The gap between an image on a screen and an object on a wall is where most generative art stops. Screens are free and forgiving. Board costs money and does exactly what you told it to, including the parts you got wrong.
This is the whole pipeline, in order, including the steps that don't work.
Writing the system
A generator is a program with two halves. There's a seed, which is a single value that determines every random choice the program will make, and there's a set of parameters with defined ranges: layer count, pattern density, palette, aspect ratio, how far a curve can stray from its neighbor.
Those two halves describe a space. Not a picture, a space of possible pictures, and designing that space is the actual authorship. I'm deciding what can appear and how far it can go, which is closer to writing rules for a game than drawing.
The parameters are exposed in the URL, so a piece can be generated with specific values. Complexity, colorfulness, palette, and size all arrive as query parameters, which means the same code produces a 16 by 16 square or a 30 by 12 panoramic panel depending on what it's told.
Here's the constraint that shapes everything: the output has to be vector. Real paths with coordinates, closed shapes, geometry that can be separated into planes. That rules out a large amount of what looks good on a screen, and it's the reason the work looks the way it does.
Running it, and mostly throwing things away
I run the generator and look at what comes out. Then I run it again.
Dozens of outputs for every one that survives. The ratio is not something I'm proud of or embarrassed by, it's just what the process is. Most of what appears is competent and inert: correctly following the rules, arriving nowhere.
Then occasionally something holds. I can't give you a rule for what that is. It's closer to recognition than judgment, and by the time I notice I've made the decision, the decision is already made. I've tried to describe the moment several times and the description always comes out worse than the experience.
What I can say is that the physical version is always in mind while I'm looking. An output that would make a good print can still be wrong here, because I'm looking at it as a stack of planes and asking whether the depth will do anything.
Decomposition
This is the step nobody sees and where the real translation happens.
A finished composition gets broken into layers. Each layer is a separate set of closed paths that will become one physical plane of mat-board, and the stack of them assembles into the relief. Deciding which elements belong on which layer is a design decision with consequences: it determines where the shadows fall, how deep the piece reads, and whether the whole thing holds together structurally once it's cut.
The count runs from about twelve layers to twenty or more depending on the piece. More layers means more depth and more places for error to accumulate.
Then there's the kerf. A laser removes material as it cuts, a fraction of a millimeter of board that turns into smoke and disappears. Every path has to account for it, because a shape cut without compensation comes out slightly smaller than the shape you drew, and twenty layers of slightly is a visible problem.
Cutting
The machine is a 100 watt CO2 laser. It sounds like a sewing machine that moves at twenty inches per second, and it fills the room with a smell somewhere between singed wood and a snuffed candle. I've written about what that room is actually like at more length.
Each layer gets its own file and its own run. Minutes per layer, depending on complexity.
The pieces come off the honeycomb bed warm, with edges that are smooth and very slightly compressed. The material that falls away, the interiors of every cutout, piles up underneath. Some of it is beautiful on its own and I keep a jar of it for no reason I can defend.
Assembly, where the machine stops helping
Twenty cut layers are not a sculpture. They're twenty pieces of board in the right shapes and no particular order.
Assembly is PVA glue, the archival adhesive used in bookbinding, applied with a fine tip. Too much and the board warps overnight. Too little and the bond fails inside a week. The right amount is something I learned by feel over years and cannot express as a volume.
Each layer gets placed, pressed flat, and checked for registration against the one below it. Alignment has to hold within about a hair's width, and error accumulates: a layer that's off by a small amount early puts every layer above it off by that amount plus its own. There's no correcting it later. There's only cutting a new one.
The finished stack goes under weight overnight. In the morning it's a single object.
What fails
Three things, roughly in order of how often.
The composition doesn't survive translation. It looked right on screen and the depth does nothing, or the layer separation reads as noise instead of structure. Usually I catch this at decomposition. Sometimes I don't, and I find out after cutting.
Registration drifts. Recoverable early, fatal late.
Adhesive misjudged. A warped layer is not repairable, so the piece becomes material for figuring out what went wrong.
None of that is unusual and none of it is interesting except that it's real. A process with no failure rate is a process where nothing is being decided.
Why the pipeline is the point
Somebody who buys one of these is buying the survivor of a system that could have produced an unimaginable number of alternatives, chosen by a person and then made by hand. The code did the generating. It has no idea what it made, and it did none of the cutting, stacking, or judging.
If you want the shorter version of why this is a different thing from typing a description into a model, I've written that separately. If you want to know what the result does on a wall once light hits it, that's here.
Shawn Kemp makes one-of-a-kind generative artworks fabricated as dimensional layered relief sculptures in Bellingham, Washington. The current work is here.