render-glassy-matte-grwm

Assemble a multi-scene GRWM beauty-demo ad from a config — a locked-identity creator applies ~5 makeup/skincare products step by step in a chosen setting, a separate ElevenLabs voiceover narrates the routine, and every scene cut is snapped to the VO's product-name word-starts, with ~5 Playwright product overlay cards on the product-name beats, a ducked music bed, burned captions, and a flat-lay end card. This capability is the FREE, deterministic assembly — the Whisper-driven re-cut + hard-concat, the Playwright card render + card composite, the VO + music mix, the caption burn, and the flat-lay end card.

This is the multi-scene beauty demo, distinct from the single-take apparel outfit-reveal (ugc-grwm, one Seedance reference-to-video call with native lip-sync and minimal post). Here the timeline is driven by a SEPARATE VO and the scenes are re-cut to its word-starts.

scripts/config.example.json is the worked example (Fernwick Skin "5-Step Glassy Matte Skincare Routine", ~32s 1080×1920 9:16, 12 VO-snapped cuts + 5 product cards); scripts/PIPELINE.md maps every config block to its source step and scripts/README.md documents the free assembly.

Choices

The creative calls come from the recipe's choices, asked of the user before any paid step. The demo's picks are examples, never defaults. This capability only assembles what those choices produced; it hardcodes none of them.

• creator — who does the routine (assets.creator_anchor, the look in the scene beats). The demo used a young woman.
• voice — the VO voice (vo.voice_id / vo.voice_name). The demo used ElevenLabs Karri.
• setting — where the routine happens (assets.vanity_world). The demo used a sage-green vanity.
• tone — how the VO talks (vo.script, vo.style). The demo was upbeat and friendly.
• music — the bed under the VO (music.prompt; "none" = VO only, skip the music mix). The demo used a warm acoustic bed.

Card colours come from palette (the brand kit), not the demo's pale sage + fern green.

Run

This is the FREE, deterministic assembly stage — it spends nothing. The paid inputs are separate capabilities — the SEPARATE narration VO (create-music-elevenlabs, or a user-supplied mp3; word-level Whisper timestamps set the timeline), ~7 Seedance scene clips one per product step (create-video-fal), the ~5 white-bg product cutouts + the flat-lay end-card still (create-image-gpt-image-fal), and the ducked music bed. Given the VO + .words.json + one clip per step + the ~5 product cutouts + the music bed, render-glassy-matte-grwm re-cuts each clip to its VO word-start window, hard-concats on the cut, renders + composites the product cards on the product-name beats, mixes the VO over the ducked music, burns the captions, and appends the flat-lay end card → the master. Re-cuts reuse the existing VO / clips / cutouts and spend nothing.

Contract (the free assembly)

• The SEPARATE VO drives the timeline — Whisper it first. The narration is a separate track (not a native take). Its word-level timestamps set every cut; the atempo'd VO ends shorter than the plan expects (a 1.15× VO landed ~27.5s), so time every window to the word-starts, never to a pre-planned grid.
• Scene cuts snap to the "step N" word-start; cards snap to the product-NAME word-start. Cut to the next product when its step is announced; the card animates in ~1s later when the NAME is spoken. Both happen. ~12 cuts over ~32s (cuts/10s ≈ 3.75).
• Hard-concat with a re-encode. Hard cuts on the VO word-starts, no dissolves; re-encode the concat -c:v libx264 -crf 20 — -c copy corrupts the duration when zoompan/PNG clips are in the chain.
• Product cards — Playwright, real cutout, PDP-verified tagline. Playwright renders the card template at 2× scale (real white-bg cutout thumb + name + PDP tagline). The cutout must match the REAL product, not the packaging shape a Seedance scene hallucinates; the tagline is verified against the brand PDP (AI flat-lays hallucinate sublines). Composite each card onto the master snapped to its product-NAME word-start, 1s fade-in, held until the next product is named. PNG overlay inputs need -loop 1 -t <dur> — without it the PNG emits one frame at t=0 and the fade/enable filters silently no-op (cards go invisible).
• VO leads the ducked music bed. Mix the SEPARATE VO on top of the ducked music (the VO is the lead), loudnorm I=-14. When choices.music is "none", loudnorm the VO alone. If the host ffmpeg lacks a filter, apad/atrim to length before the mix.
• Captions — clean-white, override the preset. Clean-white captions from the VO's Whisper words, overridden to 3 words/cue, ~3.0% font, ~20% margin, NO pill, NO shadow (the default 5-words/4.5%/18% reads too dense). Burn last. If the host ffmpeg lacks libass, render the cues as timed PIL PNG overlays composited with ffmpeg overlay=…:enable='between(t,st,en)' at the same placement.
• Flat-lay end card. Append the flat-lay still (ken-burns hold ~4s) — a gpt-image-2 flat-lay of the ~5 products; do NOT trust its AI-rendered sublines for the card taglines.
• FFmpeg composite, deterministic, FREE. Re-cut, hard-concat, render + composite the cards, mix the VO over the ducked music, burn the captions, append the end card → a 1080×1920 30fps h264+aac master (~32s). No paid calls, no keys.