Reference — The dynamic-prompt catalog & data-build pipeline

The {#name} generators in data/dynamic-prompts/ are where most of the original creative effort went: ~113 little plugins that each return a prompt fragment, composed through the DSL (prompt-dsl.md). This page is the catalog and authoring idiom — how the generators are built, how they compose, the v1→v2 story, and how the word-lists they pull from are generated from raw sources. For the sigil mechanics see prompt-dsl.md; for the engine see ../systems/core-engine.md.

As of 2.3.0 the v2 generators are sorted into category folders under data/dynamic-prompts/v2/ (scene / subject / fragment / style / prompt / user), v1/ stays frozen, and {#name} resolves by path suffix (shared resolveName) so every reference stays short and category-independent. The full parity design (sidecars, _force-prefix, the verification seam) is in dynamic-prompts-architecture.md.

The authoring idiom

A generator is export default function (settings, imageSettings, upscaleSettings) { … return string }, plus optional export const full / export const suggestion_exclude. The house style is probabilistic accretion: start from a base phrase, then append fragments under independent coin-flips, mixing literal text with nested DSL tokens. E.g. city.js:

let prompt = "city, streetview, {city}";
if (_.random(0.0, 1.0, true) < 0.5) prompt += ", {building-style}";
if (_.random(0.0, 1.0, true) < 0.5) prompt += ", cityscape";
prompt += ", {#nature}, {#weather}, reflective street, wide shot";
return prompt;            // export const full = true

Because the output is re-fed through the pipeline (expansion/dynamic-prompt run twice, then list), a generator freely emits {#other-prompt}, {list}, and <expansion> tokens and lets the engine resolve them. This is the composition backbone: full scene prompts pull in partial helpers. (The sigil is brace-delimited {#name} as of 2.4.0 — see dynamic-prompts-architecture.md.)

Full vs partial (the classification that drives everything)

export const full = true marks a prompt that stands alone as a complete scene; its absence marks a partial fragment meant to garnish other prompts. promptFilesAndSuggestions.js reads this flag to split the catalog, and it drives both {#random}-style suggestions and the web Generate page's grouped token picker ("Full Dynamic Prompts" vs "Partial Dynamic Prompts"). suggestion_exclude keeps a valid prompt out of random suggestions.

• Full scene generators (≈full): landscape, city, castle, cave, mountains, space, ship, vehicle, ruins, room, school-room, store-interior, storefront, house, log-cabin, great-tree, great-bridge, beach, park, zoo, underwaterscape, knight, futuristic, the portrait-* family, retro-poster, vibrant-art, and the publicprompts.art set.
• Partial fragment helpers (no full): color, glow, neon, eerie, mystical, nature, weather, water, wildlife, fx, expressive, general-state, room-state, ice, lava, crystal, settlement, spaceship, portrait. These return , fragment strings and are the building blocks fulls compose from (and the garnish prePrompt() sprinkles into suggestions).

The proposed textual DPL for authoring these generators without JS — both assets/mockup/ revisions, the start/middle/end layering, and the open design questions — is analyzed in dpl-language.md.

The entity polymorphism

entity.js is a small type system: it picks one of {animal, danbooru character, colored flower, instrument, mythological creature, tree, person} and optionally adds emotion/hair/clothes for humanlike results, gated by an entity-class filter. Thin wrappers specialize it: animal.js ("animal"), person.js ("human"), living-entity.js ("living"), entity-name.js (name-only). The publicprompts.art templates embed {#entity} / {#person} / {#living-entity} / {#entity-name} as their "<name>" slot, so one template yields wildly different subjects each run.

The random prompt generators

These live in the v2/prompt/ category (force-prefixed, so they show/insert as {#prompt/…}).

v2/prompt/artists.js and v2/prompt/fx.js are the two prompts the pipeline auto-appends ({#artists}, {#fx}) when autoAddArtists / autoAddFx are on (resolved by suffix, so the force-prefix is display-only); v2/prompt/d.js composes danbooru anime tag streams ({d-general} + {d-character} + {d-meta}). (Renamed from the old engine/ folder in 2.5.0: danbooru→d, random→random-words, and *-prompt→*.)

v1 vs v2 — the decomposition story

data/dynamic-prompts/v1/ (33 frozen modules, addressed as {#name-v1}, always treated as full, and they force autoAddFx/autoAddArtists off because they bake those in) are the original monolithic generators. They inline private helpers like maybeAddColor(), multiColor(), and entityBasicKeywords() and hard-bake color/weather/time (several import the shared entityBasicKeywords from ../v2/subject/entity.js). The categorized v2/ set is the refactor: the same scenes re-expressed by extracting those into shared composable sub-prompts ({#color}, {#weather}, {#nature}, {#general-state}, {#room-state}, the entity family). v1 is kept verbatim so old looks stay reproducible — reading a v1/v2 pair side by side is the clearest window into the project's "compose from building blocks" design turn (Jan 2023; see ../context/history.md).

v2/user/ holds community contributions (addressed {#user-name}, always full) — currently beach-merk.js by Merk, which notably composes siblings as direct function imports (${city()} from ../scene/city.js, ${nature()} from ../fragment/nature.js) rather than {#city} tokens, an alternative to token-based composition.

Provenance worth preserving: the 3d-*, comic, sticker, funko-3d-print, fluffy-animal, needle-felt, silhouette, psychedelic, space-hologram, gold-pendant, lowpoly-3d-isometric, sports-logo, anime-irl templates are credited "taken from publicprompts.art and modified to be more dynamic" — the original swapped their fixed <name> subject for the {#entity} system. Keep these credits (also in list-credits.md).

The data-build pipeline (lists from raw sources)

The {list} word-lists in data/lists/*.txt are generated, not hand-maintained, by one-off scripts in data/ from large raw sources (kept under version control as the inputs):

These run manually (they process.chdir to the repo root and write into settings.listFiles); they are not part of normal generation. The raw sources and their licenses are credited in list-credits.md. The hand-authored lists (cities, building styles, rooms, clothes, etc.) live alongside the generated ones in data/lists/.

Expansions (<name>) the generators rely on

data/expansions/ holds the text macros generators splice in (now nested into category folders, referenced by bare name via path-suffix resolution): dap (deviantart, art station, pixiv), detail/legacy (masterpiece, highres, … HDR; detail/ is force-prefixed), rays (god ray, light shaft, volumetric lighting), detail/legacy-person, pixelart, candlelight, coffecup, flower-pic (itself {flower}, {flower}, {artist}), and underwater-anime-irl (which nests {#anime-irl}) — proof that expansions can themselves contain lists and dynamic prompts. See expansions-architecture.md.