{
  "_comment": "Perplexity concept-10 'Bloomberg terminal' split-screen creator ad — THE WORKED EXAMPLE. Copy to config.json and edit. A ~40s 9:16 two-zone ad where an AI creator anchor (VEED Fabric lip-sync) fills the bottom ~48% and real Perplexity Computer UI screen-recordings run uncropped in the top ~52%, a 3px brand-color divider between. The creator delivers the WHOLE VO cold-to-camera and each top clip PROVES its VO line's claim. Pipeline: (1) lock the script into 6 scenes + ElevenLabs eleven_v3 with-timestamps VO + per-scene timing.json; (2) gpt-image creator anchor + ONE VEED Fabric 1.0 @720p lip-sync take of the whole VO (~$5.90); (3) source 16:9 top clips + the scene-1 hook graphic; (4) assemble.py two-zone composite (contain-fit + blurred-cover fill + 3px divider + creator slice), hard-concat + end card; (5) transcribe the ASSEMBLED cut + burn serif-accent word captions. Master 1080x1920 h264+aac 30fps, ~40.37s. See scripts/PIPELINE.md + clients/perplexity/ads/concept-10-bloomberg-terminal/HOW_TO.md.",
  "brand_name": "Perplexity",
  "campaign": "Perplexity Computer",
  "title": "Perplexity Computer — the analyst that just does it",
  "width": 1080,
  "height": 1920,
  "fps": 30,
  "aspect_ratio": "9:16",
  "duration_sec": 40.37,
  "layout": {
    "_comment": "Two-zone split. TOP ~52% (top_height 998) runs the 16:9 UI screen-recording contain-fit + a darkened blurred cover-scale fill of the SAME clip (never black/charcoal bars — trap L3). Brand-color divider (keep stacked heights EVEN: 998+4+918=1920, or libx264 rejects the odd frame — trap L6). BOTTOM ~48% is the creator lip-sync framed head-to-shoulders: scale-to-width x a small ZOOM (~1.15-1.2) then crop with a DOWNWARD offset so the face is upper-middle and the shoulders enter (a plain crop-toward-top shows only the head and cuts the shoulders — trap L7). Requires a MEDIUM chest-up anchor (see creator._comment).",
    "canvas_w": 1080,
    "canvas_h": 1920,
    "top_height": 998,
    "top_pct": 0.52,
    "bottom_pct": 0.48,
    "divider_px": 3,
    "divider_color": "#20808D",
    "top_fill": "blurred-cover",
    "bottom_fit": "cover-crop-top"
  },
  "creator": {
    "_comment": "ONE AI-creator anchor lip-synced to the WHOLE VO in ONE VEED Fabric 1.0 take @ 720p (480p reads soft). THE ANCHOR MUST BE a PHOTOREAL, candid MEDIUM chest-up shot at natural webcam distance (whole head with headroom + shoulders + upper chest, real lived-in room, anti-AI skin cues) — NOT a plain-background close-up headshot (reads as AI) and NOT a phone-selfie / arm-extended pose copied from another format (wrong for a fixed talking-head, and too-close can't be reframed). Watch the source video and match its creator distance. gpt-image-2, quality high, 1024x1536. VEED handles photoreal fine. VEED Fabric is a NO-ATOM paid step: image_url = the anchor, audio_url = voiceover/vo-arlo.mp3; ~$0.15/sec (~$5.90 for the 39s VO). Run its calls SEQUENTIALLY with sleep 3 — veed/fabric-1.0 storage-auth 403s under parallel load (trap L1). The example explored a 2nd voice (Brad) which DOUBLED the lip-sync spend (~+$6) — lock the voice first. Re-run VEED only when the anchor IMAGE changes; a crop/zoom tweak is a free re-assemble.",
    "anchor": "avatar/creator-anchor.png",
    "anchor_model": "gpt-image-2",
    "anchor_shot": "medium chest-up, natural webcam distance, headroom + shoulders — not a tight close-up, not a phone selfie",
    "lipsync_engine": "veed/fabric-1.0",
    "lipsync_resolution": "720p",
    "lipsync_provider": "fal",
    "lipsync_output": "avatar/concept-10-bloomberg-terminal-creator-arlo.mp4",
    "voice": "arlo",
    "voice_locked": true
  },
  "voiceover": {
    "_comment": "The creator delivers the ENTIRE VO cold-to-camera — the VO IS the audio bed (no music). ElevenLabs eleven_v3 with-timestamps -> voiceover/vo-arlo.mp3 + timestamp JSON. Per-scene start/end sliced from the char timestamps into voiceover/timing.json. speed 1.2 (creator-native VO reads ~1.15-1.2x).",
    "provider": "elevenlabs",
    "model": "eleven_v3",
    "endpoint": "with-timestamps",
    "speed": 1.2,
    "script_txt": "voiceover/script.txt",
    "outputs": {
      "audio": "voiceover/vo-arlo.mp3",
      "word_timings": "voiceover/vo-arlo.timestamps.json",
      "timing": "voiceover/timing.json"
    }
  },
  "scenes": [
    {
      "id": "S01",
      "role": "HOOK",
      "start": 0.0,
      "end": 5.602,
      "uses_top_clip": false,
      "vo": "Okay, this used to need a twenty-five-thousand-dollar-a-year terminal and an actual analyst to run it. Watch this.",
      "top_visual": "Scene-1 hook graphic — a kinetic '$25,000/year terminal + an analyst' stat title-card (HTML-rendered, NOT a demo clip).",
      "top_clip": "assets/scene-01-graphic.png",
      "top_start": 0.0,
      "top_end": 5.602,
      "notes": "Scene 1 is the designed hook graphic, not a product clip."
    },
    {
      "id": "S02",
      "role": "DEMO",
      "start": 6.105,
      "end": 14.725,
      "uses_top_clip": true,
      "vo": "I gave Perplexity Computer one prompt — basically, is NVIDIA undervalued, pull the comps, build the model, make it visual. And it just went.",
      "top_visual": "Perplexity Computer taking the one prompt and kicking off — the UI accepts the request and starts working.",
      "top_clip": "assets/scene-02-clip.mp4",
      "top_start": 0.0,
      "top_end": 8.6,
      "notes": "The clip shows the exact prompt going in and the tool starting."
    },
    {
      "id": "S03",
      "role": "DEMO",
      "start": 15.093,
      "end": 20.82,
      "uses_top_clip": true,
      "vo": "It spun up this whole team of little agents, pulling live market data across, like, nineteen AI models at the same time.",
      "top_visual": "The agents spinning up, live market data pulling across ~19 models simultaneously.",
      "top_clip": "assets/scene-03-clip.mp4",
      "top_start": 0.0,
      "top_end": 5.7,
      "notes": "Set top_start/top_end to the on-message window — never loop a short clip (trap L5)."
    },
    {
      "id": "S04",
      "role": "DEMO",
      "start": 21.092,
      "end": 27.244,
      "uses_top_clip": true,
      "vo": "And in a few minutes, it literally built a full valuation model AND deployed an interactive dashboard I can actually click around in.",
      "top_visual": "The full valuation model building + the interactive dashboard being clicked around in.",
      "top_clip": "assets/scene-04-clip.mp4",
      "top_start": 0.0,
      "top_end": 6.1,
      "notes": "The clip proves the model + dashboard claim as it is said."
    },
    {
      "id": "S05",
      "role": "DEMO",
      "start": 27.825,
      "end": 31.23,
      "uses_top_clip": true,
      "vo": "Oh — and it set itself to re-run automatically every time earnings drop.",
      "top_visual": "The automation / auto-re-run-on-earnings setting being enabled in the UI.",
      "top_clip": "assets/scene-05-clip.mp4",
      "top_start": 0.0,
      "top_end": 3.4,
      "notes": "One feature, shown."
    },
    {
      "id": "S06",
      "role": "PAYOFF_ENDCARD",
      "start": 31.55,
      "end": 39.358,
      "uses_top_clip": true,
      "vo": "Like, look at that. It didn't explain how I could do the analysis. It didn't give me steps. It just did the whole thing. That's Perplexity Computer.",
      "top_visual": "The finished dashboard into the Perplexity Computer end card (held on the last SHARP frame ~3s).",
      "top_clip": "assets/scene-06-clip.mp4",
      "top_start": 0.0,
      "top_end": 7.8,
      "endcard_clip_end": 7.8,
      "notes": "Note the last SHARP second before any fade-to-black as endcard.clip_end (trap L4)."
    }
  ],
  "endcard": {
    "_comment": "The Perplexity Computer end card held on the last SHARP frame ~3s. If the end-card clip fades to black, hold the last sharp second (clip_end), NOT the black tail (trap L4).",
    "clip": "assets/scene-06-clip.mp4",
    "clip_end": 7.8,
    "hold_sec": 3.0
  },
  "captions": {
    "_comment": "Word-level captions burned from the ASSEMBLED cut (transcribe master-precaption.mp4 with local faster-whisper, NOT vo-arlo.mp3 — concat drops inter-scene silence so the ad timeline != the VO timeline; trap L2). Sentence-aware chunking. concept-10 shipped serif-accent (after a kinetic-pop pass). Kept: master-precaption cut + captions/serif-accent.ass sidecar for restyle without re-render.",
    "source": "assembled-cut",
    "engine": "faster-whisper",
    "style": "serif-accent",
    "chunk": "sentence-aware word-level",
    "ass_sidecar": "captions/serif-accent.ass",
    "keep_precaption": true,
    "keep_ass_sidecar": true
  },
  "audio_mix": {
    "_comment": "The creator VO is the entire audio bed — no separate music to duck under. Body audio = the concatenated creator VO slices; the end card holds under the VO tail.",
    "duck_under_music": false,
    "loudness_lufs": -14,
    "bed": "creator_vo"
  },
  "post_production": {
    "vo": { "default": "on", "note": "the creator's ElevenLabs VO IS the audio bed; the lip-sync drives the mouth (no music)" },
    "captions": { "default": "on", "note": "word-level captions burned from the ASSEMBLED cut (serif-accent), tracking the ad timeline not the raw VO" },
    "divider": { "default": "on", "note": "a 3px brand-color divider between the two zones" },
    "endcard": { "default": "on", "note": "the Perplexity Computer end card held on the last sharp frame ~3s" }
  }
}

# Pipeline — split-screen creator

How `config.example.json` maps to the real production steps. This capability
ships a **config + this map** — the FREE assembly + captions stage. The worked
example (Perplexity concept-10 "Bloomberg terminal") was produced by the per-scene
paid steps plus the split-screen assembler + the caption burn.

The steps run **in order** because each depends on the last: the script sets the
scene count, the VO + its per-scene timing drive the bottom-zone slices, the
locked voice + anchor drive the ONE VEED Fabric lip-sync take, the brand clips
fill the top zone, assembly stitches the two zones, and captions come from the
ASSEMBLED cut (not the raw VO).

## Field → source-step map

| Config field | Phase | Source step | Paid? |
|---|---|---|---|
| `voiceover.*`, `scenes[].vo` | 1 VO | ElevenLabs `eleven_v3` with-timestamps (`create-vo-elevenlabs`) → vo mp3 + timestamp JSON | **PAID** (~$0.06) |
| `scenes[].start/end`, `voiceover.outputs.timing` | 1 VO | slice per-scene start/end from the char timestamps → `timing.json` | free |
| `creator.anchor` | 2 Creator | photoreal MEDIUM chest-up anchor, `gpt-image-2` (`create-image-gpt-image-fal`) | **PAID** (~$0.10) |
| `creator.lipsync_*` | 2 Creator | **VEED Fabric 1.0 @ 720p — a NO-ATOM paid step** (`image_url` = anchor, `audio_url` = vo mp3) → creator lip-sync mp4 | **PAID** (~$5.90 / 39s) |
| `scenes[].top_clip` + `top_start/top_end` | 3 Top | brand-owned 16:9 UI screen-recordings (S02–SN) + the scene-1 hook graphic | (brand-supplied) |
| `layout.*`, `endcard.*`, assembly | 4 Assembly | the split-screen assembler — two-zone composite + hard-concat + end card → `master-precaption.mp4` | free |
| `captions.*` | 5 Captions | transcribe the ASSEMBLED cut (local `faster-whisper`) → word-level ASS burn → `concept-<n>-final-ad.mp4` | free |

## 1. Voiceover (config `voiceover`, `scenes[].vo`) — PAID

**Lock the script first** — ~6 scenes, one VO sentence per scene, each naming one
claim. Call ElevenLabs `eleven_v3` on the **with-timestamps** endpoint → the vo
mp3 + character-level timestamps. Slice per-scene start/end boundaries into
`timing.json` (`{"scenes":[{"scene":N,"start":s,"end":e}]}`). Creator-native VO
reads ~1.15–1.2×; apply `atempo` if it runs fast. **Lock the voice here** — a
second voice doubles the VEED spend downstream. This is upstream of the capability.

## 2. Creator anchor + lip-sync (config `creator`) — PAID, no-atom VEED step

A **photoreal MEDIUM chest-up** `gpt-image-2` anchor at a natural webcam distance —
whole head with headroom, shoulders + upper chest, real lived-in room, anti-AI skin
cues — **not** a plain-background close-up headshot and **not** a phone-selfie pose
(borrow another format's realism, never its arm-extended framing). Watch the source
video first and match its creator distance; get the anchor framing right BEFORE the
lip-sync, since a too-close image can't be reframed in the composite. Then the
**whole VO** lip-synced in **one VEED Fabric 1.0 take @ 720p** (`veed/fabric-1.0`,
`image_url` = the portrait, `audio_url` = the vo mp3) → the creator lip-sync mp4.
**VEED Fabric is a no-atom paid step** — there is no `create-lipsync-*` atom; it's
an external FAL call (~$0.15/sec, ~$5.90 for a 39s VO). Always 720p — 480p reads
soft in the bottom zone. Run the lip-sync call(s) **sequentially with a `sleep`** —
`veed/fabric-1.0` storage-auths 403 under parallel load. Upstream of the capability;
the lip-sync take is a supplied input.

## 3. Top-zone clips + hook graphic (config `scenes[].top_clip`)

One brand-owned 16:9 UI screen-recording per demo scene (≥1920×1080), each showing
exactly what its VO line says (the #1 quality lever). Render the scene-1 hook
graphic (kinetic title-card / stat claim) separately. Confirm the end-card clip;
note the last **sharp** second before any fade-to-black as `endcard.clip_end`.

## 4. Ad-spec + assembly (config `layout`, `scenes[].top_*`, `endcard`) — FREE

Author the ad-spec (canvas 1080×1920, `top_height` ~998, `divider_color`, per-scene
`top_clip` + `top_start`/`top_end`, `endcard.clip` + `clip_end`) and run the
split-screen assembler. Per scene: TOP zone = the clip **contain-fit** + a darkened
**blurred cover-scale of the same clip** filling the letterbox margins (never
charcoal/black bars); a 3px brand-color divider (keep stacked heights EVEN, e.g.
998+4+918=1920, or libx264 rejects the odd frame); BOTTOM zone = the creator lip-sync
framed head-to-shoulders (scale-to-width × ZOOM ~1.15–1.2, then crop with a downward
offset so the face is upper-middle and the shoulders enter — NOT a plain crop-toward-top,
which cuts the shoulders). Tune zoom/offset visually (free re-assemble). **Never loop a
short clip** — window it and the assembler speed-fits it. Hard-concat the scenes; body audio =
the concatenated creator VO slices; append the end card holding the last sharp frame
~3s → `master-precaption.mp4`. Deterministic, $0.

## 5. Captions (config `captions`) — FREE

**Caption the ASSEMBLED cut, not the raw VO** — concat drops inter-scene silence, so
the ad timeline ≠ the VO timeline; only the final cut's audio yields correct caption
timing. Transcribe `master-precaption.mp4` with local `faster-whisper`, build
word-level cues (sentence-aware chunking), and burn the ASS in the chosen style
(`serif-accent`, `kinetic-pop`, `neon-glow`, `clean-bubble`) → `concept-<n>-final-ad.mp4`.
Keep the `-precaption` cut + the `.ass` sidecar so captions restyle without
re-rendering the composite. If the host ffmpeg lacks libass, render the cues as timed
PIL PNG overlays composited with ffmpeg `overlay=…:enable='between(t,st,en)'` instead.

Re-cuts (re-timed windows, a swapped blurred-fill, new caption chunking, a restyle, a
longer end-card hold) reuse the existing VO / lip-sync / clips and cost **$0** — only
steps 1–2 spend.

# render-split-screen-creator scripts — the FREE assembly

`render-split-screen-creator` is the **deterministic, $0 assembly + captions
stage** of the split-screen creator format. The paid stages (the VO, the
gpt-image creator anchor, and the VEED Fabric 1.0 whole-VO lip-sync) are separate,
upstream steps — `create-vo-elevenlabs`, `create-image-gpt-image-fal`, and a
paid **VEED Fabric lip-sync (a no-atom step)**. This capability spends nothing —
it takes the creator lip-sync take + the per-scene VO timing + one 16:9 top clip
per scene + the scene-1 hook graphic + the end-card clip and stitches the finished
master. Re-cuts (re-timed windows, a swapped blurred-fill, new caption chunking, a
restyle, a longer end-card hold) reuse the existing VO / lip-sync / clips and cost
**$0**.

`config.example.json` is the worked example (Perplexity concept-10 "Bloomberg
terminal", ~40s 1080×1920). `PIPELINE.md` maps every config block to its source
step. This README documents the FREE assembly pieces that
`render-split-screen-creator` owns.

## 1. Two-zone composite — top ~52% product clip, bottom ~48% creator

Per scene the assembler builds a 1080×1920 frame in two zones. The TOP zone
(`top_height` ~998, ~52%) runs the real 16:9 product/demo clip **contain-fit**
(uncropped) — its letterbox margins filled with a darkened **blurred cover-scale of
the same clip** (never charcoal/black bars — a flat bar reads cheap). The BOTTOM
zone (~48%) is the creator lip-sync framed **head-to-shoulders**: scale-to-width × a
small **ZOOM** (~1.15–1.2) then crop with a **downward offset** so the face sits
upper-middle and the shoulders enter (a plain crop-toward-top shows only the head and
cuts the shoulders — and needs a **medium chest-up** anchor to begin with, not a tight
close-up). A 3px brand-color divider separates the zones — keep stacked heights **even**
(998 + 4 + 918 = 1920), or libx264 rejects the odd frame.

## 2. Windowing — one claim per scene, never loop a short clip

Each top clip is windowed (`top_start`/`top_end`) to the on-message segment that
proves its VO line. If a clip is shorter than its scene, **do NOT loop it** (that
replays into a sparse/black tail) — set the window and the assembler **speed-fits**
it to the scene length.

## 3. Hard-concat + end card — last sharp frame, no black tail

The scenes are **hard-concatenated** (no dissolves). The body audio is the
concatenated creator VO slices timed per `timing.json` — the sung/spoken creator VO
IS the bed, no separate music. The end card is appended holding the last **sharp**
frame ~3s; if the end-card clip fades to black, hold the last sharp second
(`endcard.clip_end`), not the black tail.

## 4. Captions — from the ASSEMBLED cut, never the raw VO

Captions come from transcribing the **assembled** cut (`master-precaption.mp4`) with
local Whisper — **never the raw `vo.mp3`**. Concat drops inter-scene silence, so the
ad timeline ≠ the VO timeline; only the final cut's audio yields correct caption
timing. Build word-level cues (sentence-aware chunking) and burn the ASS in the
chosen style (`serif-accent`, `kinetic-pop`, `neon-glow`, `clean-bubble`). Keep the
`-precaption` cut + the `.ass` sidecar so captions restyle without re-rendering the
composite. If the host ffmpeg lacks libass, render the cues as timed PIL PNG overlays
composited with ffmpeg `overlay=…:enable='between(t,st,en)'` instead.

## 5. FFmpeg composite

FFmpeg stitches the master — the two-zone composite per scene (contain-fit +
blurred-cover fill + 3px divider + creator slice), hard-concat, append the end card
holding ~3s, mux the creator VO, and `loudnorm I=-14`. Output is a 1080×1920 h264 +
aac master. Deterministic, no paid calls, no keys — the VEED Fabric lip-sync is a
supplied input, produced upstream.

# Smoke Test

Given the creator lip-sync take (a supplied VEED Fabric mp4) + the per-scene VO
timing, one 16:9 top clip per scene with the scene-1 hook graphic, and the
end-card clip, `render-split-screen-creator` assembles the master — the two-zone
composite (top ~52% product clip contain-fit + blurred-cover fill, bottom ~48%
creator slice, 3px divider), hard-concat on the scene, an end card held on the
last sharp frame ~3s, the creator VO muxed, then word-level captions burned from
the ASSEMBLED cut → 1080×1920 h264+aac (~40s).

Pass when the assembly runs to a valid MP4 and:
- one two-zone scene per VO scene, hard-cut on the scene (no crossfades); the top
  clip is contain-fit with a darkened blurred-cover fill (no black bars); the 3px
  brand-color divider is present;
- each top clip is windowed (never looped into a sparse/black tail); the bottom
  zone is the creator lip-sync slice per the per-scene timing;
- captions are word-level, transcribed from the ASSEMBLED cut (not the raw VO —
  concat drops inter-scene silence), burned in the chosen style;
- the end card is on the last sharp frame (no fade-to-black tail); the creator VO
  is the entire audio bed — no separate music;
- **no paid call is made** — the VO, the anchor, and the VEED Fabric lip-sync take
  come from the paid upstream steps (create-vo-elevenlabs / create-image-gpt-image-fal
  / the no-atom VEED Fabric lip-sync); this assembly is $0 and a re-cut reuses the
  existing assets.