heygen-com/hyperframes-community-skills

p5-paint-animation

Turn a text prompt, a photo, or a live video clip into hand-made-looking p5.js animation — pencil handwriting that writes itself, photos repainted as brushstroke art with paint-on animation and camera moves, and short videos (iPhone Live Photos included) re…

Vedi sorgente
Documento Skill originale

Contenuto dal repository con titoli, esempi, codice, tabelle, link e immagini preservati.

p5 Paint Animation

Three flows, one engine. Pick by input:

InputFlowOutput
Text prompt / phraseA: Write-onHandwriting animates onto ruled paper, pen-tracked for camera
PhotoB: Repaint + paint-onBrushstroke painting, still or painting-itself film, camera pull
Video / Live PhotoC: Living paintingEvery frame repainted over a locked stroke lattice

When to use / when not

Use for: hand-made-looking motion — handwriting that writes itself, photos as brushstroke paintings (still or animating), short clips as living paintings. Do NOT use for: crisp typography or brand-accurate text (use real fonts), photoreal output, footage longer than ~10s (render cost scales per frame ≈ 5–20s each), or live/interactive pages (this renders offline MP4/PNG).

Requirements and side effects

  • Requires: Node 22.12+, ffmpeg on PATH, ~500MB disk (Chromium).
  • sh scripts/setup.sh (run manually, once) uses the committed lockfile to

download pinned dependencies from the npm registry: puppeteer 25.9.0, p5 2.3.2 (LGPL-2.1), and p5.brush 2.2.1 (MIT). Puppeteer downloads its pinned Chrome for Testing build from Google's download service. Read-only downloads; no credentials, no cost, nothing uploaded.

  • Renders launch sandboxed headless Chromium on a blank page. Network and local

file requests from sketches are blocked; sketches receive inputs as data URLs.

  • Writes only inside paths you pass (out.png / out.mp4, plus a sibling

.track.json and, with --keep-frames 1, a _frames_* directory).

  • Optional, only for flow C subject matting: `npx hyperframes

remove-background` (downloads the HyperFrames CLI from npm; local CoreML inference, no data leaves the machine).

Verify success

Render any command below, then open the output: the last frame must show the complete artwork (text fully written / painting fully covered), and re-running with the same --seed must produce identical bytes on the canvas. The smoke test used for review: node render-anim.mjs sketches/handwriting.anim.js /tmp/t.mp4 --size 1600x900 --seed 7 --inject 'PHRASE="go make something"'.

Setup (once)

sh scripts/setup.sh

Installs the exact dependency tree recorded in scripts/package-lock.json. All commands below run from scripts/.

Flow A — prompt → write-on animation

node render-anim.mjs sketches/handwriting.anim.js out.mp4 \
  --size 1600x900 --fps 24 --seed 7 --hold 12 \
  --inject 'PHRASE="your text here";SIZE=170'
  • Full alphabet: lowercase, caps, digits, ? . , - / @. Wraps automatically.
  • Dials via --inject: SIZE, INK (hex), BRUSH ("2B" pencil, "crayon",

"charcoal"), PAPER=1 for blank paper.

  • The run writes out.track.json — per-frame pen position. For a

camera-follows-pen move, crop each frame along the smoothed track (exponential follow, gain ~0.14, camera y LOCKED to the baseline — never follow the pen tip's y or the camera bounces on every ascender).

  • For scraggly/messy writing raise the per-letter jitter in the sketch; for

word-sync to a voiceover, transcribe first and start each word's strokes at its spoken time.

Flow B — photo → painting (+ paint-on film + camera)

Still:

node render.mjs sketches/repaint.anim.js out.png \
  --size 720x900 --seed 3 --image photo.jpg --inject "STYLE=0;DETAIL=3"

Paint-on film (the painting paints itself, coarse-to-fine, painter-ordered — calm regions first, the face last):

node render-anim.mjs sketches/repaint.anim.js out.mp4 \
  --size 720x900 --seed 3 --image photo.jpg --fps 14 --hold 40 \
  --inject 'STYLE=0;DETAIL=3;SUBJECT="0.5,0.5,0.44,0.55"'
  • Crop to the subject first. Face size in frame is destiny: selfie-distance

faces resolve beautifully; small faces stay figures. Portrait 4:5 or 9:16.

  • Dials: STYLE 0 pastel / 1 gouache / 2 charcoal-sketch. DETAIL 1 painterly,

2 faces resolve, 3 soft features (noses), 5 portrait-grade (adds a second fine lap). LOOSE 0.7–1.5 stroke wildness. SAT color compensation (~1.1).

  • SUBJECT="cx,cy,rx,ry" (image fractions) paints everything inside that

ellipse through all passes FIRST and drops a subject-done frame marker. Camera pull-out recipe: hold tight on the subject, then at the marker frame ease wide with ffmpeg zoompan (cosine ease, ~2s): zoompan=z='if(lte(in,M),1.9,1.85-0.85*(0.5-0.5*cos(PI*min((in-M)/28,1))))' (upscale 2x lanczos first so the tight phase stays sharp).

Flow C — video / Live Photo → living painting

node render-video.mjs clip.mov out.mp4 \
  --size 720x900 --fps 12 --seed 7 --inject "STYLE=0;DETAIL=2"
  • The stroke lattice is built ONCE from frame 1 and locked; only colors and

orientations resample per frame. Strokes hold still — the painting never "boils" while the subject moves.

  • Live Photos: get the paired .mov (share-sheet Options → Live Photo ON,

or Photos.app → Export Unmodified Original). If it arrives as bare JPEG the motion was stripped in transit — ask for a 2–3s normal video instead; that's better input anyway.

  • Motion triage first. Live Photos include the phone swinging. Measure

per-frame motion (ffprobe -f lavfi "movie=in.mov,signalstats" → YDIF), paint only the calm window with --start/--dur, and ping-pong the result (forward+reverse concat) for a seamless breathing loop.

  • Big camera motion? WORLD=1 tracks the camera by image correlation and

lets strokes live in world coordinates — the camera pans across one growing painting. Translation-only; fast zooms drift.

  • Subject-only painting: matte frames first (`npx hyperframes

remove-background), pass the matted PNGs with --frames DIR and MASKED=1, and set BG` to your paper color.

  • Dark footage: grade with gamma only (eq=gamma=1.2). Never boost

saturation before painting — shadow chroma noise becomes literal teal paint.

QC gates (all flows)

  • Read the actual output frames; never ship on intent. Check: faces resolve,

no bare-canvas gaps, first/mid/last frames coherent.

  • Same seed = identical render. Iterate one dial at a time against a fixed seed.
  • For anything joining a larger film, verify seams numerically per

motion-doctrine — this skill produces material, not final grammar.

references/flows.md has full worked recipes; references/gotchas.md is the list of failure modes already hit so you don't hit them again.

dallo stesso repository

Altri Skills

Tutti gli Skills
heygen-com
Community

duo

Composite any two HTML screens into a locked photo plate of hands holding an open foldable phone — a 1448×1086 still with two blank 495×849 / 498×849 screen slots (left and right) — to make side-by-side meme and comparison videos: TikTok vs Reels, ChatGPT vs Claude, "me vs my friend", before/after, two apps racing. Trigger on: "foldable phone meme", "put X on the left screen and Y on the right", "two-screen comparison video", "the dual-screen phone meme", or any request that names two apps/feeds

installazioni
2
GitHub Stars
149
Aggiornato
24 set
heygen-com
Community

prod-by-claude

Use when an agent needs to make a HyperFrames video of a Strudel live-coding track, where the code itself is the picture and every token lights up on the note it plays. Trigger on: "strudel video", "live coding video", "show my code playing", "make a video of this track with the code", "prod by claude". Covers composing the track, recording it offline from pinned Strudel, and building the 1080x1080 video (typed opener, parts powering on, real Strudel theme switches, graffiti ASCII drops, typed s

installazioni
2
GitHub Stars
149
Aggiornato
24 set
heygen-com
Community

session-story

An agent makes an animated film of an average session with its own user, beginning to end, told with that user's real messages: whatever their routine is (code, writing, data, design, video). The agent sits at a desk in a room decorated from what it knows about the user, and the monitor shows their kind of work. Requests glide in as paper planes and unfold. Corrections are a 3D cartoon mallet that bonks, then turns to show the words on its face. Praise is a butterfly towing the words on a ribbon

installazioni
2
GitHub Stars
149
Aggiornato
24 set
heygen-com
Community

vox-explainer

Create a 60–90 second, collage-style HyperFrames explainer from a hiding-in-plain-sight topic or supplied documents and links. Use for requests such as 'make a plain-sight explainer', 'Vox-style history of X', 'why is X everywhere?', or 'turn this document into an explainer'. Covers topic or source routing, research, script, static design approval, voiceover timing, motion continuity, build, and numeric quality gates.

installazioni
2
GitHub Stars
149
Aggiornato
24 set