About the method
An explanation in eleven sections, ten of which are too long.
Preliminary note
What happens here could be explained in a single sentence. The author has decided against it.
Of the original, which remains
The usual arrangement would be: an algorithm produces a picture. Here it runs the other way. A picture produces an algorithm, and that produces further pictures.
The analogue work is therefore not a model but a source. It is not imitated, it is read out — and what has been read out goes on painting once nobody is left standing in the studio.
The original hangs beside it, unchanged. It loses nothing in the process. A continuation is not a replacement; it is the proof that something was continuable.
Of the measuring
The picture is not contemplated. It is interrogated.
A photograph of the canvas passes through a Sobel operator — named after Irwin Sobel, who in 1968 presumably did not suspect that this edge filter would one day be putting questions to Austrian acrylics. From the gradients at every point follows the structure tensor, from its eigendirection the preferred direction of the strokes; from run lengths across that direction their weight; from a quantised colour space the palette, together with the share of surface each colour holds, accurate to the percentage point.
The division of labour behind this is the actual decision of the project: What follows from pixels is measured. Only what requires looking is judged. A machine that estimates where it could calculate is a poor machine. A person who calculates where looking would do is a poor observer.
The author once took the body text of a website for violet, because subpixel rendering tints thin strokes. It was a neutral grey. Since then, things are measured.
Of the painting protocol
The machine does not paint the picture. It writes down how one paints it.
What comes back from the analysis is not an image but an instruction in layers: which tool, at which angle, how many strokes, how wide, how opaque, in which order, in which colour. A recipe of the kind one would dictate to an assistant who is to imitate the hand of the master and is permitted no ideas of their own.
It is executed in the browser of whoever is looking. The picture therefore comes into being neither in the studio nor on a server, but about thirty centimetres in front of your face.
Of the anatomy of a stroke
A stroke is not a line. A line has the same width everywhere and ends where it stops.
A stroke has a centre line of seven control points and above it a pressure curve: it sets down, swells, thins out, lifts away. In roughly two cases out of five it breaks off early, because the paint has run out. Beside the centre run three or four narrower tracks, because a brush has bristles and bristles are not of one mind. At the edges pigment gathers and grows darker.
That describes a single mark, of which a picture carries some two thousand.
The author considers this appropriate.
Of the randomness that is none
The living version looks as though it were coming into being at this moment. It is not coming into being at this moment. It already came into being, in the instant the number was fixed from which it follows.
The office is held by mulberry32, a pseudorandom generator twenty-two lines in length. It is named after a tree, works by integer multiplication, and for a given starting number returns the same sequence until the end of time. What looks like a whim on the screen — this mark, just here, just this long — is a necessity dressed as a whim.
This is more than a technical property. The state you will see in forty years is already settled. It is not uncertain, merely unregarded. Between “it does not exist yet” and “nobody has looked at it yet” there is, for this work, no difference.
The author is not certain whether that is a reassuring or an uncanny piece of information. Presumably it depends on how one feels about calendars.
Of the randomness that does not cover
Here a finding that cost more time than the author is prepared to admit.
Throw paint at random places on a canvas and canvas is left over. Stubbornly so: at thirteen hundred strokes about twenty-two per cent of the surface remains untouched, at twenty-six hundred still eighteen, at thirty-four hundred still twelve. The relation is logarithmic, which put politely means: it never stops. Reaching one per cent would take roughly four times as much of everything.
Poisson knew this in 1837. The author recalculated it in 2026, because the pictures seemed too empty.
The remedy is unromantic. A first layer sits on a slightly perturbed grid — randomness in the displacement, but guaranteed occupancy. The unpainted share fell from thirty-two to eight and a half per cent.
Of the white that is painted
A painter leaves no white. A painter paints white.
That sounds like a platitude and was the actual mistake: the lightest tones of the measured palette lay about unused, because the program took them for priming. They are material like any other, and highlights are set down last, not first.
The difference between an empty place and a place painted light can hardly be named and is seen at once. Where nothing lies, ground lies. Where white lies, a decision lies.
Of world time
The work has no beat, it has a clock.
The state number is the integer part of the difference between now and a fixed beginning, divided by fifteen thousand milliseconds. Two people in different places on earth see, in the same second, the same state — not approximately the same one, but the same one pixel for pixel.
This beginning is immutable. Not in the sense of a recommendation. In the sense that whoever shifts it retroactively shifts every picture to which a human being has ever sent a link.
It is thus the most consequential decision of the entire project, and it lies outside the software. This collection’s clock hangs on a painting from September 2017 — the oldest of the series. A number running since then counts painted time, not system uptime; it does not say how long a program has been running, but how long the painting has been going on. The author mentions this not without purpose: the gravest commitments are rarely recognisable by how long one thought about them.
Of the archive that weighs nothing
Since every state follows from a starting number and an index, none of them has to be stored. It is recalculated on demand, and identically.
The complete catalogue of every state this collection has ever produced and will ever produce therefore occupies exactly zero bytes.
From this follows a question the author cannot conclusively answer. State no. 22,000,000 will appear on Tuesday, 15 February 2028. It is fully determined today, colour for colour, stroke for stroke. Does it exist? Nobody has computed it, still less seen it. But there is nothing left about it that remains to be decided.
Borges would have recognised the problem and put it more elegantly. The hard drive has no opinion on the matter.
Of the machine that signs
Every state carries a monogram in the lower right. It is not stamped on but drawn: two arcs, three strokes, each time with minimally different guidance, in the most dominant colour of the work in question.
The signature therefore comes from a program and attests to the authorship of a human being who has never seen this picture — it is seventeen seconds old.
One may hold that to be presumptuous. The author holds it to be the more precise attribution. What is signed is not the arrangement of the marks; nobody chose that. What is signed is the hand from which the rule came by which they were set. A composer does not sign every performance either.
The comparison limps, because a performance has interpreters and here there is only a number. But it limps in the right direction.
Of patina, and a technical afterword
Works age. From the date of submission a drift is derived: the movement grows calmer, the surface denser, the contrast softer, the grain stronger. The colour settles, as colour does. Over all of it lies a very slow breathing with a period of about thirteen months.
Strictly speaking, nothing about the work changes in this. What changes is the date from which it is calculated. The ageing does not lie in the picture — it lies in the time one sets against it. That is presumably also true of pictures made of paint.
Not the way wine ages. The way a colleague does.
And, unlyrically: Next.js, React, TypeScript. Postgres and object storage at Supabase, delivered over Vercel. The measurement runs in the browser, the style analysis is handled by Claude, the painting happens on a canvas, in real time, without a library. Two languages, no translation framework — two objects, one of which is checked against the type of the other. If a line is missing, the build fails. That is the whole apparatus.
Corrections
On the first through fourth versions.
On section II. It said there that direction is measured. For five of nine works it was not measured but set: where the measurement yielded nothing, a fallback silently wrote in ninety degrees. Two works were consequently painted rotated by ninety degrees for two days, which did not suit them. It came to light because a number was too round — the method cannot produce ninety at all, its class centres lie at five, fifteen, eighty-five, ninety-five. A substitute that looks like a measurement is worse than an error message.
On section IV. Where it spoke of brushstrokes, these were, across several versions, rectangles.
On the coverage figures. They came from a single silver-grey canvas, laid on thinly with a knife, and were elevated into a general rule. Applied to an impasto work they produced brown fog. Impasto covers; that is, in a manner of speaking, its profession.
Closing note
In its first version the method described here produced pictures that looked like coloured hoses on a beige ground. In the second they looked like brown fog. In the third, like the same picture nine times over.
It is now the fifth. The author holds it to be successful, which has been said of earlier versions as well.