoded to match the master's encode so the joint is invisible), or the rebuilt Act C/D if that option is chosen.
6. Audio — mux the untouched source track; hard-trim master to 31.44 s.
7. Deliverables — remake_720.mp4 (master), remake_360.mp4 (source-matched), sxs_compare.mp4 (original | remake hstacked, shared audio).

Verification — how we score "perfect"

• Side-by-side sync cut — the hstacked comparison video is the primary artifact of the whole experiment; every review happens against it.
• Cut-sheet diff — paired stills at every cut point ±0.1 s, original over remake, one contact sheet.
• Caption timing — every caption in/out within 0.1 s of the original (word-timestamp check, scripted).
• Identity hold — Roxie recognisably the same person across all 11 clips and 9 simultaneous grid panels; the grid frame is the acid test.
• Set hold — chair, doors, handle position, carpet identical across shots (locked-camera illusion preserved).
• Scorecard — per-shot 1–5 on motion / identity / set / texture, logged in the evidence folder with the generation IDs, so the writeup can say precisely where current models succeed and fail.

Cost & approvals — full run (v2, 2026-08-28)

Spent so far: 56 cr (4 stills + the two-arm W1 pilot). Settled along the way: strategy = motion transfer, wardrobe = sage unitard, no ChillFit watermark, captions English. Remaining spend to a finished human section:

Run order

1. Batch 1 — W2–W5 (96 cr): four motion-transfer clips off the locked recipe, reviewed as side-by-sides against their source segments before the grid spends anything.
2. Batch 2 — G1–G4 + W1 redo (~120 cr): grid clips (with the upscaled-crop caveat checked clip by clip) plus the W1 redo if batch 1's fixes prove out.
3. Finish pass (no video credits): per-shot muscle strips via the extraction recipe, English captions, VO + demucs music bed, grid xstack, body concat at the source cut points, tail splice, 360×640 delivery encode + side-by-side master.

Open decisions