render-editorial-motion-podcast is the deterministic, $0 assembly stage of the\neditorial-motion podcast-clip format. The paid stages — the two Whisper calls (~$0.04 each) and\nthe per-beat Nano Banana keyframes — are upstream (create-image-fal + a transcription step);\nthe real podcast MP3 is clipped from source with free ffmpeg. This capability spends nothing on\nthe motion layer: it takes the clipped audio + words.json + the per-beat keyframes + the real\nbrand wordmark PNG and stitches the finished master. Re-cuts (new beat windows, added micro-cuts,\nre-timed caption ranges, a swapped end card) reuse the existing audio / keyframes and cost $0.
config.example.json is the worked example (Klarify \"Rat Park\", ~40.8s 1080×1920). PIPELINE.md\nmaps every config block to its source step. This README documents the FREE assembly pieces that\nrender-editorial-motion-podcast owns.
The motion is deterministic ffmpeg ken-burns on static keyframes, never generative i2v.\nSeedance/Kling are photoreal-trained and invent naturalistic middle states (a photoreal drawing\nhand, rats morphing into humanoids, hallucinated clouds) that collapse the 2-tone discipline.\nTwo helpers: render_kb <src> <out> <dur> [zoom_end] (push-in — scale to 2×, crop,\nzoompan z='min(zoom+step,zoom_end)', trim=duration, -t dur) and render_kb_out (pull-back,\nzoom_start down to 1.0). Never -loop 1 with zoompan d=N — d runs per input frame per\nloop and balloons the duration; feed a single image and clamp with -t + trim.
Each beat's segments are hard-concatenated on the beat; long beats split into micro-cuts (target\n8–10 distinct visual moments — anything held > 6s reads static even with active zoompan). No\ncrossfades — they ghost two drifting cages through each other. Each beat's visual STARTS within\n~0.5s of its spoken line (marked from the Whisper timings upstream), never ahead, never lag.
\nThe narration MP3 is muxed as the master audio (-map 0:v:0 -map 1:a:0) — no generated song.\nThe narration is a real clipped podcast line (preferred) OR an approved generated VO\n(create-vo-elevenlabs); clip-vs-generate is the recipe's STEP-0 intake decision (if no source\nepisode is supplied, ASK the user — don't fabricate or silently skip). The video runs a ~1.5s\nsilent hold past the audio on the end card (fade the first/last 0.3s so it doesn't hard-black-flash).
Captions are frosted-subtle (Whisper word-timestamps on the master) ON only while the speaker\nis mid-sentence — silent/reflective beats and the end card stay uncaptioned. Three mandatory\nrules, all learned the hard way:
end = min(last_word_end + ~0.15, next_start - 0.03)). Two boxes must never be on screen at\nonce; an end-tail bleeding into the next window stacks two boxes at the same spot (the #1\ncaption bug). Verify zero remaining window overlaps before burning.look_pack.caption_safe_area). If a finished keyframe's subject intrudes\ninto the caption band, deterministically shift the subject up into the empty top space\n(PIL: paste up ~0.24H onto a canvas pre-filled with the exact paper color sampled from a clean\ncorner, so the vacated bottom is clean paper) — never let the box sit on the subject.ass/subtitles), but check ffmpeg -filters first:\nmany builds (Homebrew) ship without libass/drawtext. If absent, use the deterministic overlay\nfallback — render each caption line as a transparent PNG (frosted rounded box + white text,\nPIL) and composite with the ffmpeg overlay filter using timed enable='between(t,st,en)'\nwindows. Identical frosted-subtle look, no libass required.The brand lockup is composited with PIL from the brand's REAL wordmark PNG: stretch a 1px\ncenter column → 1080×1920 + GaussianBlur(80) for the gradient bg; a feathered radial-alpha\nellipse crop of the brand mascot pasted centered-upper; wordmark + tagline drawn with a system\nfont. The brand text is never AI-rendered — a diffusion model garbles a wordmark\n(\"therapits\"). Output is a 1080×1920 h264 + aac master. Deterministic, no paid calls, no keys.