render-cgi-sizzle

Assembles the 3D-CGI app sizzle video format: a gold-trimmed phone floats in a smoky-black studio while six app features demo one per beat — each beat bursts REAL App Store UI elements out of the phone in 3D with amaranth rim-light + bokeh, then everything collapses back into the screen at a climax ("200+ classes. One app.") + a brand end card over a premium-tech bed. It reads as an Apple-keynote product film, not UGC and not a physical-product shoot.

This capability ships the recipe (scripts/config.example.json) + a config→step map (scripts/PIPELINE.md) + the FREE-assembly how-to (scripts/README.md). It documents the FREE, deterministic assembly between the paid model calls:

• nano-banana CGI plates (paid, separate cap) render only the phone shell + smoky-black studio + amaranth rim-light + bokeh + placeholder burst shapes; the screen is left a blank warm glow on purpose, and every plate is --anchored on beat 1 so the phone/studio stay identical across beats.
• PIL real-UI compositing (FREE) — auto-detect the bright phone-screen bbox in each plate and feather the REAL App Store screenshot into the bezel → scene-NN-composite.png; bake real burst-out overlays (e.g. climax instructor portrait tiles, rim-light baked before rotation) around the plate. The on-screen UI + faces + wordmark are ALWAYS real assets — never AI-rendered, so no claim is invented and nothing reads fake.
• Kling 3.0 i2v steady-float (paid, separate cap) drives each composite; the burst-out pops/settles while the phone + screen stay locked. Any beat Kling garbles drops to a FREE Ken-Burns FFmpeg push-in (zoompan, heavier on the climax) — the shipped demo used this path for the feature beats.
• Assembly + finalize (FREE) — dice + intercut concat (timeline locked from the measured VO durations), audio mix (sidechain-duck the music under VO, loudnorm -14 LUFS master), PIL brand end card (real wordmark, never AI), then 1.15x speed + anti-AI grain master.

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

Run

Config-and-PIPELINE capability (no re-built runnable pipeline here). Copy scripts/config.example.json → config.json, edit the brand/beats/screens, and follow scripts/PIPELINE.md:

VO first (locks the timeline) → nano-banana CGI plates → PIL screen composites → burst-climax overlays → Kling i2v clips (Ken-Burns fallback per garbled beat) → PIL end card → captions → sidechain-ducked mix → 1.15x speed + grain.

Output: 1080x1920, ~22.6s H.264 (+ AAC music). 6 feature beats + PIL end card.

Contract

• REAL UI, always PIL — never AI. Every app screen, instructor face, and the wordmark is the real asset composited via PIL. AI plates only ever render the phone shell, studio, bokeh, and placeholder burst shapes. This is the format's whole credibility and the guard against invented claims.
• Kling for the float, Ken-Burns fallback per beat. If a Kling beat garbles the burst-out UI, distorts the phone, or animates the screen, fall that beat to a FREE Ken-Burns push-in — never ship a garbled beat.
• Anchor every plate on beat 1 so the phone/studio read identical across beats (one shoot).
• Timeline locked from the measured VO durations, never planned word counts.
• The paid steps — nano-banana plates, Kling 3.0 i2v beats, ElevenLabs VO + music — are separate capabilities (create-image-fal, create-video-fal, create-vo-elevenlabs, create-music-elevenlabs); the recipe orchestrates them and gates the spend.