← writing
1 September 2026

The grain of randomness: what decides a repetition

A generator of repeated fields is usually controlled by a single number, the quantity of randomness. Yet that is the less important parameter. What the image looks like depends mainly on the level at which the randomness applies, and that level appears nowhere in the code.

The starting point is a conversation with a model, published as is, in which a generative drawing is requested in the style of five artists in turn. The answer always has the same form: three or four geometric shapes and a single slider, called chaos, disorder, intensity, accumulation or obliteration depending on the artist. That slider always controls the same thing, a quantity of randomness.

A quantity of randomness says almost nothing about what will be seen. The same amount produces very different fields depending on where it applies: the position of each element, the interval between two elements, whether an element is present, or its size. That level has no name in the generated code, and it is what determines the result. We will call it the grain.

01What the single slider conflates

A generator written in one go mixes the four levels. A single line of code draws a number and applies it to position, size and opacity at once, so the slider acts on everything at the same time. The result can be adjusted but not interpreted: when disorder increases, it is impossible to tell which of the four decisions produces the visible effect.

The generator below separates them. The four grains use the same stream of numbers and consume the same amount of it: at equal quantity, only the target of the randomness changes.

02Four grains, one quantity
generator

Set the quantity of randomness, then change the grain without changing that quantity. The four grains use the same draw and differ only in the level at which it applies, which is enough to produce four very different fields. The names given indicate which artist worked with a procedure, not a particular work.

grain

family

A generated field of repeated shapes, on a light or dark ground depending on the chosen surface.

The experiment is simple: leave the quantity as it is and change the grain. Nothing else changes, and the field becomes unrecognisable. Then set the quantity to zero: the four grains then produce the same field, position by position, whatever the seed. This reference point makes the rest measurable. The third slider, the hand, is covered in section five.

03Same dose, unrelated departures

The measured departure is the distance between where the rule puts an element and where it ends up. At sixty per cent quantity, on the rows family, two hundred draws give this.

grainmean departurelargest departuresource
position1.15 px2.29 px396 shapes
drift1.96 px7.59 px396 shapes
presence0 px0 px319 shapes
weight0 px0 px396 shapes
Two hundred draws per row, quantity fixed at 60 %, density at 55 %, rows family. The figures come from the generator above and describe the model.

Two grains move no element at all and are still visible on the field, one by removing seventy-seven elements, the other by changing only their size. Between the two grains that do move elements, the largest offset varies by a factor of three. The quantity is the same from one row to the next.

04Drift is a random walk

The two grains that move elements differ on one point only: the first draws an offset for each element, the second adds that offset to the previous one. That addition changes the nature of the result. Over four hundred draws, the offset of the position grain stays steady, around 0.8 pixel, whatever the element’s rank in the row. The offset of the drift grain grows as the square root of the rank: the ratio between the offset and that root stays at 0.35, to within a hundredth, from the second to the forty-sixth element.

That is the characteristic of a random walk, and it has a practical consequence. No step is large: each element is placed roughly where it should be relative to the previous one. But the error is never corrected, and by the end of the row the offset from the model is nearly three times that of the independent grain, for the same quantity. A rule can therefore stop being followed without any local deviation being visible.

This is what distinguishes a hand from a machine. A hand placing marks side by side does not position each one on an invisible grid, but relative to the previous one. Its irregularity accumulates, and a drawn line drifts from its axis without any single interval being wrong. A generator that draws each offset independently never produces this effect, whatever the quantity.

05The hand, independent of the grain

A field whose randomness bears entirely on one parameter looks like a diagram: the marks are perfect rectangles, all in the same ink and the same weight. Disorder is already present; what is missing is material. The third slider adds it: rotation, weight, opacity. It never changes position or presence, and that separation can be checked.

With the hand at its maximum, the mean offset stays at zero for the grains that move nothing, and unchanged for those that do: with the drift grain set to twenty per cent, the offset is 0.65 pixel whether the hand is at zero or at a hundred. The two settings are therefore independent, which justifies keeping them apart. A sheet can be covered in tremor without moving the rule by a pixel, while a small amount of the right grain is enough to break it.

06What each family lets the grain show

The four families apply the same grains to four different repetition techniques, each inspired by an artist who worked with it. They do not all show each grain equally well. The grid is where the position grain is clearest, because a square lattice gives the eye a reference; drift there becomes hard to tell apart from rotation. The rows are the reverse: they are the only family where drift is clearly visible, since a single line shows its accumulation.

The dots correspond to the presence grain, because a field of dots is judged less by where each dot sits than by how much ground remains between them. The discs are a special case, for a reason unrelated to grain: the technique relies on the effect of their overlap more than on their position. The groups of discs are placed on the golden angle, with no rows or columns, and the filled circles blend where they intersect instead of hiding each other. This family uses seven hues reserved for this figure, because the site palette, monochrome with one red, cannot produce simultaneous contrast.

07The paper/ink switch, reused

The field opens full screen, with the same controls as the tool built for the inverted drawings: left and right arrows change family, up and down change surface, the space bar draws again, and escape closes the view. On a touch screen, a horizontal or vertical swipe has the same effect, depending on the dominant direction of the gesture.

Something was lost along the way. On a scanned old sheet, inverting the image requires a filter: a matrix computes a negative, tints the luminance and restores the red so that vermilion stays vermilion. The filter is needed because the paper is data whose treatment has to be decided. Here there is no paper. A generated field is a list of shapes, each indicating which of three inks it uses, and the switch reassigns three properties without any calculation. It is simpler, but the tool loses something: on the old drawings, it revealed things because it worked on material that existed before it. Here, everything displayed was decided by the sliders.

08Other grains, and the selection

These four grains are not the only possible ones. Randomness can also affect the order of elements, the number of rows, the length of a line, or apply at two levels at once with different quantities, which is probably the most common case in a real drawing. The generator applies only one at a time, because combining them would prevent the demonstration, and that mixture was the flaw to correct.

This article says nothing about the quality of what is produced. A randomly drawn field is rarely successful, a selection still has to be made, and the sliders do not make it. Nor does it say anything about what matters most in the works mentioned: scale, material, and the fact that a sheet covered in ink marks is looked at from close up. An image three hundred pixels wide cannot show that.