---

name: render-absurdist-explainer description: Assemble an absurdist animated-explainer video ad (~38s, 9:16) from per-scene i2v clips + their measured VO windows — retime each clip to its VO, re-encode every segment to identical 30fps/libx264/yuv420p so the concat demuxer never drops frames, concat, build a REAL-product PIL end card (never AI) held static, mix VO (loudnorm I=-14) under music (loudnorm I=-26, volume 0.62, amix normalize=0), and burn libass captions last. FREE deterministic assembly (bash-free, Python + ffmpeg + PIL); the recipe supplies the clips, VO, music, product photo, palette, and caption table and gates the paid keyframe/clip/VO/music calls to their own capabilities. Use for the absurdist-explainer format. status: active

render-absurdist-explainer

The free, deterministic renderer for the absurdist-explainer video ad format — an animated spot that explains one problem and how the product fixes it. The format is the mechanics: show the problem in a funny, exaggerated cartoon way, teach the product's ownable mechanism visually, show what the problem costs, bring the product in, land the fix as the climax, pay it off, end on a real-product card. One narrator voice carries the whole spot. How the story is told (problem → fix, how it works, a day in the life, the product as hero, or a villain arc) is the user's choice, made upstream. This capability is the FREE assembly stage only. All generative work (nano-banana keyframes, Seedance i2v clips, ElevenLabs VO + music) happens upstream in the recipe and is handed to this capability as files.

It ports the compose recipe validated on two earlier reference runs (both happened to use the villain arc). The recipe is deterministic — iterate the cut for free, re-roll only the offending paid beat.

Choices

The creative calls are made upstream by the recipe's choices and arrive here only as files and config values. This renderer is story- and style-agnostic: it never assumes a story shape, a cast, a look, a narrator or a music style.

• story_shape — how the story is told: problem → fix (no villain) / how it works / a day in the life / the product as hero / villain arc (the problem as a cartoon villain who schemes and loses). Reaches this capability as the per-scene caption text and the VO files. Asked of the user; the demo used the villain arc.
• cast — who the cartoon characters are (the customer, the product as a character, a mascot; a personified problem only on the villain arc). Reaches this capability only inside the clips. Asked of the user.
• narrator — who speaks the single VO track. Arrives as the scenes[].vo files. Asked of the user; the demo used the villain (villain arc only).
• visual_style — the art style of the i2v clips (scenes[].clip). Asked of the user; the demo used Pixar-style 3D. Nothing here depends on it.
• narrator_voice — the voice cast for the VO. Asked of the user; the demo used a characterful male villain voice.
• music — the bed at music_bed. Asked of the user; the demo used whimsical pizzicato + woodwinds + xylophone. The mix constants below apply to any style.

What it does (the deterministic recipe)

1. Per-scene retime. Each i2v clip is retimed to its measured VO window (scale=1080:1920:force_original_aspect_ratio=increase,crop=1080:1920,fps=30,setsar=1, then tpad=stop_mode=clone if the VO is longer than the clip, else a trim). Each window is snapped to a whole number of frames first. compose.py and make_captions.py use the same rule, so video cuts, VO windows and caption cues share the same cut times and a caption never outlives its cut.
2. Identical re-encode. Every segment is re-encoded libx264 -crf 18 -pix_fmt yuv420p -r 30 even if already correct — a framerate mismatch makes the concat demuxer silently drop frames.
3. Concat all scene segments + the end card via the concat demuxer (-c copy).
4. Real-product end card. build_endcard.py composites the REAL retail product photo over the brand palette (flat, or sampled from the photo's own edge pixel) with a typeset wordmark + claim rows + CTA pill in PIL ImageDraw.text — never an AI cartoon bottle, never AI-rendered brand text. compose.py holds it static over the dwell. Set end_card.zoom_to above 1.0 (for example 1.04) only if you want a slow zoom. Text that would be hard to read on its background (a dark brand colour on a dark photo) is drawn in white or near-black instead, with a warning. If end_card.vo is set, that spoken line plays from the start of the end-card window, and the dwell is stretched to at least the line's duration + 0.5s.
5. Mix. VO bus loudnorm I=-14 TP=-1.5, music bus loudnorm I=-26 TP=-3 then volume=0.62, amix inputs=2 duration=first normalize=0, so the music is ducked under the VO.
6. Master pass. The mix is measured, gained to -14 LUFS, limited, encoded to AAC and measured again. The pass repeats until the encoded audio is at -14.5..-13.5 LUFS with a true peak <= -1.5 dBFS. compose.py prints the measured loudness and peak.
7. Captions last. make_captions.py emits a libass .ass (one cue per scene, Arial 64 white / 6px outline / MarginV=330, start = scene_start + 0.08s, suppressed on the end card). compose.py burns it as the final filter so captions sit on top.

Scripts (free — Python + ffmpeg + PIL, no bash, no paid calls)

• scripts/build_endcard.py — PIL composite of the real product photo + typeset brand layer (wordmark / product line / claim rows / accent CTA pill). Reads the same config.json. Run this FIRST so end_card.image exists before compose.py.
• scripts/make_captions.py — emits the per-scene libass .ass from the SAME scene table compose reads, so caption windows stay in lockstep with the cut. Run before compose.py (or point config.captions_ass at nothing to skip captions).
• scripts/compose.py — the assembler: per-scene retime + identical 30fps re-encode → concat -> static end card (voiced if end_card.vo is set) -> VO/music loudnorm mix -> master loudness pass -> burn captions -> master mp4.
• scripts/config.example.json — the shape of the config the recipe binds. Its values are a labelled worked example (the demo build: a villain-arc story where Rust loses to Glidefern, a made-up bike-chain wax). Captions, end-card copy and palette come from the user's brand and choices — never copy them as defaults.

Inputs (all via --config + a runtime work dir — NO hardcoded paths)

config.json carries: scenes[] (each {id, clip, target_sec, vo, caption, atempo?} where target_sec is the measured VO window), end_card{product_image, image, dwell_sec, zoom_to?, wordmark, product_line, claims[], cta, background?, vo?}, brand_palette {primary, primary_lite, accent, grey}, music_bed, music_volume (default 0.62), atempo (compose-stage VO speed-up, default off; the reference runs used 1.3 when the VO read slow), captions_ass, and caption_style. See config.example.json.

• end_card.vo (optional) is the path to the end card's spoken line. It is laid at the start of the end-card window. The dwell becomes the larger of dwell_sec and the line's duration + 0.5s. Leave it out for a silent end card (music only). end_card.atempo overrides the global atempo for this line.
• Paths may be absolute, including Windows paths (C:/...), captions_ass too.
• The config is read as UTF-8 (with or without a BOM). Save it as UTF-8 if any text has non-ASCII characters.
• Fonts. The end card uses DejaVu (Linux), Arial (macOS), or Arial / Segoe UI (Windows). If none is found it falls back to Pillow's built-in font at the right size.

Craft rules (load-bearing — faithful to the source molecule)

• The end card is the REAL product photo, composited — never an AI cartoon bottle. Both reference runs shipped an AI bottle first and had to re-shoot with the real photo.
• No AI-rendered brand text anywhere. Wordmark, claims, CTA, motif — all PIL ImageDraw.text. AI draws the world + characters only.
• Re-encode every segment to 30fps before concat, even if already correct, or the concat demuxer silently drops frames.
• target_sec is the MEASURED VO duration (ffprobe each VO mp3), never a planned word count — VO drives the per-scene timing.
• Mix constants are validated - VO -14 LUFS, music -26 LUFS then volume 0.62 to 0.70 (the two reference runs), amix normalize=0. Master target -14.5..-13.5 LUFS, true peak <= -1.5 dBFS. The master pass enforces this and prints what it measured.
• Caption start = scene_start + 0.08s, suppressed on the end card (its typeset copy carries the message — two text layers at one spot are both unreadable).

Known gaps

• Captions are one cue per scene. The whole sentence is on screen from the start of the scene, before most of it has been spoken. Captions are not timed to the words. This is not fixed yet.

Requires

watch (QC the final master - confirm every character's look holds, the single narrator voice carries the whole spot, no AI brand text leaked into a cartoon background, the end card is the real product, and duration is within 0.1s of the summed windows plus the end-card dwell). The recipe gates the paid create-image-fal (keyframes), create-video-fal (Seedance i2v), create-vo-elevenlabs, and create-music-elevenlabs calls to their own capabilities — this capability itself makes NO paid calls.