A collection of thousands of distinct images is not drawn one by one. It is generated from a set of layered components combined under rules, and the rules do more work than the artwork.
Layers are combined in a fixed order
An artist produces variants for each visual slot: background, body, clothing, headwear, and so on. Each slot holds several alternatives.
The generator picks one option per slot and stacks them in a defined order, so items in front are drawn over items behind.
The number of possible combinations grows multiplicatively, which is why a modest number of hand-drawn components yields tens of thousands of distinct outputs.
Weights create rarity deliberately
Options are assigned selection probabilities rather than being chosen uniformly. A trait weighted low appears in few items and becomes the scarce one.
Rarity is therefore designed in advance, not discovered afterwards. The artist decides what will be scarce before a single item exists.
Because selection is random within those weights, the realised counts differ slightly from the intended ones, which is why published rarity tables are computed after generation rather than before.
Constraints prevent nonsense combinations
Random layering produces conflicts: headwear that clips through a hairstyle, or two items that occupy the same space.
Generators handle this with incompatibility rules that exclude certain pairings, and with conditional layers that only appear alongside specific choices.
The rules also introduce bias, since excluding a pairing changes the effective frequency of both traits involved, and careless constraints distort the intended rarity distribution.
Deduplication and ordering matter
The generator must check that no two outputs are identical, which happens more often than intuition suggests when some traits are heavily weighted.
The order in which items are numbered is also significant, because collectors attach meaning to low identifiers, and a naive sequence can correlate number with rarity.
Shuffling before assignment, and committing to the shuffle in a verifiable way, is what separates a fair distribution from one where insiders knew which mints to take.
Reveal mechanics protect the distribution
If metadata is published before minting completes, buyers can inspect which identifiers hold rare traits and target them.
Collections therefore withhold metadata until minting closes, then reveal it all at once, often using an on-chain random value to fix the mapping.
Without that commitment, the party controlling the metadata could assign rare traits after seeing who minted what, which is a failure that has occurred often enough to make the safeguard standard.