render-food-product-sizzle

Assemble a food-product sizzle ad from a config: a wordless macro-tabletop photorealistic sizzle for a physical food / CPG product — tactile sunlit tabletop photography in a warm tungsten kitchen register, ~4 dynamic macro scenes (hands tearing, a flat lay, a partial-face bite, a box / pack hero) flowing into a static end card, carried by a non-diegetic acoustic music bed + a few diegetic SFX with NO voiceover. This capability is the FREE, deterministic assembly — normalized concat, the anti-AI grain pass, the audio (music bed + SFX) composite, the PIL end card, and the optional serif stat-callout pills.

scripts/config.example.json is the worked example (Lineage Provisions "Beef Sticks Sizzle", ~14s 1080×1920 9:16, ~4 macro scenes + a static end card); scripts/PIPELINE.md maps every config block to its source step and scripts/README.md documents the free assembly.

Run

This is the FREE, deterministic assembly stage — it spends nothing. The paid inputs are separate capabilities: ~4 photographic macro keyframes (create-image-fal, Nano Banana; the box / pack hero grounds on the real product PNG); one locked-off, anti-shake i2v clip per keyframe (create-video-fal, Seedance); and a non-diegetic acoustic / bluegrass bed (create-music-elevenlabs). Given the ~4 clips + the music bed + the diegetic SFX + the real logo PNG + the real product PNG, render-food-product-sizzle normalizes fps / SAR, concats the body clips, applies the anti-AI grain pass, composites the audio (bed + SFX at their cue points), composites the static PIL end card, burns the optional serif callout pills, and muxes → the master. Re-cuts reuse the existing keyframes / clips / music and cost $0.

Contract (the free assembly)

• Wordless — the music bed is the audio, no VO. A non-diegetic acoustic bed (no vocals) IS the bed; do not add a spoken voiceover. The brand name + claim land on the STATIC end card, never in the body.
• ~4 macro scenes, concat in order. Normalize fps / SAR across the body clips and concat them in their scene order (tear → flat-lay → bite → box-hero by default); the box / pack hero shows the REAL label (grounded on the product PNG upstream — the assembly must not re-render it).
• Anti-AI grain pass, applied globally. Apply eq=contrast=1.06:saturation=0.93,hqdn3d=1.5:1.5:3:3,noise=alls=8:allf=t+u across the whole video — the noise on a food macro is load-bearing for the tactile / photographic read, otherwise the sizzle looks AI-smooth.
• Diegetic SFX on the tactile beats. Mix a crisp ~120ms snap on the fiber tear and a ~180ms tear on the box-open at their measured cue points — a couple of short hits, not a wall of sound. Time each to its beat, not a round number.
• Music bed with no sparse intro. The supplied / generated bed opens sparse — the upstream step trims the ~2.5s intro so it kicks in from frame 0; the assembly loudnorms + fades in / out to the master length.
• Static end card via PIL from the real logo + product PNG — never AI-render brand text. Solid / ivory bg + the real logo PNG (upper third) + the real product PNG (centered, soft shadow) + a serif heritage headline + a CTA, held ~3s WITH the music still playing under it (fade the tail — no silent tail). A diffusion model garbles a wordmark and the packaging. On macOS pick a serif with the middle-dot glyph (use ·).
• Optional serif stat-callout pills at beats. Ivory-or-brand-color pill + serif type at choreographed windows. Write any % string to a textfile and use ffmpeg drawtext textfile= + expansion=none — a raw % is read as a strftime spec and renders garbage.
• FFmpeg composite, deterministic, FREE. Concat the body clips, grain-pass, composite the audio (bed + SFX), append the PIL end card, burn the callouts, mux with a fade tail, loudnorm → a 1080×1920 h264+aac master (~14s). No paid calls, no keys.