算法艺术创作
algorithmic-art
通过代码和可调参数创作生成艺术与交互式视觉效果。
render-whiteboard-explainer【设计与创意】Assemble a whiteboard explainer video ad from a beats file — a narrator explains one idea while a real whiteboard fills up in black marker, with a hand-lettered title, bulleted rows down the left, a drawing filling the right, and a payoff that letters the closing line and rings the claim. The board is a PHOTOGRAPH and the ink is multiplied onto it, s…
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
The renderer for the whiteboard video ad format: a voice explains one idea while a board fills up in marker, the way a graphic recorder works a room. Every mark lands on the word it belongs to, and the board is wiped and reused between sections.
Siblings in the chat-UI family are render-imessage-chat, render-chatgpt-chat and
render-apple-notes-chat, where a screen is the creative. Reach for this one when the creative
is someone explaining a thing by drawing it.
The board is a photograph. A drawn board reads as a drawn board, and a plain white frame reads as a sketch film. Both were built and both were rejected. One generated still of a real blank whiteboard in a real room is the plate, and the ink is multiplied onto it so the surface sheen and the window light come through the strokes rather than sitting on top as flat black. The handheld drift is generated over both layers at once, so nothing can slide against anything.
The drawings are generated as line art, then traced. Do not hand-code them, and do not trace painterly images. Art traced from a photograph carries every contour the photograph had and reads as clip-art; art generated as a marker drawing traces to a handful of confident paths. An open hand came back as a single continuous stroke. Chain each generation off the first so the line weight matches across the set.
There is no colour. Black marker throughout. Do not add an accent.
The recipe asks these before any paid step; this renderer only draws what the beats say.
"new_board": true on the beat that should
open a board; any such mark turns the automatic split off. Use it when the split puts a
line on the wrong board, such as the brand name landing on the problem board.Brand facts come from the brand kit.
Every mark lands on its own word. The voiceover is the clock. An anchor that resolves to the wrong word does not look wrong, it silently reorders the video: one that matched an early word instead of a late one once dragged a whole section to the front, and the ad played its ending first with nothing reporting a problem. Anchors are resolved once, shared by the solver and the renderer, and an ambiguous one is refused rather than guessed at.
Nothing is hand-placed. Anything that refers to something else — an arrow, a note, a ring — takes its target as an argument and works out its own position. Anything sharing space with a drawing is bounded by that drawing's box. Two things placed at coordinates chosen independently will eventually meet.
Copy a reference's grammar, never its content. The reference board bullets its rows with a drawn eye and fans emphasis marks beside its phrases, but those belong to that board's brand. Reproduced on another script they are decoration that means nothing. Filler has to be about something: a drawing of what the line refers to, never abstract specks.
Two reviews, both human. The storyboard still costs seconds and is where relevance is
judged — whether a mark is about the line it sits under, whether the payoff says the closing
line. No automated check decides that. Then the finished cut, watched end to end with the
watch skill.
It runs on any machine. The two board fonts ship in assets/fonts (see the NOTICE.md
there), and the image tool is found beside this skill. The caption pill uses Segoe Script
Bold where Windows has it and Permanent Marker everywhere else, so captions look slightly
different off Windows.
The caption pill ends above y=1520. review-finished-ad flags the bottom 400px, so the
default caption_y is 1440 and the layout keeps every row, drawing and the payoff above the
pill. An episode.json written before this keeps its own caption_y; delete the key and
re-solve to move it.
The board photo carries no logo. The prompt forbids a maker's badge. Look at the plate anyway: a real manufacturer's name on the frame is a third party's mark in a brand's ad.
The voice comes with its word timings. create-vo-elevenlabs returns audio only, and this
format needs a timing for every word. scripts/gen-voice.py --project <dir> reads vo and voice
from beats.json and calls the proxy's /with-timestamps route, writing voice/vo.mp3 and
voice/words.json, which make-episode.py requires. A dry run prints the character count and the
cost; --yes spends. It routes through the GooseWorks proxy (scripts/media_proxy.py), never the
speech provider directly, and refuses a payload it has already paid for.
The build refuses to continue when any of these fail, and names what broke:
The board photo and the drawings are one-time per brand. A second video for the same brand costs only the voiceover. Everything else — the layout, the lettering, the tracing, the render and the master — is free and local. Rendering a half-minute takes under three minutes.
Measured against the filmed reference this reaches about a fifth of its ink cover, because the reference is a time-lapse of a much longer drawing session while every mark here waits for its word. Filler narrows that and does not close it. If a brief needs a genuinely full board, the script has to be longer or the marks have to stop waiting for words.