render-multiworld

Assemble a silent, music-led "multi-world product tour" ad (≈27s, 9:16) — a tour of three distinct "third-place" worlds (chosen per video), one per product/variant, that lands on a Pinterest-style brand end card. Each world is a two-shot pair: a ~4.5s WIDE kinetic-calm ARRIVAL (the environment dominates, the product stays small) hard-cutting to a ~3.5s top-down MACRO product MOMENT (the sealed product nested with its companion prop). Variant identity is carried by the world + companion prop, not by product color. No VO, no captions in the scenes — one music bed carries the whole thing.

This capability is the FREE, deterministic assembler. The paid steps — the six per-world clips, the AI flat-lay end-card background, the ElevenLabs music bed — are separate capabilities (see the gap below for the clips); the recipe orchestrates and gates them.

Choices

The calling recipe asks the user these before any paid step; this capability only assembles what they produced. The worked example (the fictional brand Kettle & Fern) values are examples, never defaults.

• worlds — the three places toured → config.worlds, config.scene_grid (demo: reading nook / glasshouse / farmers market).
• visual_style — the worlds' look → config.visual_style, worlds[*].palette, end_card.nb2_prompt (demo: sunlit-natural / handmade).
• music — the bed's genre/mood → config.music.prompt (demo: warm acoustic-pop).

Nothing in the assembly keys on world slugs or names: clips are ordered by scene_grid[].scene, and the end-card copy is whatever end_card.overlay holds.

Run

1. Trim clips (FFmpeg, FREE) — trim each per-world clip to its scene_grid[].duration_sec (arrival 4.5 / macro 3.5), re-encode to the master spec (720×1280, 24fps, yuv420p, scale+pad, audio stripped). Trimming the macro so the top-down portion dominates also hides any label misrender at the clip's upright tilt extreme.
2. End card (Playwright/HTML, FREE) — screenshot end_card.html over the AI flat-lay BACKGROUND, then FFmpeg-encode to a dwell_sec (3.0s) static clip. Headline (overlay.headline, approved brand copy — the demo's was "STEEP SLOWER."; Inter 900), one handwritten Caveat product label + hand-drawn SVG arrow per product, Playfair wordmark + URL. End-card text is HTML, NEVER AI-rendered — the AI step produces the background only.
3. Hard-cut concat (FFmpeg, FREE) — concat the six trimmed clips in scene order (S01 arrival → S02 macro → … → S06 macro) + the end-card clip. Hard cuts (no dissolves), normalized to one fps/codec first so concat-copy is safe.
4. Music mux + web encode (FFmpeg, FREE) — mux the single instrumental bed onto the silent concat with afade in/out + loudnorm I=-16:TP=-1.5:LRA=11, AAC 192k, clamped to 27.0s, explicit single-audio map so no silent scene-track leaks in → the H.264 (+ AAC) 720×1280 master.

Contract

• Deterministic + FREE (Playwright + PIL/HTML + FFmpeg); no paid calls, no AI-rendered end-card text. Iterate the cut for free by re-running the assembly.
• Silent, music-led — no VO, no captions during the scenes; on-screen text appears ONLY on the end card, HTML-composited.
• Per world = WIDE arrival (product small, environment dominant) → top-down macro (product + botanical, no hands). Hard cuts between scenes; end card is a static hold with legible HTML text; music starts at t=0 and fades the tail.
• Sealed-product rule is enforced at QC — trim/concat can't fix an opened or spilling clip; that's a re-roll (paid), not an assembly fix. Identity via world + botanical, never pack color.
• The template recipe (DB) supplies the per-video config (worlds, palettes, companion props and music mood from the user's choices; prompts; end-card copy from the brand kit); this capability is the generic assembler.

Gaps / routing notes

• Update: the current recipe makes the six clips with create-image-fal (nano_banana_2 keyframe edit grounded on the real product_image) → create-video-fal (Kling 3.0 i2v), so the gap below only applies to reproducing the demo's Higgsfield build exactly.
• The demo's clips used a Higgsfield-proxy capability that does not exist. The source molecule fires them through Higgsfield Marketing Studio (marketing_studio_video/product_showcase) grounded on imported product UUIDs, with the sealed-product safety block front-loaded on every prompt. create-video-fal is a FAL i2v proxy — a different provider and job shape — so it cannot serve this step. Wiring the clip step to templates-as-data requires a Higgsfield video proxy (Marketing Studio product_showcase, imported-product grounding, per-prompt safety block) that does not exist yet. Until it lands, the clips are generated via the Higgsfield Marketing Studio path directly (CLI/MCP) and that step is not a fetchable capability.
• The end-card background is an AI flat-lay render (text-free, all three sealed products + botanicals) that in prod routes through create-image-fal (NB2); only the HTML text overlay + encode run locally as FREE assembly.
• The music is a paid ElevenLabs call that in prod routes through create-music-elevenlabs; only the loudnorm + fade + mux run locally as FREE assembly.

See scripts/PIPELINE.md for the full config-field → source-step map and scripts/README.md for the FREE-assembly detail.