how the art is made

the art is made with ai. this page says exactly how, so nobody has to guess.

we are not claiming a hand drawn collection. we are claiming that the process is stated and the same for every unit.

image model: gpt-image-2, called through the codex cli image tool. no other generator is used for collection art.

assembly and post: sharp, in node. no external service touches the files.

every direction has one prompt file. it is fixed before generation starts and not edited between units.

each prompt carries a block of invariants that must not change: frame, camera, lighting, background, material, and a strict frontal view. it also carries a negative list of what must never appear.

sku names are metadata. they are never written into a prompt — a label like "bear" in a prompt produces a literal bear, which is not what a trading floor archetype means.

the base is generated once and never regenerated. traits are separate layers composited onto it.

this is the whole point: pixels that were not regenerated cannot drift. generating each unit whole gives a different figure every time, no matter how precise the prompt.

base drift is measured per pixel against the base, not by eye and not by bounding box. a bounding box moves when a hat is taller — that is a trait, not drift.

the measuring tool is calibrated against synthetic targets with a known offset before it is trusted. an uncalibrated ruler reported a one pixel shift as four.

target for base drift is zero, not "close enough".

choosing the direction. rejecting units. retouching defects. deciding trait weights and which combinations are forbidden.

every unit is looked at before it ships. a tool reporting success is not the same as a file that is right.

this page describes the pipeline as it stands today, not a plan. when something changes, this page changes with it.

direction not yet locked. supply not yet fixed. mint date not set.