Generative Motion

Generative Motion

The gap between a generative sketch and a generative artwork is almost never the algorithm — the algorithms are public and short. It is art direction: what the colour means, where the density sits, how slowly it evolves, and what you left out.

A tutorial renders the algorithm. A piece renders a decision.

Workflow

1. Write the intent first, in two or three sentences: what the system is (embers on

a thermal, ink in water, a colony finding food), what the viewer should feel, and the one property that earns colour. Skipping this is what produces rainbow noise. The design-direction skill's philosophy format applies directly.

1. Pick the system from [references/systems.md](references/systems.md) — each entry

has the equations, the parameter ranges that actually look good, and its failure mode. Its numbers are machine-checked: validate/attractors.py and validate/gray_scott.py verify every attractor seed and reaction–diffusion pair on CPU in seconds. Run them after changing any value, and use them to find your own.

1. Build the field and the motion with

[references/noise-and-fields.md](references/noise-and-fields.md) — noise, fbm, domain warping, curl. Most beautiful motion is a field being sampled.

1. Decide CPU or GPU early with [references/gpu-and-shaders.md](references/gpu-and-shaders.md).

This is architectural, not an optimisation — some systems simply do not express themselves below a population the CPU cannot reach, and porting later means rewriting.

1. Render it well with [references/rendering.md](references/rendering.md) — trails via

accumulation, additive blending, envelopes so nothing pops.

1. Map the field to the screen with [references/density-and-tone.md](references/density-and-tone.md).

Accumulation buffers are unbounded and displays are not; this mapping is where technically-correct pieces most often look wrong.

1. Finish it with [references/post-processing.md](references/post-processing.md) — bloom,

grain, vignette, grading. The cheapest quality per line of code in the discipline.

1. Art-direct it against [references/art-direction.md](references/art-direction.md) —

the checklist that separates a piece from a screensaver.

1. Watch it. Record with ../motion/scripts/capture-motion.py and look at the result

over a long run, not one frame. Generative work fails slowly: it looks great at 4 seconds and turns to grey mush at 60. When something is wrong, work [references/diagnostic.md](references/diagnostic.md) rather than tuning at random.

Which reference to load

| Situation | Load | |---|---| | Choosing a system; attractors, boids, physarum, reaction-diffusion, differential growth | references/systems.md | | Noise, fbm, domain warping, curl fields, flow fields | references/noise-and-fields.md | | Trails, blending, buffers, performance at scale | references/rendering.md | | Going GPU: WebGL2 GPGPU, ping-pong, the float-extension minefield, shader systems | references/gpu-and-shaders.md | | 3D: raymarched SDFs, smooth blending, domain repetition, instancing, camera | references/three-dimensional.md | | Flat white, thin outlines, or any density-to-colour mapping | references/density-and-tone.md | | Bloom, grain, vignette, aberration, feedback, colour grading | references/post-processing.md | | It runs but looks like a tutorial | references/art-direction.md | | It's black / saturating / dying / clumping / 4fps / grey mush | references/diagnostic.md |

Non-negotiables

• Colour is earned. Map it to a property the system actually has — velocity,

curvature, age, density, divergence — never to a rainbow of hue over time. If colour encodes nothing, the piece reads as a demo.

• Seed the randomness. A composition you like must be recoverable; ship the seed.
• Delta-time everything, clamped. A field that runs at double speed on a 120Hz display

is a broken piece, not a fast one.

• Evolve slower than feels right. Field evolution that reads as "alive" in the editor

usually reads as "busy" on the wall. Halve it, then halve it again.

• Density needs negative space. A frame at uniform particle density is noise. Vary it,

or crop into it.

• Verify parameters before quoting them. In attractors and reaction–diffusion alike, a

value just outside the viable band produces nothing rather than something worse, so a plausible-looking number is not a usable one.

• Long-run test. Run it for a minute before shipping — trails saturate, particles

clump in attractor basins, and energy either dies or explodes.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jangles-byte
• Source: jangles-byte/atelier
• License: MIT

Install and usage instructions live in the source repository linked above.