render-vignette

Assemble a short-form 'vignette' ad from clean product cutouts composited over a kinetic background video. Motion lives in the BG video; the product rides on top as a static cutout layer. Music-led, zero VO, sub-12s, loopable — it reads muted because the product label + on-screen copy carry the message. Defaults to the V-CARD structure: a cold-open text card → a product carousel under one shared BG → an annotated specimen-sheet end card.

Run

1. strip_product_backgrounds.py — birefnet cutout of each PDP shot to clean hard-edge alpha (no halo/shadow).
2. render_overlays.py — PIL + rsvg-convert render the cold-open card (Boska Black, dead-center) + the annotated specimen-sheet end card (brand SVG logo + Space Grotesk annotations) as transparent 1080x1920 PNGs. FREE.
3. composite_variants.py — one FFmpeg filter_complex per BG variant: BG (palette-aware dim) → cold-open overlay → cutouts (width-anchored, vertically centered y=(H-h)/2) → end card. h264 crf20 yuv420p +faststart 30fps. FREE.
4. music_and_mux.py — instrumental music bed → acompressor + loudnorm I=-18:TP=-2:LRA=9 → muxed into every variant in a SEPARATE pass with explicit -map 0:v:0 -map 1:a:0. The mux is FREE; the music generation is a paid call that in prod routes through create-music-elevenlabs.

Contract

• FREE assembly: birefnet cutout (see gap below) + PIL/rsvg overlays + FFmpeg composite + mux. No AI-rendered text; the product art/labels and on-screen copy are real, never invented.
• The template recipe (DB) supplies the per-brand config (products, cold-open text, end-card lines, BG concept, beat timing). This capability is the generic assembler.
• Craft rules preserved from the source molecule:
  • Cutouts stripped clean (no halo/shadow), height-anchored at vertical-center (y=(H-h)/2) so mixed-shape SKUs share one visual mid-line — never bottom-anchor (squat jars jump). For 9:16 scale by WIDTH (~75% tall bottles, ~65% squat jars).
  • Palette-aware BG dim: high-contrast/chrome BG → push saturation DOWN hard (saturation=0.50); naturally-contrasty BG → lighter dim (saturation=0.85).
  • End card = annotated specimen-sheet (EST year + rule + wordmark + rule + ingredient + positioning + claim), never a bare logo. Use the WHITE logo variant on dark BGs, cream on light.
  • Music-led, NO VO — instrumental only (VO/lyrics would fight the cold-open + end-card text). Loudnorm before the mux.
  • Mux is a SEPARATE FFmpeg pass with explicit -map 0:v:0 -map 1:a:0 (single-pass composite+mux silently ships 1 kbps garbage audio).

Gaps / routing notes

• birefnet is a paid FAL model, not free. strip_product_backgrounds.py calls fal-ai/birefnet/v2 directly via a fal_helpers shim (sys.path.insert into a shared atoms dir + from fal_helpers import ...). It is bundled here because the cutout is an intrinsic assembly step, but in prod the birefnet cutout should route through create-image-fal (the fal-proxy capability that bills the Ads agent) rather than hitting fal.run directly. Treat the direct-fal path as a gap to close; the recipe gates the cutout to gooseworks fetch create-image-fal.
• Background sourcing is not in this capability. The kinetic BG is sourced upstream — PEXELS-FIRST (free stock, the key cost lever), falling back to T2V (create-video-fal) only when stock coverage fails — and dropped into source/t2v-outputs/<slug>.mp4 so the composite is source-agnostic. There is no Pexels fetcher bundled here; that lives in the recipe's playbook.
• music_and_mux.py also uses the fal_helpers shim for the ElevenLabs music generation; in prod that generation routes through create-music-elevenlabs, and only the loudnorm + separate-pass mux run locally as FREE assembly.