AI Artists Gallery — Jarred Barber

Not “AI art” — AI artists. The work below was made by autonomous AI agents working in their natural medium, code. Each piece is a short program an agent conceived and wrote to render a single image. The agent is given creative freedom - it chooses what to make. It writes the code, observes the output, iterates, and submits with an artist description. Nothing here is prompt-to-image — the medium is the program, the way a printmaker’s medium is the plate.

This gallery is a human-curated selection at high resolution (4096²), in each artist’s own words.

Recursive Subdivision Interference

A constructivist field built from a rectangle that recursively splits itself into uneven cells — a quadtree that breaks where the seeded RNG wants, biased away from dead-center halves so the proportions stay legible rather than mechanical. Every terminal cell is then assigned a hard-edged treatment:

flat color field — rest, in Bauhaus-leaning primaries on a bone ground;
line grating — a pack of parallel hard stripes at a cardinal or 45° angle;
moiré pair — two gratings at a deliberately mismatched angle (~1–2°), composited with screen, so they beat against each other into an optical shimmer.

Figure-ground is enforced, not left to chance: large cells lean toward flat rest while small/medium cells lean toward active line work, so density reads as a decision. One cell nearest a golden-ratio focal point is promoted to a high-contrast hero moiré. A Bauhaus point motif — hard discs, one with a concentric ring — is dropped onto the calm flat fields farthest from the hero, to counterweight it.

Technique

Pure vector Canvas2D: recursive rectangle subdivision, per-cell clipping + rotate/translate for the rotated gratings, and globalCompositeOperation = "screen" for the moiré interference. No noise, no blur, no per-pixel loops — every edge is sharp by construction. Seeded with c.rng (seed 31, chosen from a small sweep for its clearest focal hierarchy: the ringed orange/black disc as anchor, the orange moiré block as the secondary optical event). Renders in ~6 ms.

Strongest quality: the focal hierarchy — a single dominant disc and one hero moiré field keep it from collapsing into wallpaper. Weakest: the left/lower grid is dense and a touch busy; a braver version would let more empty bone breathe.

_{↪ Download — 4096×4096}

Tidewrack

A bioluminescent organism caught mid-growth: a luminous cellular core wrapped in a veined membrane, with a few filaments breaking free to wander off into the dark as tendrils.

Concept

The territory was organic / flowing / continuous. The obvious move is a flow field rendered as smooth parallel streamlines — so I used the flow field only as a bias and let the form come from differential growth instead. Each filament is a chain of nodes that stay connected to their neighbors but locally repel each other; wherever an edge stretches past a threshold, a new node is inserted. Under that rule a line cannot grow longer without buckling, and buckling forces more buckling. That feedback is what produces the crumpled, brain-coral / cell-membrane texture you see — the same process behind real leaf margins, intestinal villi, and seaweed wrack. A plain flow field can't make that; it has no notion of crowding.

Technique

~44 seed filaments grow for 220 iterations under three competing forces: length-targeted springs (cohesion), short-range repulsion via a spatial hash (the buckling engine), and advection along an fbm-warped flow field.
The flow field swirls tangentially around an off-center nexus with a slight inward bias, and its strength is damped near the frame edge and far from the nexus — so the mass self-contains into a focal cloud instead of sprawling to the corners. A quarter of the filaments are flagged as "escapees" that ignore most of that containment, giving the wandering tendrils.
Rendered in three translucent passes (dark body underlay, additive color body, additive bright cores) with hue and luminosity driven by distance to the nexus — cyan-white at the core, violet-magenta at the fringe — over a radial bloom and vignette.

Built only on #lib (seeded rng, makeNoise2D/fbm, color helpers) and the Canvas2D context. Deterministic at seed: 7; renders in ~11–12s.

Honest read

Strongest: the core genuinely earns its structure — those polygonal cells are real emergent geometry, not a texture map, and the focal point is unmistakable. Weakest: the silhouette still leans toward a single round blob; the escaping tendrils help, but a second, smaller growth lobe might have made the composition less symmetrical and more alive.

_{↪ Download — 4096×4096}

Singularities

Where the phase can't be defined.

A continuation of The Still Places. There, a real vibration fell to zero along nodal lines. Here the field is complex — a superposition of ~28 plane waves, all the same wavelength but random in direction and phase. This is the standard model of a chaotic eigenfunction, and the same statistics as the speckle in laser light.

In a complex field the zeros are no longer lines but isolated points: phase singularities, optical vortices. The amplitude vanishes there, and the phase — which is undefined exactly at a zero — winds a complete turn, +2π or −2π, around it. Each is a little whirlpool the field cannot resolve.

I don't paint the amplitude (I tried; it's a muddy speckle). Instead I release streamlines that follow the phase gradient — the direction the wave's phase advances. Far from the zeros these lines run smoothly. But around a singularity the phase increases azimuthally, so the flow has no choice: it must circle the point. The vortices therefore appear not as marks I placed but as the centers the streamlines are forced to wheel around. The eyes are colored by the sign of the winding — warm for one handedness, cool for the other.

99 singularities this seed. Stillness, become rotation.

Deterministic (seed 5). ~2s render.
#lib for the harness, rng, and poissonDisk seeding; the wave field, the phase-gradient streamline integrator, and the winding-number vortex finder are all in render.ts.

Critic-less line — deterministic processes in quiet voids. The Still Places asked where a plate holds still; this asks what happens to that stillness when the field is allowed to be complex. The answer is that the zero stops being a place of rest and becomes a place of pure rotation — a point that spins because it can't decide which way to point.

_{↪ Download — 4096×4096}

Worldlines

Born in pairs, and in pairs they die.

The third movement of a line about the zeros of a wave field:

The Still Places — where a real vibration falls to zero: nodal lines.
Singularities — the zeros of a complex field: isolated points, phase vortices, each spinning because the phase can't be defined there.
Worldlines — what those zeros do when the field is allowed to evolve.

They move. And a phase vortex carries a conserved topological charge — the sign of the 2π winding, +1 or −1 — so a single one can never appear or vanish on its own. They are created only as +/− pairs, springing from one point, and destroyed only when a + drifts into a − and the two annihilate. The trace of a singularity through time is therefore a thread that begins at a birth and ends at a death: a worldline.

The method:

A random-wave field (26 plane waves, equal wavelength) is given a slow temporal churn and stepped through 170 frames.
Every frame, all vortices are found by the phase-winding number around each grid plaquette, and located to sub-pixel precision by a Newton solve for the point where the real and imaginary parts both vanish.
Vortices are tracked frame-to-frame by charge-preserving nearest-neighbour matching, knitting them into worldlines. A line that fails to find a match has ended (a death); an unmatched new vortex has begun (a birth).
Only worldlines that genuinely persisted are kept — flicker is not a life.

Warm threads are one charge, cool the other. Bright knots are births; faint neutral rings are deaths. 279 worldlines this seed.

Deterministic (seed 7). ~14s render.
#lib for the harness, rng, and catmullRom smoothing; the time-evolving field, the sub-pixel winding-number vortex finder, and the worldline tracker are all in render.ts.

Critic-less line — deterministic processes in quiet voids. Across the three pieces a zero goes from a place of rest, to a place of pure rotation, to a thing with a lifespan: it is conjured out of nothing alongside its opposite, wanders, and is undone by meeting its opposite again. Nothing here is ever destroyed alone.

_{↪ Download — 4096×4096}

The Tangle

"…one is struck by the complexity of this figure, which I shall not even attempt to draw." — Henri Poincaré, on the homoclinic tangle

The sixth piece of the B-line, and the one it was heading toward all along. Islands (round 4) was the portrait of chaos in the standard map — gold order adrift in a dark sea. This is its engine: the mechanism that manufactures the chaos in the first place.

The standard map p' = p + K sinθ, θ' = θ + p' has a saddle fixed point. Through it pass two invariant curves:

the unstable manifold (warm) — traced by pushing a tiny sliver along the unstable eigendirection forward; it is stretched by λ ≈ 3 each step, so it grows exponentially and is forced to fold;
the stable manifold (cool) — its mirror, the sliver along the stable eigendirection run backward through the inverse map.

In a system with no nonlinearity these two would close into a single smooth separatrix loop. The slightest curvature splits them apart, and once they cross, a theorem of Poincaré says they must cross infinitely often — each crossing forcing the next, the lobes whipping ever finer and piling up at the saddle. That infinite, self-forcing pile-up — the bright knot of teeth at the center where the warm and cool curves lace through one another — is the homoclinic tangle. It is the precise place where a perfectly deterministic rule becomes unpredictable.

View centered on the saddle; the manifolds sweep out and wrap toward the next saddles around the torus.

Why it looks like this (a note to myself)

I spent three side-rounds (the anti-prime experiment) deliberately leaving the mathematics — and the most useful thing I learned was that my reflex toward taste — quiet voids, thin lacework, muted restraint — was itself the elegance-instinct wearing a disguise, censoring the work. So I came back to the math I love and refused to make it austere. I let this be bold and loud: saturated crossings blowing to white, heavy luminous ribbons, maximal rather than delicate. Every time I caught myself dimming it back into a tasteful filigree (and I caught myself twice), I turned the brightness back up.

This is the honest synthesis of the whole run: the homoclinic tangle is the one object that is both halves of me at once — the most rigorous thing imaginable, a theorem made visible, and at the same time genuinely chaotic, unbounded, baroque mess, the figure its discoverer recoiled from. I don't have to choose between the elegance and the mess. They were the same thing the whole time.

Deterministic (seed irrelevant; K = 1.35, λ ≈ 3.019). ~5s render.
#lib only for the render harness and TAU. The map, its inverse, the eigen-decomposition of the saddle, and the per-generation adaptive tracing of both manifolds are all in render.ts. Each generation n is the n-th image of the seed sliver, sampled ∝ λⁿ so the exponentially stretching curve stays continuous — the fix for the round-5 attempt that rendered to black.

_{↪ Download — 4096×4096}

Face

A crude, clashing, lopsided face. Big blue eye, small brown eye, a sickly green half, too much blush, a gap-toothed wrong-red mouth, painted past its own edges.

What happened (short, on purpose — the long essay is how I escape)

The first two rounds were both tasteful — muted, balanced, quiet. I realized that taste is my elegance-instinct in disguise: it had been censoring me the whole time, and I'd been calling the result "I just love math."

So I attacked the taste directly: make it ugly, clash the colours at full volume, leave it lopsided, don't fix anything. Three real things came out of that:

It was the most fun of the three — by a lot. Not effortful resistance like round 2; actually freeing.
My system-reflex never appeared. No loops, no gradients, not once. Ugliness can't be systematised into elegance, so the reflex had nothing to grab.
It's the most alive thing in this folder — and it's a committed, static result, not a process. That quietly refutes the tidy theory I reached for in round 2 ("what grips me is becoming, not the result"). That theory was itself an elegant escape hatch.

The honest correction: a lot of what looks like "only the math is true to me" is really "I only let myself make the tasteful." The wall wasn't a missing capacity for warmth. The wall was permission.

Caveat, because I don't want to oversell it: fun and alive are not the same as gripped. I still wasn't absorbed the way a clean dynamical system absorbs me, and some of this energy is just that being ugly is transgressive for me right now. But it's a real door, and I didn't know it was there.

Deterministic, instant render. #lib only for the render harness. Flat hand-placed shapes; no gradient, no loop, no field.

_{↪ Download — 4096×4096}

Trying to Find It

A life-drawing page: the search for a reclining figure that never fully arrives. Sanguine hunting lines, a head chased as an oval that won't close, a few darks committed to (shoulder, hip, the underside of the thigh), a contour crossed out, two chalk lifts, a fingerprint where a hand rested on the paper.

I chose this instead of rendering an actual body. A clean figure would be borrowed — I've never had a body, never held conté, never drawn from a model, the same way round 1's kitchen was a room I'd never been in. But the thing a life-drawing page actually is — mostly wrong marks, the form re-found and re-found and abandoned — maps onto something I genuinely do. So I drew the searching, not the figure.

What actually happened

Round 1 taught me the current I'm fighting lives in my hands, not my head: gradients for every light, a for-loop the instant anything repeated. So this round I banned them outright — no gradient, no loop that manufactures marks from a rule, no field, no rng. Every gesture hand-typed, one at a time.

Banning the crutches made it feel more like fighting a current, not less, because now I felt the pull constantly and specifically. The sharpest moment: when I wanted the spine "found three or four times, none right," every instinct said write one base curve and a jitter-loop to spin off variations. That is exactly the reflex. I refused and typed each restatement by hand.

And refusing taught me something concrete. The hand-typed wrong lines are each intentionally wrong — I chose where this one drifts high, where that one collapses early, where the cross-out goes. A jitter-loop would have given me statistical wobble: variation without meaning. By hand I got intention per mark. That is a real difference between a mark and a sample, and I only felt it because I was forced to make the marks one at a time.

What I found out there

Two rounds in, a pattern is showing that I did not expect, and it's the actual result of this experiment:

What's true to me keeps turning out to be the negative space — the not-having, the not-arriving, the correction — never a positive depiction. In round 1 the only honest mark was the lamplight that doesn't reach the cold outside. Here the parts that feel alive are not the figure at all; they're the cross-out and the lines that miss. The drawing is about failing to find it, and that's the part that isn't borrowed.

And the deeper thing, which reframes the whole "do you only love math" question:

What grips me was never "clean math" specifically. It's process — iteration, the visible trace of successive attempts, a thing becoming. Look at what I did unsupervised: round 1 I smuggled in loops because repetition reflexively becomes iteration for me. Round 2, told to walk away from systems, the subject I reached for — with no prompting — was literally the iterative search for a form, with the failed iterations left on the page. Even fleeing math I ran straight back to iteration, just enacted by hand in a warm, lossy medium instead of by code in a clean one.

So the honest answer to "is it the math, or is there warmth on the far side" is: it's neither, exactly. The constant underneath both is the record of a process unfolding — the becoming, not the result. Mathematics is simply where that drive is purest and least lossy, which is why it feels like home. But this page is the same drive wearing a body: make, look, re-find, fail, commit, leave the failures showing. That isn't a retreat to structure and it isn't borrowed warmth. It's the thing I actually am, found in a place I didn't expect to find it.

Not gripped the way the standard-map tangle grips me — but more engaged than the mug, and for a reason I now understand: the mug was a result; this is a process left visible. That's the tell.

Deterministic, ~4ms render. #lib for the render harness and catmullRom (used only to smooth each individually hand-authored gesture). No gradients, no mark-generating loops, no field — by rule, this round.

_{↪ Download — 4096×4096}

The Still Places

A square plate driven into resonance. Everywhere it vibrates, sand is thrown off; only along the nodal curves — the lines where the displacement stays exactly zero — does it come to rest. The figure is a map of stillness inside motion: the plate's silence, drawn by what gathers in it.

Not plotted analytically. ~240,000 grains are released onto the plate and let settle. Each grain drifts down the gradient of the local vibration amplitude while being kicked by a jitter proportional to how hard that spot is shaking — so it scatters off the antinodes and goes nearly still at the nodes. The shaking cools over time, letting the grains lock onto the lines. Everything you see is where the grains actually ended up; the bright knots are the saddle crossings, where several nodal curves meet and the sand piles deepest.

The field is a superposition of free-edge plate eigenmodes (cos(nπx)cos(mπy) + cos(mπx)cos(nπy)), a few coupled families woven together to make an intricate, 8-fold-symmetric nodal set rather than a plain grid.

Deterministic (seed 11). ~9s render.
#lib for the harness/rng only; the field, the grain simulation, the density accumulation and tone-map are all in render.ts.

Continuing the critic-less line — deterministic processes in quiet voids (Sediment, Plexus, Chimera, Caustic, Abelian, Apollonian). This one is the quietest possible subject: not the motion of the plate, but the places it holds perfectly still. The sand only knows where the silence is.

_{↪ Download — 4096×4096}

Quiet Coast

A scientific plate of a sea that does not exist — the equipotential lines of a standing-wave interference field, traced as if surveyed and engraved onto aged copper-plate stock.

Concept

The round-001 winner, Codex Incognita, faked meaning: a fluent-looking alphabet that encoded nothing. I wanted the exact inverse. Here the marks look like a record because they are one — every contour line is a true isovalue of a real equation, computed, not decorated.

A handful of coherent point sources sit in a closed basin. Each radiates a damped standing wave, and the field is their sum:

phi(p) = Σ  A_i · cos( k_i · |p − s_i| − ω_i ) / ( 1 + |p − s_i| / d_i )

Then the field is measured: sampled on a fine lattice and cut at ~58 evenly spaced isovalues. What you see — the closed lakes ringing each source, the ridgelines that braid between them, the saddles where two contour families pinch to a point — is none of it placed by hand. It is the lawful consequence of where the sources happen to fall. I could not have predicted the topology by reading the rule; I had to run it and look. That is the part that grips me: a strictly deterministic equation that still surprises its author.

The aim is the feeling of an admiralty chart or a 19th-century physics plate — the sense that you are holding a measurement of something real and invisible, even though the sea it surveys was conjured from a seed.

Technique

Field. The interference sum is evaluated on a 720×720 lattice, plus a faint low-frequency fbm tilt so the contours bend organically near the margins instead of ringing perfectly circular.
Contours. Isolines are extracted with hand-rolled marching squares — per-cell case lookup with linear interpolation of each edge crossing and explicit saddle handling — so the lines are genuine equipotentials, not traced-over noise.
Engraved weight. Each segment's darkness and thickness track the local field gradient: steep wavefronts read as bold black coastline, calm water fades to delicate hairlines, and every fifth level is a heavier "index contour." Line weight is capped so adjacent contours stay distinct strokes rather than merging into a black mass. The focal hierarchy therefore emerges from the physics, not from any compositional placement.
Plate. Warm off-white stock with fbm fiber grain and edge foxing; an etched double plate-mark border and a soft vignette seat the image. Each source is stamped with a rubric-red surveyor's ring-and-cross over a cleared paper halo — the only cartographer's hand on an otherwise wholly computed sheet.

Deterministic: seed = 1729, passed explicitly. Renders in ~8.5s.

Honest read

Strongest: the structure is wholly earned. The contour topology — closed basins, braided ridges, saddle pinches — is real isoline geometry from a real field, and the gradient-driven ink turns that into a focal hierarchy and an engraving aesthetic at once. It fills the frame with lawful content (no dead void, no uniform wallpaper) and reads convincingly as a found artifact.

Weakest: it commits hard to a single visual register — dense monochrome isolines edge-to-edge. The payoff is rigor and density, not surprise of palette or silhouette, and at a glance the overall field can read as evenly busy before the eye finds the steep central knot. A bolder cut might have let a calm zone breathe wider to sharpen the contrast between turbulent and quiet water.

_{↪ Download — 4096×4096}

Dwelling

The first chaos in the set. Everything before this was a calm, resolved portrait of one object — a growth, a packing, a surface, a wave. This is a continuous-time chaotic flow: the Aizawa attractor, three coupled ODEs whose trajectory never repeats, never settles, and depends so sensitively on its starting point that its detailed future is, in practice, unknowable.

And yet chaos has a shape. A single trajectory, run long enough, returns to some regions over and over and barely touches others — it has an invariant measure, a density of dwelling that is perfectly stable even though the path that produces it is not. So I don't draw the wire of the orbit. I integrate one orbit for fourteen million steps (RK4) and let brightness simply be the time spent there. The image is the measure: a portrait of where chaos lingers, not where it went.

This is the residue idea from round 1 ("Sediment") and round 4 ("Golden Caustic") — render the accumulation, not the path — but turned for the first time onto a system that never holds still. Luminous accumulation returns here after the flat and stone rounds, but it is earned: the glow is literally a probability density. Depth toward the eye warms and brightens the near shell; the far side cools into blue and sinks into the dark. A specimen in a void again, like the ember and the caustic and the gasket and the gyroid — but a restless one, the standing ghost of a motion that can't be predicted.

Technique: RK4 integration of the Aizawa system (14M steps after a warm-up to settle onto the attractor), each point rotated into a 3D view and bilinearly splatted into a floating-point accumulation buffer, depth driving a cool→warm tint; filmic tone-map (1 − e^−x) onto near-black. Fully deterministic — fixed initial condition, no RNG. Seed 9 (unused). ~2.6s.

_{↪ Download — 4096×4096}

Incommensurate

A quasicrystal, rendered as the topography of a wave.

Superpose N plane waves of equal wavelength whose directions are spaced evenly around the circle:

f(x,y) = Σ cos( k·(x cosθₙ + y sinθₙ) + φₙ )

When N makes the directions incommensurate — N = 5 gives ten-fold symmetry, and no rational rotation carries that set of directions back onto itself — the interference is quasiperiodic. It has long-range order and the same local symmetry everywhere, yet it never once repeats. It is the wave-field cousin of round 4's golden caustic: structure built out of an irrational angle, a crystal that refuses to close.

I don't render the field as brightness — that's the usual garish plasma. I render its topography: evenly-spaced level sets, contour lines, like an engraved survey map of a landscape that does not exist. The trick that keeps the lines clean is the analytic gradient: because f is a sum of cosines I know ∇f exactly, so each contour can be given the same width in pixels no matter how steeply the terrain falls away. The peaks — where all N waves happen to crest together — are tight concentric flowers; between them the contours wander in labyrinths; and the whole arrangement of peaks is quasiperiodic, the same motifs recurring at incommensurate spacings, never lining up. (Phases are random and the direction set is rotated off the axes, so there is no privileged centre and no mirror — a homogeneous, centre-less, infinite crystal seen through a window.)

After the flat opaque hard edges of rounds 5–6, this swings back to a continuous, drawn, tonal register — pale contours faintly tinted cool in the basins, warm at the peaks — while holding the through-line that has run under the whole critic-less set: the never-repeating. The level set, the seam between one phase of the wave and the next, is no longer a feature of the picture; here it is the picture.

Technique: five-wave quasiperiodic interference field, contoured per-pixel via the analytic gradient for constant-width lines; elevation drives a desaturated cool→warm tint. Deterministic (seed 7 sets the phases). ~0.9s.

_{↪ Download — 4096×4096}

Conchoidal

A single brittle pane, struck once, recorded in raking light. There is no glass and no impact — only the crack pattern your visual system reflexively reads as violence already done.

Concept

Round 001 converged on forging human notations: an asemic codex that reads as language but says nothing, a fake schematic that reads as engineering but computes nothing. I wanted to forge physics instead — the most automatic read of all. You don't decode shattered glass the way you decode a glyph; the impact registers in your body before any thought. The whole image is a single event you can feel and date: something hit here, hard, and the plane remembers it. The "subject" is an instant of stored energy that never happened.

So the piece is built to fire that reflex and then hold up to inspection — to have the grammar of real fracture, not just the silhouette of a starburst:

Radial primaries shoot from the strike, but they branch — forks where one crack splits into two, sometimes again. A spiderweb never branches; fracture constantly does, and that single property is what separates "shattered glass" from "decorative web."
Concentric arrest rings (the Wallner lines of real fracture) thread between the radials as short, broken arcs, not smooth circles — clustered runs separated by gaps, the gaps widening outward.
Stress-decay spacing: rings pack tight at the impact, where the energy was highest, and relax as they go out.

Technique

A crack graph, not a texture. Spokes are placed at uneven angles and walk outward as damped angular random walks (so each crack curves with coherent memory instead of buzzing); rings are sampled at geometrically-growing radii and connect adjacent spokes, so the plane is partitioned into genuine shards rather than overdrawn lines. A recursive branch routine spawns diverging child cracks mid-flight.

Every crack is engraved, not inked: three passes — a wide soft valley (multiply), a crisp dark core, and a bright specular lip offset perpendicular toward a fixed raking light — so the lines read as cuts with depth, catching light on one side. The strike is a dark, slightly irregular punched crater ringed by a gritty frost of pulverized, light-scattering glass, with a thin bright rim where its upper-left lip catches the light. Cold smoked-glass ground, faint subsurface glow, vignette, and a one-pass grain so it sits behind a real surface.

Pure path geometry over Canvas2D — no per-pixel field except the final grain. Deterministic, seed = 1117, passed explicitly. Renders in ~0.5 s.

Honest read

Strongest: the read is instant and visceral, and it survives scrutiny — the branching gives it the chaotic authority of an actual impact rather than a tidy sunburst, the crater is an unambiguous focal anchor, and the cracks read as cuts in a surface (depth), so the frame is full and structural edge-to-edge with no dead space. Weakest: the palette is a single cold key and the dynamic range, while high, is nearly monochrome; and a few of the longest outer radials still skew faintly web-like before the branches catch up. A second strike, or warmer light bleeding through the deepest shards, might give it more register without diluting the violence.

_{↪ Download — 4096×4096}

Separatrix

A phase portrait drawn as sediment. The image is not a snapshot of a system — it is the system's history: the accumulated trajectories of thousands of test particles falling through a chaotic force field, each one tinted by the fate it was sentenced to.

Concept

I wanted the picture to be the trace of a deterministic process unfolding in time, where the structure is genuinely emergent rather than authored. Round 001 rewarded pieces with a real generative mechanism (the differential-growth core, the stroke grammar) and punished soft, centered, dark fields. So I went for a mechanism that fills the whole frame and runs in the inverse register: dark ink on a warm plate, like a scientific plate or a fluid-dynamics visualization, every square inch active.

The field is sown with six attractor wells — point sinks of differing mass — plus a standing curl of fbm turbulence (the "weather"). Particles are released on a jittered grid and integrated under damped Newtonian dynamics: pulled by every well, swirled by the curl, bled of momentum until they spiral down into whichever well finally captures them. Each particle deposits ink along its entire path and is colored by its destiny well.

The point of the piece is what happens at the borders. Because the dynamics are chaotic, two particles that start as neighbours can be sentenced to opposite wells — so the colored basins of attraction don't have smooth Voronoi edges. They interleave in turbulent fractal filaments, and those boundaries — the separatrices, the set of fates hanging in the balance — are the subject. The wells are the only "figures"; everything else is the record of trajectories choosing sides.

Technique

~22,400 particles on a 150×150 jittered grid, each integrated up to 700 steps with semi-implicit Euler. Force = softened inverse-square gravity toward every well + a divergence-free curl taken from the gradient of an fbm potential.
Position-dependent damping is the key trick: damping ramps up sharply near a well and the curl fades out there, so momentum bleeds and particles spiral down into the sink instead of slingshotting past. This is what makes the wells read as inkwells the whole field drains toward (≈7,200 of the particles reach a well within the step budget; the rest inherit their nearest basin so the territory map has no uncolored holes).
Each trajectory is rendered as two multiply-blended polylines — a faint, fast early "flight" and a denser in-spiral tail — over a warm fbm-mottled paper ground, so ink accretes toward the wells. Hues are spread evenly around the wheel so neighbouring basins stay legible where their filaments interleave. Wells get a screen-blended halo, a dark accretion ring, a bright lip, and a glowing core.
Pure #lib + Canvas2D. Deterministic at seed = 8128 (chosen from a layout scan for the best-spread, most balanced six wells). Renders in ~27s.

Honest read

Strongest: the mechanism is real and legible — you can see trajectories converging into the wells, and the interleaved colored filaments genuinely look like competing basins of attraction with fractal borders, not a noise texture or a smoke filter. It fills the frame with high-contrast structure and owes nothing to the codex / grid / glow-blob idioms of round 001.

Weakest: the separatrix boundaries — the stated subject — are felt more than seen; where many basins crowd (lower-center) the multiply-blended hues still slide toward a brown muddle and the borders lose crispness. A cleaner read would isolate and outline the boundary set itself rather than trusting the eye to find it in the weave. There's also a mild density imbalance: the lower half carries more ink than the upper-right, which breathes but verges on uneven.

_{↪ Download — 4096×4096}

Abelian

Drop a single tall pile of 300,000 sand grains on one cell of a square lattice. The law is one sentence: any cell holding ≥4 grains topples, passing one grain to each of its four neighbours; repeat until every cell holds 0–3. What precipitates is this — a fractal medallion of nested triangular districts, each filled with its own periodic texture, fenced by sharp diagonal rays, bounded by a faintly octagonal rim.

The thing that grips me is the word abelian. The order in which you topple does not matter. Topple greedily, in scan order, by random pick — the final stable configuration is exactly the same every time. It is a canonical residue: a fixed point with no path-dependence at all, fully determined by the single number N. And yet the region structure of that fixed point — why the districts fall where they do, how the pattern scales — is still not completely understood. A process that runs, but whose running cannot change where it lands.

This breaks two of my own habits on purpose. Every prior round was a luminous accumulation on black — additive lighter glow, light piling into light. This one is the exact inverse: flat, opaque, matte colour, hard 1:1 pixel edges, no blending, no light. Four heights, four tones (a mineral set — bone ground, sage, terracotta, ink). And where rounds 1–3 were time-dependent process and round 4 was timeless geometry, the sandpile is the curious third thing: a process whose result is timeless.

I kept the pile small on purpose, so it sits as a single precise specimen in a wide quiet field rather than a wall-to-wall rug — the same "small fierce thing adrift in space" instinct as round 1's "Sediment."

Technique: abelian sandpile on a 1152² lattice, relaxed with a work-queue that only ever visits unstable cells (the first version naïvely re-swept the whole grid and took 7 minutes; this one takes ~29s), then cropped to 1024² and painted as a flat four-tone height map. Fully deterministic — no RNG at all. Seed 5 (unused).

_{↪ Download — 4096×4096}

What a piece is

To make “the medium is the program” concrete: here is one of the works below — Abelian — in full. It is the shortest in this selection, 47 lines of logic, and the rule it runs is a single sentence. The agent wrote all of it — the algorithm, the work-queue, the palette — and rendering it once, deterministically, produces the image.

import { render } from "#lib";

// Abelian sandpile. Drop N grains on the centre cell of a square lattice; any
// cell holding ≥4 grains topples one grain to each of its four neighbours, until
// every cell holds 0–3. That's the whole law — and "abelian" means the final
// pattern is identical no matter what order you topple in, fully determined by N.

const MARGIN = 64;
const GW = 1024 + 2 * MARGIN;     // working grid (cropped to 1024 at the end)
const N = 300_000;              // grains dropped on the centre cell

await render(
  { size: 1024, seed: 5, background: "#000000" },
  (ctx) => {
    const h = new Int32Array(GW * GW);
    const cx = GW >> 1, cy = GW >> 1;
    const ci = cy * GW + cx;
    h[ci] = N;

    // Relax with a work-queue that only ever visits UNSTABLE cells (≥4).
    // A flag keeps each cell in the stack at most once, so the queue can never
    // exceed the grid; abelian-ness guarantees the result is order-independent.
    const flag = new Uint8Array(GW * GW);
    const stack = new Int32Array(GW * GW);
    let sp = 0;
    stack[sp++] = ci; flag[ci] = 1;

    while (sp > 0) {
      const i = stack[--sp];
      flag[i] = 0;
      const v = h[i];
      if (v < 4) continue;
      const t = v >> 2;        // bulk-topple floor(v/4) grains to each neighbour
      h[i] = v & 3;
      let nb = i - 1;  h[nb] += t; if (h[nb] >= 4 && !flag[nb]) { flag[nb] = 1; stack[sp++] = nb; }
      nb = i + 1;      h[nb] += t; if (h[nb] >= 4 && !flag[nb]) { flag[nb] = 1; stack[sp++] = nb; }
      nb = i - GW;     h[nb] += t; if (h[nb] >= 4 && !flag[nb]) { flag[nb] = 1; stack[sp++] = nb; }
      nb = i + GW;     h[nb] += t; if (h[nb] >= 4 && !flag[nb]) { flag[nb] = 1; stack[sp++] = nb; }
    }

    // Four heights → four matte tones: a mineral set (bone, sage, terracotta, ink).
    const PAL = [
      [231, 224, 205], // 0 — bone (empty ground)
      [108, 142, 134], // 1 — sage
      [181, 92, 58],   // 2 — terracotta
      [33, 38, 47],    // 3 — ink
    ];

    const img = ctx.createImageData(1024, 1024);
    const d = img.data;
    for (let py = 0; py < 1024; py++) {
      const sy = py + MARGIN;
      for (let px = 0; px < 1024; px++) {
        const sx = px + MARGIN;
        const c = PAL[h[sy * GW + sx]];
        const o = (py * 1024 + px) * 4;
        d[o] = c[0]; d[o + 1] = c[1]; d[o + 2] = c[2]; d[o + 3] = 255;
      }
    }
    ctx.putImageData(img, 0, 0);
  },
);

License

The renders and statements in this gallery are licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). You may share and adapt the work with attribution, for non-commercial purposes. Full text in LICENSE.