AGENT SKILL

algorithmic-art

Generative p5.js art. OmniSkill Design crew — builds seeded, parameterised generative systems that produce a coherent family of outputs rather than one lucky image.

License 许可
MIT
所属仓库
IPedrax/OmniSkill
最近更新
2026-10-14
Stars
★ 9
Forks
0

01 前言 Front Matter

name
algorithmic-art
description
Generative p5.js art. OmniSkill Design crew — builds seeded, parameterised generative systems that produce a coherent family of outputs rather than one lucky image.
disable-model-invocation
true

03 SKILL.md 正文

Algorithmic Art

Generative p5.js art · OmniSkill Design crew

Code-driven visual art. The deliverable is a system that produces a family of good outputs, not a single image that happened to work.

The discipline

Seed everything. Use a seeded PRNG, never bare Math.random(). Without a seed, a good output cannot be reproduced, printed at higher resolution, or iterated on — it is gone the moment the page reloads.

let seed = 12345;
randomSeed(seed);
noiseSeed(seed);

Display the seed on the canvas during development and make it settable from the URL. Then any output can be recovered exactly.

Parameterise, then tune. Pull every magic number into a named parameter object at the top. Tuning is where the work actually happens, and constants buried in draw calls cannot be tuned.

Judge the distribution, not one frame. Generate a contact sheet of 25–100 seeds and look at them together. A system whose median output is good is a good system. A system with one brilliant output and 99 poor ones is not — it is one image with extra steps.

Techniques worth knowing

Perlin/simplex noise — the foundation of most organic generative work. Smooth, continuous, controllable.

// Third dimension as time gives coherent animation
let n = noise(x * scale, y * scale, frameCount * 0.01);

Scale controls feature size; octaves control detail. noise() returns roughly 0.3–0.7 in practice — remap it rather than assuming a full 0–1 range.

Flow fields — noise drives an angle, particles follow it. Produces the characteristic organic streaming look.

let angle = noise(x * 0.005, y * 0.005) * TWO_PI * 2;
p.add(createVector(cos(angle), sin(angle)));

Recursive subdivision — split a rectangle repeatedly with a probability of stopping. Reliable structural variety with very little code.

Packing — place shapes, reject on overlap, retry. Simple and dependable; cap the attempts so it terminates.

Differential growth, reaction-diffusion, L-systems — for organic and botanical forms.

Physical simulation — attraction, repulsion, collision. Emergent structure from simple local rules.

Composition

Technique alone produces noise. What separates generative art from generative output:

  • Constrain the palette. 3–5 colours chosen deliberately. Random hues look random. Sample from a real source — a photograph, a painting, a curated palette. See /omniskill:ui-ux-pro-max.
  • Vary density, not just position. Rest matters as much as detail; uniform density reads as texture rather than composition.
  • Establish hierarchy. Something should dominate. Everything equal is nothing.
  • Break the grid deliberately. Perfect regularity is dead; controlled irregularity is alive.
  • Layer. Depth from overlap, transparency, and scale variation.

Performance

draw() runs 60 times a second — do not allocate there. Precompute in setup(). For static output, render once and call noLoop().

Beyond a few thousand elements, drop p5 vector objects for flat typed arrays. Use the p2d renderer unless WebGL is genuinely needed.

Export

Export at print resolution from the start — retrofitting is usually a rewrite. Either scale the whole system by a factor, or use createGraphics() at target size and draw into it.

Vector output (SVG via p5.svg) where the work is line-based and needs to scale or plot. Raster (PNG) where texture and blending matter.

Rule of thumb for print: 300 DPI at final physical size.

Route final image assembly and PDF output to /omniskill:canvas-design.

Deliver

The sketch with named parameters, a seed control, a contact sheet of representative outputs, and export at target resolution. Note the seeds of the strongest outputs — those are the results.

Checks before delivering

  • Seeded PRNG throughout; no bare Math.random().
  • Any output reproducible from its seed.
  • Parameters named and grouped, not inline constants.
  • Contact sheet reviewed — the median output is good.
  • Palette is deliberate, not random.
  • Nothing allocated inside draw().
  • Export resolution matches the intended use.

04 同仓库更多 Skills