render-editorial-motion-podcast

# render-editorial-motion-podcast

Assemble an **editorial-motion podcast-clip** ad from a config: a real clipped podcast audio
line carries the whole narrative and every visual beat is timed to the sentence it describes,
in a bold flat 2-tone editorial-illustration look ("a New Yorker spot-illustration that
moves"). The motion is **not generative video** but deterministic ffmpeg ken-burns on static
keyframes, so it reads as a printed page that moves. This capability is that **FREE,
deterministic assembly** — the ffmpeg motion, hard-concat, audio mux, caption burn, and PIL
end card.

`scripts/config.example.json` is the worked example (Klarify "Rat Park", ~40.8s 1080×1920
9:16, 6 beats); `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 on the motion layer.
The paid inputs are separate: the real podcast MP3 is clipped from source (free ffmpeg) with
its Whisper word timings, and one editorial-illustration keyframe per beat (chained ref images
so cage/character geometry holds) comes from `create-image-fal` (Nano Banana). Given the
clipped audio + `words.json` + the per-beat keyframes + the real brand wordmark PNG,
`render-editorial-motion-podcast` renders each keyframe as a ken-burns segment, hard-concats
on the beat, muxes the real audio, burns the mid-sentence captions, and composites the PIL end
card → the master. Re-cuts reuse the existing audio / keyframes and cost **$0**.

## Contract (the free assembly)

- **A spoken narration carries the whole spot — no generated SONG.** Mux the provided
  narration MP3 (`-map 0:v:0 -map 1:a:0`) — a real clipped podcast line (preferred) OR an
  approved generated VO (`create-vo-elevenlabs`). Never a sung/generated track. (Clip-vs-generate
  is the recipe's STEP-0 intake decision — if no source episode is supplied, ASK the user.)
- **NO generative i2v — deterministic ffmpeg ken-burns only.** Animate each static keyframe
  with `zoompan` (push-in / pull-back, 1.0→~1.06×, 24fps); Seedance/Kling are photoreal-trained
  and invent naturalistic middle states that collapse the 2-tone look. Never `-loop 1` with
  `zoompan d=N` (it balloons the duration); feed a single image and clamp with `-t` + `trim`.
- **Hard cuts on the beat — no crossfades.** Crossfades ghost two drifting cages through each
  other; hard-concat each beat's segments and split long beats into micro-cuts (target 8–10
  distinct visual moments). Each beat's visual STARTS within ~0.5s of its spoken line.
- **Captions from Whisper word-timestamps, ON only mid-sentence.** Burn `frosted-subtle`
  captions while the speaker talks; leave silent/reflective beats and the end card uncaptioned.
  THREE mandatory rules (each bit us in prod — bake them in):
  1. **NON-OVERLAP** — clamp every line to END before the next STARTS
     (`end = min(last_word_end + ~0.15, next_start - 0.03)`). Two boxes must never stack at the
     same spot; an end-tail bleeding into the next window is the #1 caption bug.
  2. **SAFE AREA** — captions sit in the lower third, so the keyframe's subject must stay in the
     upper ~75% (see the recipe's `look_pack.caption_safe_area`). If a finished keyframe's subject
     intrudes into the caption band, deterministically **shift the subject UP** into the empty top
     space (PIL: paste up ~0.24H onto a canvas pre-filled with the exact paper color from a clean
     corner) — never let the box sit on the subject.
  3. **BURN ENGINE** — prefer libass (`ass`/`subtitles` filter), but **check `ffmpeg -filters`
     first**: many builds (Homebrew) lack libass/drawtext. If absent, use the deterministic
     **overlay fallback** — render each line as a transparent PNG (frosted rounded box + white
     text, PIL) and composite via the ffmpeg `overlay` filter with timed
     `enable='between(t,st,en)'` windows. Same look, no libass.
- **End card via PIL from the real wordmark PNG — never AI-render brand text.** The lockup is
  composited deterministically (stretched-gradient bg + feathered mascot crop + wordmark +
  tagline with a system font); a diffusion model garbles a wordmark ("therapits"). The video
  runs a ~1.5s silent hold past the audio on the end card (fade first/last 0.3s).
- **FFmpeg composite, deterministic, FREE.** Ken-burns each keyframe, hard-concat, mux the real
  audio, burn the captions, hold on the end card → a 1080×1920 h264+aac master. No paid calls.