{
  "_comment": "Lineage Provisions 'Beef Sticks Sizzle' food-product sizzle — THE WORKED EXAMPLE. Copy to config.json and edit. A ~14s warm-tungsten macro-tabletop 9:16 sizzle for a physical grass-fed beef-stick product: WORDLESS (no VO), driven by a non-diegetic bluegrass bed + diegetic snap/tear SFX. Pipeline: (1) ~4 photographic macro keyframes via create-image-fal (the box hero grounds on the REAL product PNG); (2) one locked-off anti-shake create-video-fal i2v clip per keyframe (tear / flat-lay / bite / box-hero); (3) an ElevenLabs Music ~22s bluegrass bed trimmed of its 2.5s sparse intro + two ffmpeg-lavfi diegetic SFX; (4) a STATIC PIL end card (real Lineage logo PNG + real product PNG + serif headline + CTA) + ffmpeg filter_complex concat + anti-AI grain pass + audio composite + 5 serif stat-callout pills. Master 1080x1920 h264+aac, ~14.04s (24fps). See scripts/PIPELINE.md for the source scripts.",
  "brand_name": "Lineage Provisions",
  "campaign": "Beef Sticks Sizzle",
  "title": "Grass-fed. Air-dried. Nothing else.",
  "width": 1080,
  "height": 1920,
  "fps": 24,
  "aspect_ratio": "9:16",
  "duration_sec": 14.04,
  "look_pack": {
    "name": "WARM_TUNGSTEN_KITCHEN",
    "palette_anchors": [
      "ivory #F6F1E7 background tone",
      "ranch brown #3B2A1E shadows",
      "dried-blood red #7A1F1F on the beef",
      "honey-gold highlights",
      "warm tungsten window daylight, slightly desaturated"
    ],
    "style_descriptor": "Tactile macro-tabletop food photography, warm tungsten kitchen register, linen + wood + butcher paper + cast iron surfaces, warm window daylight raking from camera-left, slight 35mm film grain. Photographic, NOT cinematic, NOT smooth. iPhone-shot daytime kitchen feel, NOT studio polish. Sharp focus on the product. 1080x1920 vertical.",
    "negative_tail": "NO studio polish, NO bokeh, NOT shallow DOF, NO smooth digital 3D render, NO AI-illustration look. NO invented label text, NO captions, NO brand wordmarks rendered by the model (the brand appears only on the composited end card). NO faces beyond a partial jaw / hands crop. Vertical 9:16 composition.",
    "_note": "ONE look pack drives all 4 keyframes so the sizzle reads as one spot. Palette sampled from the Lineage packaging + site (brand-research/visual-identity.md). The 'no invented label text' negative is load-bearing — the box hero grounds on the real product PNG so the label is real."
  },
  "keyframe_engine": {
    "atom": "create-image-fal",
    "aspect_ratio": "9:16",
    "product_ref": "S04 (box hero) passes the real meat-sticks-classic.png (+ the variety pack) as media[role=image] with 'reproduce the label EXACTLY, invent no text'; S01/S02/S03 pass no product ref",
    "_note": "See scripts/.hf_S04.json — it threaded the real box + variety-pack PNGs as media_input and instructed the model to reproduce the label exactly. NEVER include readable brand text in a prompt — the model garbles it into typos."
  },
  "clip_engine": {
    "atom": "create-video-fal",
    "duration_sec": 4,
    "resolution": "720p",
    "aspect_ratio": "9:16",
    "submission": "sequential",
    "anti_shake_tail": "Camera is locked off, tripod-stable. NO shake, NO wobble, NO earthquake. Slow, deliberate motion only.",
    "_note": "Locked-off, anti-shake. Submitted SEQUENTIALLY (gen_clips_hf.sh) to avoid the burst-credit reserve gate. Per-scene durations match tStart/tEnd — S01/S02 3s, S03/S04 4s."
  },
  "music": {
    "atom": "create-music-elevenlabs",
    "provider": "elevenlabs",
    "model": "music_v1",
    "request_length_ms": 22000,
    "trim_intro_sec": 2.5,
    "target_length_sec": 17.04,
    "fade_in_sec": 0.4,
    "fade_out_sec": 1.0,
    "loudnorm_i": -16,
    "prompt": "Upbeat warm bluegrass instrumental. Picked banjo as the lead voice with a supportive flat-picked acoustic guitar and very light brush percussion. Friendly Americana, ranch-honest, heritage food-brand register. NO vocals, NO fiddle solo, NO heavy drums. Mid-tempo, major key, walking forward groove. The banjo and guitar should start IMMEDIATELY with a clear rolling phrase — do NOT begin with silence or a long sparse intro. Builds slightly in energy toward the end with one tasteful banjo flourish around the 16-second mark.",
    "force_instrumental": true,
    "outputs": {
      "raw": "audio/music_raw_v2.mp3",
      "bed": "audio/music-bed-v2.mp3"
    },
    "_note": "gen_music_v2.py requested 22s to allow the 2.5s sparse-intro trim; build_master_v2.sh step 1 trims start=2.5, loudnorm I=-16, fade in 0.4s + fade out 1.0s, clamps -t to the FULL master length. target_length_sec MUST equal duration_sec (body, 14.04) + end_card.hold_sec (3.0) = 17.04 so the bed plays under the end card and fades on the tail — audio_mix.music_over_end_card is true, so a 14.04 bed would leave the 3s end card silent. No artist names (ElevenLabs Music ToS filter)."
  },
  "sfx": {
    "engine": "ffmpeg-lavfi-anoisesrc",
    "script": "make_sfx.sh",
    "cues": [
      { "name": "snap", "wav": "audio/snap.wav", "dur_ms": 120, "cue_t": 1.5, "note": "crisp white-noise crack at the fiber tear (S01), highpass 900 / lowpass 5500, acompressor" },
      { "name": "tear", "wav": "audio/tear.wav", "dur_ms": 180, "cue_t": 11.0, "note": "pink-noise pull at the box-open / hand reach (S04), highpass 350 / lowpass 4000" }
    ],
    "_note": "Two short diegetic hits synthesized with ffmpeg lavfi anoisesrc (make_sfx.sh) at measured cue points."
  },
  "scenes": [
    { "id": "S01", "role": "MACRO_TEAR", "t_start": 0.0, "t_end": 3.0, "uses_product_ref": false, "keyframe_prompt": "Slow-motion macro: two weathered hands hold a dark grass-fed beef stick at a break line on butcher paper, coarse meat fibers visible, a few crumbs. Warm window daylight from camera-left.", "motion_hint": "The two weathered hands tear the dark grass-fed beef stick apart; real coarse meat fibers stretch and separate at the break line; a few crumbs fall straight down with gravity; hands stay in frame, daylight steady from camera-left." },
    { "id": "S02", "role": "FLAT_LAY", "t_start": 3.0, "t_end": 6.0, "uses_product_ref": false, "keyframe_prompt": "Top-down flat lay of three beef sticks fanned on butcher paper next to a single sprig of rosemary and flaky salt. Warm daylight, slight film grain.", "motion_hint": "Gentle slow 2% top-down push-in on the flat lay; a tiny breath of air moves one rosemary leaf slightly; salt grains catch the light; otherwise still life." },
    { "id": "S03", "role": "BITE", "t_start": 6.0, "t_end": 10.0, "uses_product_ref": false, "keyframe_prompt": "Side profile, ONLY the jaw and lower lip in frame (eyes and forehead cropped OUT), a man about to bite a beef stick. Background defocused warm ivory-brown, consistent stubble shadow.", "motion_hint": "The man bites cleanly into the beef stick — a clean tear, then a slow chew; background stays defocused warm ivory-brown; stubble shadow stays consistent." },
    { "id": "S04", "role": "BOX_HERO", "t_start": 10.0, "t_end": 14.04, "uses_product_ref": true, "keyframe_prompt": "Locked medium shot. The EXACT Lineage Provisions 100% Grass-Fed Beef Sticks Original box (from the reference PNG) stands upright on a light oak kitchen counter with a linen runner. A man's hand enters from the right edge, fingers reaching to lift one beef stick out. A second variety pack box (from the reference) sits in soft focus behind. Reproduce the box label EXACTLY as in the reference — invent no text, no captions, no stickers.", "motion_hint": "The hand enters from the right, fingers gently grasp one beef stick and lift it slowly straight up out of the box; the box itself stays put; the variety pack stays still in soft focus behind; daylight from camera-left holds steady." }
  ],
  "end_card": {
    "engine": "pil",
    "script": "make_end_card.py",
    "background": "#F6F1E7",
    "brand_logo_png": "brand-assets/logo/lineage-logo.png",
    "product_png": "brand-assets/products/meat-sticks-classic.png",
    "logo_scale": 0.62,
    "product_scale": 0.62,
    "headline": "Grass-fed. Air-dried. Nothing else.",
    "headline_font": "Times New Roman / Georgia (serif heritage)",
    "cta": "30% off your first subscription  ·  lineageprovisions.com",
    "hold_sec": 3.0,
    "_note": "make_end_card.py composes scene-S05.png entirely in PIL — ivory #F6F1E7 bg + the real Lineage wordmark PNG (upper third) + the real meat-sticks-classic product PNG (centered, soft shadow) + the serif headline + the CTA. NEVER AI-rendered. The demo's Memorial-Day promo added the '30% off' tag. Uses a middle dot ' · ' (macOS serif em-dash gap)."
  },
  "stat_callouts": {
    "enabled": true,
    "style": "serif-pill",
    "font": "Georgia Bold / Big Caslon (serif)",
    "_note": "The demo burned 5 serif stat-callout pills (ivory or dried-blood-red pill, serif type) at choreographed beats via build_master_v2.sh drawtext. % strings ('100% GRASS-FED', '30% OFF') were written to textfiles (c1.txt / c5.txt) with expansion=none — drawtext reads a raw % as a strftime spec.",
    "callouts": [
      { "text": "100% GRASS-FED + GRASS-FINISHED", "t_start": 0.6, "t_end": 2.4, "pill": "ivory" },
      { "text": "20G PROTEIN", "t_start": 2.7, "t_end": 4.7, "pill": "dried-blood-red" },
      { "text": "0 SEED OILS · 0 FILLERS", "t_start": 5.0, "t_end": 7.0, "pill": "ivory" },
      { "text": "ANCESTRAL RECIPE / 54 iterations · 2 years", "t_start": 7.5, "t_end": 9.5, "pill": "ivory" },
      { "text": "30% OFF FIRST SUBSCRIPTION", "t_start": 13.5, "t_end": 14.04, "pill": "dried-blood-red" }
    ]
  },
  "grain_pass": {
    "enabled": true,
    "filter": "eq=contrast=1.06:saturation=0.93,hqdn3d=1.5:1.5:3:3,noise=alls=8:allf=t+u",
    "_note": "Anti-AI-look grain applied globally in build_master_v2.sh so the sizzle doesn't read AI-smooth."
  },
  "audio_mix": {
    "loudness_lufs": -16,
    "duck_under_vo": false,
    "music_over_end_card": true,
    "audio_fade_out_sec": 1.0,
    "_note": "The bluegrass bed is the full audio bed — no VO to duck under. snap.wav (~1.5s) + tear.wav (~11s) mix in at their cue points. The bed plays UNDER the end card and fades out over the tail — no trailing silence."
  },
  "post_production": {
    "music": { "default": "on", "note": "non-diegetic bluegrass bed (no vocals), sparse intro trimmed, loudnorm + fades" },
    "sfx": { "default": "on", "note": "diegetic snap (fiber tear) + tear (box-open) at measured cue points" },
    "grain": { "default": "on", "note": "anti-AI-look grain (eq + hqdn3d + noise) applied globally" },
    "stat_callouts": { "default": "on", "note": "5 serif stat-callout pills at beats (protein / ingredients / provenance / promo)" },
    "end_card": { "default": "on", "note": "static PIL end card (real logo PNG + real product PNG + serif headline + CTA), music under it" }
  }
}

# Pipeline — food-product sizzle

How `config.example.json` maps to the real production steps. This ships a **config + this map**, not
a bundled runner: the worked example (Lineage Provisions "Beef Sticks Sizzle") was produced by
per-project scripts (`gen_keyframes` / `.hf_S0*.json`, `gen_clips_hf.sh`, `gen_music_v2.py`,
`make_sfx.sh`, `make_end_card.py`, `build_master_v2.sh`). Drive the whole run via a control-plane
video orchestrator, or reference those steps directly.

The five steps run **in order** because each depends on the last: the photographic keyframes seed
the i2v clips, the clips set the body length, the music bed + SFX are cut to the master, the PIL
end card composites from the real logo + product PNG, and assembly stitches all of it with the
grain pass + audio composite + optional callouts.

## Field → source-step map

| Config field | Phase | Source step / script | Paid? |
|---|---|---|---|
| `look_pack.style_descriptor`, `look_pack.negative_tail`, `look_pack.palette_anchors` | 1 Keyframes | threaded verbatim into every keyframe prompt | (defines cost) |
| `scenes[].keyframe_prompt`, `scenes[].uses_product_ref`, `keyframe_engine` | 1 Keyframes | `create-image-fal` macro stills, 9:16 (box hero grounds on the real product PNG) | **PAID** |
| `scenes[].motion_hint`, `clip_engine` | 2 Clips | `create-video-fal` i2v, locked-off anti-shake, 3–4s | **PAID** |
| `music.prompt`, `music.request_length_ms`, `music.model` | 3 Music | ElevenLabs Music ~22s acoustic bluegrass | **PAID** |
| `music.trim_intro_sec`, `music.target_length_sec`, `music.fade_in_sec`/`.fade_out_sec`, `music.loudnorm_i` | 3 Music | trim the 2.5s sparse intro, loudnorm, fades | free |
| `sfx.cues` | 3 SFX | ffmpeg lavfi `anoisesrc` → `snap.wav` / `tear.wav` at cue points | free |
| `end_card` | 4 End card | PIL — bg + real logo PNG + real product PNG + serif headline / CTA | free |
| `stat_callouts` | 4 Assembly | ffmpeg `drawtext` serif pill callouts at beats | free |
| `grain_pass`, `scenes[].t_start/t_end`, `audio_mix` | 4 Assembly | ffmpeg `filter_complex` concat + grain + audio composite + end card | free |

## 1. Keyframes → `create-image-fal`  (config: `scenes[].keyframe_prompt`, `look_pack`, `keyframe_engine`)  [PAID]

Per scene, build the prompt as `look_pack.style_descriptor` + `scenes[i].keyframe_prompt` +
`look_pack.negative_tail`, at 9:16. The box / pack hero (`uses_product_ref: true`) passes the REAL
product PNG as `media[role=image]` with "reproduce the box label EXACTLY, invent no text". Texture-only
shots (tear, flat-lay, bite) pass no product ref. The "NO studio polish, NO bokeh, NO invented label
text" negative is load-bearing. Review all before step 2.

## 2. Clips → `create-video-fal` i2v  (config: `scenes[].motion_hint`, `clip_engine`)  [PAID]

Per keyframe, build the prompt as `scenes[i].motion_hint` + `clip_engine.anti_shake_tail` and call
image-to-video off the clean keyframe, 9:16, at the scene's duration (3–4s). **Locked-off,
anti-shake** — every prompt ends with "Camera is locked off, tripod-stable. NO shake, NO wobble.
Slow, deliberate motion only." Submit **sequentially** (burst-credit reserve). The motion IS the
tactile beat: the tear parts the fibers, a gentle push-in on the flat lay, a clean jaw-line bite,
a hand lifting one stick straight out of the box.

## 3. Music + SFX → ElevenLabs Music + ffmpeg lavfi  (config: `music`, `sfx`)  [PAID music]

ElevenLabs Music requests ~22s of upbeat acoustic bluegrass (banjo + guitar + light brush
percussion, NO vocals, no artist names). Then **trim the ~2.5s sparse intro** (`atrim=start=2.5`),
`loudnorm`, and fade in / out to the master length. `make_sfx.sh` synthesizes two diegetic hits with
ffmpeg lavfi `anoisesrc` — a crisp ~120ms `snap.wav` (white noise, fast envelope) at the fiber-tear
cue (~1.5s) and a ~180ms `tear.wav` (pink noise) at the box-open cue (~11s). Time each to its beat.

## 4. End card + assembly → `make_end_card.py` + `build_master_v2.sh`  (config: `end_card`, `stat_callouts`, `grain_pass`, `audio_mix`)

- `make_end_card.py`: composes the STATIC end card entirely in PIL (NO AI) — the ivory (or
  brand-color) bg + the real logo PNG (upper third, ~62% width) + the real product PNG (centered,
  soft shadow) + a serif heritage headline + a smaller CTA line. **Never AI-render the lockup.**
  On macOS pick a serif with the middle-dot glyph (use ` · `).
- Assembly (`build_master_v2.sh`): normalize fps / SAR, concat the ~4 body clips, apply the anti-AI
  grain pass GLOBALLY (`eq=contrast=1.06:saturation=0.93,hqdn3d=1.5:1.5:3:3,noise=alls=8:allf=t+u`),
  composite the audio (music bed + `snap.wav` + `tear.wav` at their cue points, loudnorm), append
  the static end card holding ~3s WITH the music under it (fade the tail — no silent tail), and burn
  the OPTIONAL serif stat-callout pills at their beats → `edits/master-final.mp4` (1080×1920, 24fps,
  h264 + aac, ~14s). **Write `%` strings to a textfile and use `textfile=` + `expansion=none`** —
  ffmpeg `drawtext` reads a raw `%` as a strftime spec.

Re-cuts (new callout timing, swapped end card, re-timed SFX cues, grain-pass toggle) reuse the
existing keyframes / clips / music and cost **$0** — only steps 1–3 spend.

# render-food-product-sizzle scripts — the FREE assembly

`render-food-product-sizzle` is the **deterministic, $0 assembly stage** of the food-product sizzle
format. The paid stages (the ~4 photographic macro keyframes, the ~4 locked-off i2v clips, the
non-diegetic music bed) are separate capabilities — `create-image-fal`, `create-video-fal`,
`create-music-elevenlabs`. This capability spends nothing: it takes the ~4 clips + the music bed +
the diegetic SFX + the real logo PNG + the real product PNG and stitches the finished master.
Re-cuts (new callout timing, a swapped end card, re-timed SFX cues, a grain-pass toggle) reuse the
existing keyframes / clips / music and cost **$0**.

`config.example.json` is the worked example (Lineage Provisions "Beef Sticks Sizzle", ~14s
1080×1920). `PIPELINE.md` maps every config block to its source step. This README documents the FREE
assembly pieces that `render-food-product-sizzle` owns.

## 1. Static end card — PIL from the real logo + product PNG, no AI text

The end card is composed **entirely in PIL** (NO AI): the ivory (or brand-color) bg + the real logo
PNG (upper third, ~62% width) + the real product PNG (centered, with a soft shadow) + a serif
heritage headline + a smaller CTA line. The brand text is **never** AI-rendered — a diffusion model
garbles a wordmark and the packaging. On macOS pick a serif with the em-dash / middle-dot glyph
(use ` · `). It holds ~3s after the body with the music still playing under it.

## 2. Diegetic SFX — ffmpeg lavfi, on the tactile beats

Two short diegetic hits are synthesized with ffmpeg lavfi `anoisesrc` (NO AI): a crisp ~120ms snap
(white noise, fast envelope, highpass / lowpass, `acompressor`) on the fiber tear, and a ~180ms tear
(pink noise, low-end roll) on the box-open. They mix in at their **measured cue points** (~1.5s and
~11s in the demo), not round numbers — a couple of hits, not a wall of sound.

## 3. Music bed — no vocals, no sparse intro

The supplied / generated acoustic-bluegrass bed (no vocals — no VO to duck under) opens sparse; the
upstream step trims the ~2.5s intro so it kicks in from frame 0. The assembly `loudnorm`s it and
fades it in / out to the master length, playing it UNDER the end card with a fade tail (no silent
tail).

## 4. FFmpeg composite

FFmpeg stitches the master: normalize fps / SAR across the ~4 body clips, concat them in scene order,
apply the anti-AI grain pass GLOBALLY (`eq=contrast=1.06:saturation=0.93,hqdn3d=1.5:1.5:3:3,noise=alls=8:allf=t+u`
— load-bearing for the tactile / photographic read), composite the audio (music bed + `snap.wav` +
`tear.wav` at their cues), append the PIL end card, burn the OPTIONAL serif stat-callout pills at
their beats, mux with a fade tail, and `loudnorm`. Output is a 1080×1920 h264 + aac master (~14s).
Deterministic, no paid calls, no keys.

**`%` in a callout** — ffmpeg `drawtext` reads a raw `%` as a strftime spec; write the callout string
to a textfile and use `textfile=` + `expansion=none` (see the source `build_master_v2.sh` `c1.txt` /
`c5.txt`).

**No libass needed** — the sizzle body carries no burned captions (it is wordless); the stat-callout
pills are `drawtext` / `drawbox` overlays, not subtitle events.

# Smoke Test

Given the ~4 macro-scene i2v clips (locked-off, anti-shake — hands tearing, a flat lay, a bite, a
box / pack hero, the hero showing the REAL label), a non-diegetic acoustic music bed, a couple of
diegetic SFX (snap, tear), and the real logo + product PNG, `render-food-product-sizzle` assembles
the master: normalize fps / SAR, concat the clips in scene order, apply the anti-AI grain pass,
composite the audio (bed + SFX at their cue points), append a static PIL end card, burn the optional
serif stat-callout pills, and mux → 1080×1920 h264+aac (~14s).

Pass when the assembly runs to a valid MP4 and:
- ~4 macro scenes concat in order; the box / pack hero shows the REAL label (no invented packaging
  text — grounded on the product PNG upstream, not re-rendered here);
- the anti-AI grain pass is applied globally (the sizzle does not read AI-smooth);
- the diegetic SFX land on their tactile beats (snap on the tear, tear on the box-open);
- the music bed carries under the whole video with no sparse dead intro and a fade tail (no silent
  tail); it is the entire audio bed — NO separate VO;
- the optional serif stat-callout pills burn at their beats (`%` strings via `textfile=` +
  `expansion=none`);
- the end card is a static PIL composite (real logo PNG + real product PNG + a serif headline + CTA)
  — never AI-rendered text;
- **no paid call is made** — the keyframes, clips, and music come from the paid capabilities
  (create-image-fal / create-video-fal / create-music-elevenlabs); this assembly is $0 and a re-cut
  reuses the existing assets.