ohmdob/podcast-motion-animatic

Podcast Motion: Animatic

Status: one scripted film rendered and approved by its producer: 2 minutes, two drawn hosts, cards, blueprint, 16:9. On real footage and at 9:16 it has only been read on test frames, so show those and get a decision before any long render, then record the result in the show's STYLES.md. Read podcast-playbook first: it holds the sound, the cut, the gates and the master.

• Idea: a storyboard brought to life. Each point in the talk is a panel, held and moved in stepped motion, under the real voice.
• Best for: talks about things you can draw, and audio-only shows.
• Extra step: none, and no matte.
• Costs: the pictures, unless the show already has them. It reuses any stills it finds.

Faces stay natural

Panels draw ideas and objects, never the people talking. An illustrated guest is the one thing this look must not do: the guest's credibility is the point of the show, and a stylised face has already been rejected on sight on a real show. So never give it a picture of a person, and don't ask a generator to draw one. Between panels you see the real speaker, in natural colour.

Hosts who are already characters are the one exception, when the producer asks: the reason above is about a real person's credibility. They take the speaker's place, and panels still show only objects and words. See references/scripted-show.md in podcast-playbook.

Build

python3 scripts/build.py episodes/EPxx/full --ratio both --frames
python3 scripts/build.py episodes/EPxx/clips/Cnn_name --ratio both --frames

scripts/ is this skill's folder. The film's folder needs full_edit.py --render done first, with faces.jsonl and pass1.json in build/ (references/full-episode.md in podcast-playbook). --frames writes a contact sheet of real frames to judge. It writes variant_animatic/ and variant_animatic_9x16/. Render and master as in podcast-playbook.

Panels

Every card in cards.json is a panel.

• A card with a picture (opts.panel, else opts.still or opts.obj): the picture as a pencil sketch, with the card's title under it and each row replacing the last as it lands.
• A card with none: its words, lettered in the frame.

It looks for assets_src/panels/<name>, then Studio Cut's stills, then Pop Studio's objects, as .png or .jpg. A named picture it can't find goes on as words, and the build reports it.

Sketching is local and free, and it's what makes this cheap: any photograph, or a picture from any tool, comes out in the same hand, so mixed sources still look like one storyboard. If your pictures are already drawn the way you want, set "sketch": false.

Getting pictures follows references/storyboard-and-assets.md: quote first, and with no Ideogram, ask the producer which alternative to use. One picture serves several cards, so a 20-minute episode can run on about 20 panels. Keep to one clear object, because a busy photograph sketches into noise:

Studio photograph of one single plain white cup, centred, on a plain pale background, soft even light, strong clear outline. No text, no letters, no logos, no people, no hands.

Save as assets_src/panels/cup.png, and write {"panel": "cup"} on the card.

What makes it read as drawn: the moves run at six steps a second (steps()), and cuts are hard. Smooth motion turns it into a slideshow, so leave that alone.

Cards, motion and theme

What one producer asked for, one request at a time, and all of it held up. Each is off until the show asks, and the code is in scripts/animatic_cards.py, whose top lists every option.

"looks": {"animatic": {"layout": "card", "blueprint": true, "paper": "#0A1A33", "ink": "#E6F4FF", "pencil": "#9CC4E4", "accent": "#FFB547"}}

• "layout": "card" keeps the speaker on screen. A full-frame panel hides them, and the producer asked for image cards and explaining text like Pop Studio's. The card takes the side away from whoever is talking: heading first, rows landing on their words and staying, then a picture filling what the words leave. 16:9 only so far. Type behind the hosts, Pop Studio's other half, needs a cutout, so it isn't here.
• opts.motion moves the picture with the story. Built in: still, drop, arc, fly, drift, sway, hop, swarm, flipbook. A slow push-in on every panel wasn't enough for the producer; a ball being thrown tells the story as well as showing it. Write your own in moves.py: examples/hail-mary/moves.py has one line told in three drawings, and a ball from early in the film landing at the end.
• "blueprint": true themes it. Take the theme from the story, and ask: a navy grid suited a space film. It inverts every sketch and lays it over with screen, so moved pictures show no edge.
• Draw what a generator can't give you cleanly. A flipbook of 18 drawings, a walking alien and a solid ship cost nothing, and read better than sketched photographs. examples/hail-mary/draw_art.py shows how. A drawing of someone else's character is still theirs, so raise it before publishing.
• "wide" in paper.json gives a two-shot. Hosts who trade short lines never get one from the planner, and a viewer should see both early and at the end.

Mistakes already paid for

• A card must land on its own framing. The planner holds a shot for at least a second, so a card could hide half a face for 0.9 s. In card layout the shots now cut exactly on card edges and on each change of speaker.
• A sliver of the other host at the frame edge looks like a mistake. The card gets a solid strip out to the edge behind it.
• Picture too small in its window: give it what the words leave, then use zoom. A busy row of text over a tiny apple looked unfinished.
• Flip a picture that flies backwards. The ship's nose pointed the way it had come.
• Check that what hangs meets what it hangs from. Strings stopping short of the puppet were on screen for 10 s before a fresh reviewer said so.
• A lettered card is the weakest one on a film with moving pictures. Keep them for a number, and draw something for the rest.

No camera

For a show with audio only, make a plain video from the recording, and the whole pipeline runs as usual:

ffmpeg -f lavfi -i "color=c=0xF6F7F5:s=1920x1080:r=25" -i EPxx.wav -shortest -c:v libx264 -crf 18 -pix_fmt yuv420p -c:a aac 1080p/EPxx.mp4

Then set "audio_only": true in paper.json. Between panels the frame holds the chapter's name. faces.swift finds no faces, which is fine. Speakers come only from the transcript, so correct who in pass1.json. Keep a panel every few seconds: with nothing else moving, a long gap is a frozen picture, and it fails the master's check at 12 s.

What's on screen

• 16:9: the speaker fills the frame between panels; a panel is a loose pencil frame on plain paper, with one line of caption below it.
• 9:16: the frame sits at y 170 to 1070, with the caption below it inside the app's clear area, and the show's mark below that. The chapter's name takes the caption's place between panels.
• Numbers and verdicts are in red. There is no flood: a red word in a pencil frame is already the loudest thing on screen.
• Paper is a cool off-white, on purpose: cream paper with serifs is the look every show should avoid.

"looks": {"animatic": {"paper": "#F6F7F5", "ink": "#1C2126", "pencil": "#59626A", "accent": "#D3261F", "sketch": true}}

Handwriting is Itim, which holds Thai and Latin.

What to check

• Read each caption and lettered panel at phone size.
• Does the sketch read as the thing it's meant to be? Look at each one. A picture that doesn't, on a health or legal point especially, is worse than words: letter it instead.
• A panel covers the speaker completely, so don't put one over the moment their face is doing the work.
• The same picture repeating is fine for a few cards, and gets tired past that.
• Fine on purpose: timeline_track_too_dense, because a chapter holds many timed panels.
• Step through each card's arrival frame by frame. The first second of a card is where framing and entrance faults hide, and samples every few seconds miss them.