INTERACTIVE GENERATIVE ART STUDY / 01
Can a flock paint?
Small choices, shared motion, accumulated marks. Watch a flock turn movement into a painting.
ABOUT THE STUDY
A flock that leaves a record
What is it?
Each small triangle is a boid: an agent that steers with its neighbours. Here, its path lays down translucent pigment. The marks accumulate into an image that changes with the flock size, colour mode, and your pointer.
How was it designed?
The study separates motion, intent, and painting style. Flocking physics keeps the agents moving together. Six bounded intents—follow, explore, rejoin, evade, reinforce, and quiet—change their speed, turning, and brush marks. Three palettes give the same system different visual moods.
What can I try?
Compare Plain flock with Browser rules, change the palette and pigment, move your pointer through the group, then save a PNG. A seed labels each new painting so its starting arrangement is identifiable.
The model-driven version
The complete local source also includes optional Ollama and Jev modes. In those modes, a model selects intents for eight leaders about every four seconds and nearby boids inherit the choices. The Ollama version was tested on an RX 6700 XT; Jev uses OpenRouter through a localhost server with a spending cap. Those modes are experiments and are not connected to this public page.
What we learned
What we saw
The boid physics, palettes, and brush renderer already produce most of the visible image. In informal viewing, the model-directed results were less different and less compelling than hoped. That is an observation, not a controlled artistic verdict.
Why the influence is small
This prototype gave the model seven coarse fields per boid (an id and six state readings) and asked for one of six intent labels for eight leaders about every four seconds; nearby boids inherit those labels. The original idea imagined richer awareness of the canvas and several bounded choices. In the current code, “explore” adds wandering rather than seeking empty canvas, and “reinforce” changes stroke weight rather than finding a promising mark to build on. The model therefore has limited influence on composition.
What the tests showed
The local tests established that the connection works and measured latency and cost. They did not establish whether model decisions can improve this artwork.
Next experiment
Keep seeds comparable and force intents manually to see whether the mapping from behaviour to marks is legible. Then add richer spatial or global canvas context, and meaningful painting actions, before comparing rules, Ollama, and Jev. Simply calling the model more often is not the first fix.