Art and audio direction

How Halvesper looks and sounds, and how a tiny team produces it with a fully AI-generated pipeline. Builds on 00-canon.md (HD hand-painted-look sprites, layered equipment, titans at nightfall, light against the dark, AI-generated art and audio). Supersedes the art, UI and audio sections of the retired docs/design.md (deleted from git), which were written for pixel art.

Names come from 01-world.md and 02-atlas.md; controls, hotbar counts and touch sizes come from 03-systems.md; this document draws them. 09-glossary.md lists who owns what.

What went wrong before, and what the mood tests show

docs/design.md produced a 16 px pixel-art frame: square tile terrain, lollipop trees, a flat purple gradient sky and two tiny people. Foster’s verdict was “a cheap little Terraria game, nothing unique original or cool”. The reasons, so we do not repeat them:

1. The grid was visible. Every surface was a 16 px tile. Nothing broke the grid, so the world read as a voxel sandbox.
2. Nothing was big. The largest thing on screen was a house. No scale, no awe.
3. The theme was a color, not an image. “Dusk” was a purple sky gradient. Nothing in the frame was only possible in this world.
4. Generic everything. Stock tree, stock brick, stock dirt. Any asset pack.

The two AI mood sheets (concepts/sheet-dusk.jpg, sheet-beast.jpg, quick renders, not targets) are useful mainly as a list of corrections:

Key-art references for the chosen direction

The renders in img/ are AI references for mood, composition and color, not final art and not a style to copy stroke for stroke. They are much closer to the target than the mood sheets. What they share and get right: the half-sunk orange sun on the horizon, warm lantern light against a violet-blue dusk, tiny figures for scale, and holds built onto living backs. What they share and get wrong: a smooth airbrushed CG finish instead of visible painted brushwork, and titans that are still too small (barn- or hill-sized, not mountain-sized).

Mossback. Right: moss on the shell, warm windows, a ladder up to the hold, people the size of a toe at its feet, the sun on the horizon. Wrong: the houses are a third of the shell’s height, so the tortoise is about four stories tall; at titan scale the hold is a whole town and a house is a crumb on the shell ridge. The hold is a cluster, not a town. The face is good: ancient and plain, not a mascot.

Greyhorn. Right: the hold strapped on with saddle straps, buckles and bags (signature 5 in one image), forges glowing inside the buildings, canyon framing, ground lanterns. Wrong: horse-to-house scale again; no heartfire glow through the beast’s own body (its ribs or chest should burn faintly); the fur reads as a dark mass with no form.

Ystra. Right: the best scale of the set (the scholars’ city is small on the hide), the cloud sea, white towers that suit mages, glyph markings on the hide. Wrong: the glyphs glow cold white-blue; if they are heartfire seams they must glow warm (cold light is the Lampghosts’). The finish is the most CG of the set: a 3D render, not a painting.

The Moth Queen. Fails (redo is decision D10 in 07-decisions.md). Her hold, Lampblack, is woven into her fur, and this bazaar sits on the ground under her. She is also tent-sized, a big bug, with no hold and no awe. Keep the cream wing glow, the strings of paper lanterns and the night palette for how she looks until level 68: Lampblack dusts her wings each Glow with warm wing-scale powder so they read cream from a distance. Redo: silk houses, rope stairs and hammocks woven into the fur of her thorax, lantern strings along her wing struts, the whole hold tilting with each wingbeat, seen from far enough away that she spans the sky. A second sheet shows her undusted, wings shining cold coldlight: the level-68 reveal (01-world.md §1.5).

The fallen titan. Right: the body as terrain (ribs you could walk through), dying ember light inside the ribcage, Lampghosts crawling over it, ruins at the edges, one lantern for scale. Wrong: it is a humanoid skeleton. The fallen titan is Wend, a beast built like a heron crossed with a millipede, lying on its side with its eye open (01-world.md §4.5); regenerate it (decision D9). A beast’s skull, many jointed legs and rib-arches read as awe without grimness.

The four classes. Right: chibi proportions, every class carries or stands by a lantern, silhouettes read apart (helmet and shield, flowing hair and flame, hood and bow, cloak and hood). Wrong: the Shade’s dark face with pale eyes reads as a Lampghost and a borrowed look; the Shade wears a hood and mask with real colour instead (proposed: a deep plum or slate-teal hood with a stitched mask, the face visible above it); thick uniform black outlines and flat cel shading, the mobile-gacha look we are avoiding; front-facing poses (gameplay is three-quarter side view); the Warden’s weapon is a lantern staff, not a lantern-halberd; lanterns hang from nothing.

The Lampghosts. Right: familiar creature shapes (a wolf, a ghost, a knight, a slime), eerie and cute at once, which fits pillar 3 better than sharp, menacing shapes. Wrong: the black, empty bodies with big glowing round eyes read as another game’s shadow creatures, not ours (Foster, 2026-10-04); the eyes glow pale yellow-green; the purple slime is a normal purple monster, not a Lampghost. Keep the shapes, drop the black bodies and orb eyes: Lampghosts are coloured coldlight with small faces and a stolen ember inside (signature 4).

A gameplay map. The closest to a real screen. Right: a titan silhouette walking in the far background, sagging wooden platforms with rope rails, ladders, lantern posts at platform ends lighting the walkable edges, foreground silhouettes at the screen corners. Wrong: the titan is too near and too small (it should span half the sky, fainter); the whole scene is a teal-green forest with no sun and no dusk, and teal is too close to the Lampghosts’ cold light; the purple-pink magic orb reads as a Lampghost; the characters are small flat cel figures with thick outlines; the hotbar is a generic frame.

Visual identity: five signatures

A screenshot is recognizable as Halvesper if it shows at least two of these. Every map must show at least two; every marketing shot shows three.

1. A titan in every sky

Every outdoor map has a titan somewhere in its parallax: the one carrying you, the one walking past on the horizon, or the dead one lying across the far hills. They are never shown whole and never in the gameplay layer. A titan’s leg fills a screen; its head is a hill. Scale comes from cropping, atmospheric fade and parallax speed (far layers at 0.03–0.08 of camera speed), and from tiny lit windows of a hold on its back. Each titan’s heartfire glows through seams in its body (between shell plates, along Greyhorn’s ribs, in Ystra’s gill slits, through the Moth Queen’s wing veins): the warmest light in any frame where it appears.

The titan in a hunting map’s background is in the place the server says it is. If Mossback is two regions east, it is a faint silhouette on the eastern horizon; if it is walking through, its feet come down in the mid layer and dust falls on the gameplay layer. The world map is never quite the same twice (pillar 1), and you can see that from any map.

2. The sun that will not set

One low sun, always on the western horizon (Lampward; it is how players navigate). Maps are laid out so west is screen-left or behind the scene. It is a big orange disc half-sunk into the horizon (as in every key-art reference), never higher. How much of it you see is the depth gauge of the world:

The disc is visible from any high or glare map (holds on titans, the Cloudshelf, the Sunfoot), and the sun does not move until Morrow’s birth (08-scale.md §1).

Every surface lit by the sun is lit from that side, low and raking: long shadows stretch away from it, rim light catches the sun-side edges. No noon light, no light from straight above, no blue daytime sky anywhere in the game.

3. Light pools are the composition

Below the holds, darkness is real. The scene is darkened by a per-map ambient level, and light sources carve readable areas out of it: player lanterns, heartfire braziers, glow fauna, Lampghost light. A party moving through a dark map is a moving bubble of warm light; a solo player’s radius is smaller. Map designers compose with light first: safe ledges are lit, ropes have a lantern hook at the top, danger sits at the edge of a pool. In holds the opposite holds: everything is warmly lit and the dark is only the sky.

Every Lanternbearer carries a lantern, always visible on the character (hip, hand, shield boss, staff tip or bow limb depending on class). It is the one piece of equipment every silhouette shares and it is a cosmetic slot.

4. The Lampghosts are made of stolen cold light

Everything in the world is painted in warm light, except the Lampghosts, which are made of the cold light they steal. A Lampghost creature is a coloured, semi-translucent body of coldlight (proposed): each kind has its own cool tint (sea-glass aqua, sea-green, frost blue, teal, violet, periwinkle; never black, never warm amber, which is the player’s fire), with soft inner gradients brighter at the core and edges that trail and fade into cold light, so the scene shows faintly through the thin parts. A thin lighter rim with a soft halo outside it draws the shape. Faces are real faces: small eyes and a mouth with expressions, so they read cute-eerie, never a blank void with glowing orbs. They “wear the shapes of what they consumed”: a Lampghost bat is the shape of a bat made of cold light. Eerie and cute at once, never gruesome.

What reads. This is a grind game: a Lampghost must read at a glance anywhere on a 1920x1080 screen, on dark ground and light. A flat black fill does not (it vanished on Mossy Field’s dusk ground); the tinted glow of the body, the rim and the halo carry the shape. Out of light (Shrouded) the body and halo dim, but the body stays dense enough to read and the rim stays strong: the dark changes how a Lampghost is hurt, not whether it can be seen.

The stolen ember (proposed). Every Lampghost carries a small warm spark of stolen fire visible inside its body: the one tell no other creature has (Cinderkin are made of fire; they do not hold a spark in a cold body). It is never a cut-out, never a teardrop or keyhole shape, and never centred in the chest as an emblem: it sits where the kind would hold something, the low belly, the throat or a tail tip (Puddling low belly, Mossnip and Thornpig belly, Dandlehop tail puff, Dimwing and Afternooner throat). It grows and brightens with how much fire the Lampghost holds and dims as you knock fire out of it: a Puddling’s is a firefly, the Herdmother’s a hearth, Wend’s Absence’s a buried sun. It stays clear of the face and visible in every pose (a point on the rig). At 48 px the ember is at least 3 px across.

Their light is cold (cyan-white to violet). Cold light means absence, not “enemy”: it also lights the Moth Queen’s true wings, the Stay’s caged farms, a Shade’s cracked lantern and Morrow’s veins. What marks a hostile Lampghost is the coloured coldlight body plus the stolen ember, not cyan alone. A screenshot with a small glowing creature of teal or violet light, a warm spark caught inside it, in a warm painted world is this game.

5. Holds that walk

Hold maps are built on a creature’s back and look it: buildings strapped and chained to shell plates, horn or blubber; rope bridges between moss ridges; the ground has texture that is clearly hide, shell or fur. The far background scrolls on its own, slowly, because the town is moving (2–6 ref px/s on the farthest layer, independent of the camera). Every 8–12 s a footfall: the whole scene bobs 2–3 ref px over 400 ms, hanging lanterns swing, dust sifts down, a low thud plays. Ystra’s hold bobs slower and rolls instead; the Moth Queen’s hold dips with each wingbeat.

Style rules

Shape language

Line

• Lines are a darker, more saturated step of the local color, never black. The darkest line value anywhere is #120F1C.
• Line weight decays with depth:

• The top edge of every walkable surface has the strongest line and a lighter lip. This is how a player knows what they can stand on.

Rendering

• Painted, not vector: two to three tonal shapes per form plus soft gradients and a visible gouache-like brush texture at 2x master resolution. Edges are hard on the light side of a form, soft on the shadow side.
• Characters and monsters are painted with a neutral top-front key light so they can be mirrored. Sun-side rim light is added at runtime by a rim shader (below), so it is always on the correct side after flipping.
• Environments are painted with their map’s sun side baked in.
• Texture detail falls off with distance: actors and the gameplay layer have the most; far layers are flat color masses with soft value steps.

Lighting model: painted light plus a light buffer, no normal maps

How it works: 1. A SubViewport at 1/4 of screen resolution draws soft additive light sprites (lantern pools, heartfire, spell glows, Lampghost glows) on top of the map’s ambient color. Each light is one textured quad, so 300 lights cost about what 300 particles cost. 2. The scene is multiplied by that buffer in a full-screen pass, then light sources themselves (flames, lantern glass, eyes) are drawn unlit on top with an additive bloom sprite. Only the UI is drawn after. 3. Per-map values: ambient color, ambient level (holds 0.9–1.0, first lands 0.7, middle 0.45, deep 0.25, boss arenas set by the boss), sun side, rim color. 4. A rim shader on actors samples their own alpha offset toward the sun side and adds a thin rim of the map’s rim color. In dark maps the rim comes from the nearest strong light instead (one direction per actor, updated at 10 Hz). 5. No shadow casting. Contact shadows are soft painted ellipses under actors. 6. A per-region color grade (a 3D LUT strip, 16³) is the last scene pass.

Light budget: unlimited soft lights in the buffer; at most 32 drawn per frame on mobile and web, prioritized as self, party, bosses and hazards, then nearest. Other players’ lanterns beyond that draw a small glow sprite only.

Value and color structure

• Two lights in the world. Warm heartfire (amber to white-gold) means people, safety and the player’s own power. Cold coldlight (cyan-white to violet) means absence. No Lampghost emits amber (the Cinder Eye is not a Lampghost). Friendly cold light is rare and always tied to absence in the lore: the Stay’s cage farms, the Moth Queen’s undusted wings, Morrow’s veins, and the Shade class, whose effects are void-black with a warm edge plus a thin cold crack, so they still read as the player’s. Decorative cold light with no lore reason (blue mushrooms, teal magic) is never used.
• Warm light, cool dark. Lit planes push warm and lighter (value 65–95%); shadows push cool and darker (value 15–40%), never neutral gray.
• Contrast belongs to the gameplay layer. Far layers sit within 20% of the sky’s value; mid layers within 35%. Actors and walkable edges carry the full value range.
• Saturation belongs to light and actors. Backgrounds are desaturated; light sources, player gear and monster accents are the most saturated things on screen.
• Grayscale test. Every map, desaturated, must still let a player find themselves, every monster, and every walkable edge within one second.

Global anchor palette

These do not change by region.

Region palettes (three key regions plus the holds)

Each region defines sky, sun, terrain, foliage, accent and ambient. Hex values are the mid step of each ramp; the palette tool extends each ramp with one lighter and two darker steps.

Lantern Town, on Mossback (starting hold)

Mossy Field (first hunting land, below Mossback’s route)

A deep land (Lowlight’s fens and the deep strata of late regions)

The other holds (ambient 0.9–1.0, each with its own key accent)

Wend’s Fall (the dead titan, region 9): bone #CFC2A8, ash #5A5560, cooling heartfire veins #FF6A3D fading to #7A2A30. The only place where warm light is dying and the Lampghosts have moved into it.

Never

• Pixel art, nearest filtering, or visible square tiles.
• Pure black (#000000) lines or shadows; pure white (#FFFFFF) anywhere except the core of a crit flash.
• Blue daytime sky, overhead sun, midday light, or a sun disc above the horizon.
• The smooth airbrushed CG finish of the reference renders; shipped art shows brushwork.
• Cold light with no absence behind it (blue mushrooms, teal magic, cold friendly glows in the world).
• Black or flat-void Lampghost bodies, big round glowing orb eyes as their identity, a keyhole or teardrop hole through any body, a Lampghost without its stolen ember, or a stolen ember on anything that is not a Lampghost.
• Gore, blood, dismemberment, realistic skulls with flesh. Monsters “pop” into light motes; Lampghosts unravel into specks.
• A whole titan in one frame, or a titan in the gameplay layer.
• Mascot faces on titans.
• Text baked into art (except signage in the world, drawn as glyph shapes without real words so nothing needs localizing).
• Tracing or prompting from MapleStory, Terraria, Hollow Knight, Ghibli or any named living artist.

Resolution and scale

One reference frame

All sizes in this document are ref px: pixels at a 1920×1080 reference screen at zoom 1. The sim keeps its own units; 1 sim unit = 4 ref px. A tile is 8 units (game/sim/tiles.gd SIZE := 8), so 32 ref px; the player’s 12×22 collision box is 48×88 ref px, and today’s 480×270-unit view becomes a 1920×1080 frame, so the sim does not change. Collision moves from the tile grid to footholds (04-architecture.md §5.2); the tile survives only as a measuring unit and in the text test maps.

A character is about 10% of screen height on desktop, close to MapleStory’s proportion, so a map shows enough to read platforms and other players.

Project settings

• Base viewport 1920×1080, stretch mode canvas_items, aspect expand, linear filtering with mipmaps on all art. project.godot is set this way since the look pass, the camera at zoom 4 (480×270 units), and the gate’s frames are 1920×1080.
• Horizontal view is capped at 2560 ref px (ultrawide players see extended backgrounds, not more of the gameplay layer).

Per platform

Phones see about the same width as desktop and three quarters of the height, so a character renders about 150 physical px tall on a 1080-tall phone screen. Zoom is not competitive in PvE; if PvP is added, its maps lock zoom 1.0 on every platform.

Pixel density

• Masters are kept at 2x (2 texels per ref px), downsampled from the generation size. 1x is a downsample (Lanczos) done by the export script, never by hand.
• Sprites are never scaled in-engine by more than ±10% except VFX, titans, and far backgrounds.
• Far background layers (parallax < 0.3) are stored at 0.5x and upscaled; they are soft by design, and this cuts their memory by 4.
• Titans are stored at 0.5x (far) or 1x (near pieces).

Texture memory budget

Characters

Proportions

• 2.4 heads tall: head 48 ref px tall (with hair), body and legs 64. Feet sit on the collision floor; the head overhangs the collision box by about 24 px.
• Three-quarter front view facing right; facing left is a horizontal flip. Asymmetric gear is fine (it flips with the body); text and emblems are symmetric.
• Big eyes in the upper-middle of the face, small mouth, no nose line. Expressions do most of the work, so faces get their own swap set.
• One body base for everyone. Body type is a choice of two silhouettes that share the same rig and the same gear pieces (gear parts are cut to fit both, with an optional variant only for tops). Skin tone is a gradient-map dye with 12 presets and a free picker.

Skeletal cutout, not frame-by-frame

With 300+ gear pieces at launch and cosmetics forever after, frame-by-frame paper dolls would need tens of thousands of frames that all line up, which no image model can hold consistent. Cutout is the only option a tiny team can sustain, and it is the only one an AI pipeline can align automatically (see Characters: the hard part). We pay for it in animation craft:

• Rigid parts on plain Node2D bones, animated with AnimationPlayer. No Skeleton2D/Polygon2D mesh skinning on characters (CPU-skinned on every platform, and 150 skinned characters per map is not viable on web). Capes, long hair, scarves and tails use a 3–5 segment verlet chain for secondary motion.
• Sprite swaps carry the drawing quality. Each limb has 2–3 painted angle variants (front arm: extended, bent, foreshortened toward camera); hands have 5 poses (open, fist, grip, point, cup for the lantern); faces swap per expression. Swaps happen on key poses so motion still has drawn “frames”.
• Pose-to-pose with holds. Attacks snap between strong key poses with 1–2 frame holds, sampled at 30 fps for combat actions (stepped, not smooth), 60 fps interpolated for locomotion. Squash and stretch through bone scale (±12% max).
• Smear and trail overlays. Weapon swings add a generated smear overlay (3–4 frames) that hides the rigid-part look at the fastest moment.
• Native Godot, not Spine. Spine’s tools are better, but its runtime needs a license per user and a runtime outside stock Godot. The canon fixes stock Godot, and creators will later author against the same rig, so the rig and its animations must live in our own data format. Revisit only if the style-target milestone shows AnimationPlayer authoring is too slow.

First cut (look pass). The starter character is a side-view cutout cut from one Z-Image Turbo painting, dressed with a Klein edit, plus an armless Klein edit that gives the torso clean shoulders (tools/art/player/, prompts and seeds in player.json). Fifteen parts plus lantern, sword and bow props share one 512×256 atlas at 2x master. The rig draws every part from one node (game/render/player_view.gd) with poses keyed in code (player_art.gd), not Node2D bones and AnimationPlayer, so the paper-doll order stays one flat list; move the keys into AnimationPlayer data when animations are hand-authored. There is no back view yet, so climbing uses the side view.

Animation timing is data: each skill in the sim has windup, active and recovery ticks; the attack animation is retimed to fit them, with the impact key pose on the first active tick. Animation never drives gameplay.

Paper-doll layers and draw order

Back to front, facing right. The back arm and back leg are the character’s left side (away from camera).

Climbing uses a back view of the same rig (head, torso, limbs from behind). Each hat, top, bottom, cape and hair includes back-view parts; shoes, gloves and faces do not need them (hidden or shared).

Hair rule: each hairstyle ships full, under_hat (cropped at the crown) and back_only variants; the hat’s rule picks one.

Combining gear without combinatorial art cost

1. Cosmetic override per slot. Every slot has an equip item and an optional cosmetic item; the cosmetic’s parts render, the equip item’s stats apply.
2. Dye regions instead of recolors. Gear is stored in value with up to three dye regions stored in the R, G and B channels of a mask. A gradient map per region gives tiers and dyes from one generation (e.g. a tier-3 helmet is the tier-1 helmet plus a new trim part and a different gradient). Target: one generated base yields 3–5 shipped items.
3. Runtime bake. When a character’s look changes, a one-off pass composites each body part with everything on it (arm + sleeve + glove, dyed) into a per-character atlas: about 12 parts in one texture. The rig then draws one texture per character, so characters batch, no per-part shader runs per frame, and dye is free at runtime.
   • Bake sizes: 1x at 512×512 (about 1 MB) for self, party and the 20 nearest players; 0.5x at 256×256 for everyone else in the map. At 150 players this is roughly 21 × 1 MB + 130 × 0.25 MB ≈ 54 MB.
   • Bake runs on idle frames, a few characters per frame; an unbaked character draws from shared atlases meanwhile.
4. Equipment icons are renders. Each gear item’s icon is rendered from its own parts on a mannequin pose with a fixed light; no separately painted icon.

Animation list (player)

Durations at 1x speed. Combat timings are defaults; skill data overrides them.

About 30 body animations at launch, each keyed once on the rig. Class advancements add skill poses by reusing these with new VFX, not new animations; budget 4 new animations per advancement at most.

Monsters, bosses and titans

Size classes

Monsters use the same cutout tech as players with fewer parts. Bosses alone may use Polygon2D skinning (one or two on screen, not 150).

The existing slime, cave bat and skeleton are regenerated under these rules as the Puddling, Dimwing and Afternooner (02-atlas.md); a first single-sprite pass of all three (per-kind tinted translucent body, face, rim and an ember scaled by held fire in game/render/lampghost.gdshader) is in the game, waiting on the cutout rig. The Cinder Eye becomes a field-boss-class creature drawn as a warm fire that wrongly hungers for lanterns: it is the one enemy that uses heartfire colors, which is exactly what makes it unsettling.

Natural versus Lampghosts

• Natural creatures are painted like everything else, warm-lit, rounded, cute.
• Lampghosts follow signature 4. A Lampghost version of a natural creature reuses its rig and shape with a tinted coldlight body, rim, a new face and a stolen ember: one rig, two monsters, half the art cost of a new species.
• Lampghost elites show both: patches of painted creature with coldlight spreading through it, marked by cold cracks, and the stolen ember in the coldlight part. Without the ember they are cut.

How titans are drawn

• A titan is a set of layered painted pieces, not a sprite: far silhouette (0.5x, one color mass plus heartfire seams and a few lit hold windows), mid body parts (legs, head, tail as separate pieces for slow cutout motion), and near pieces used only in hold maps and titan-pass events.
• Motion is slow: a leg lift and plant takes 8–12 s; the head turns over 20 s. Silhouettes are animated with the same cutout rig at very low speed, so a walking titan is about a dozen animated pieces.
• Partial reveals do the work: in a hold, the titan’s head is visible at the far left and its tail at the far right; in a canyon, only a passing leg, then dust; in Wend’s Fall, the player walks on its ribs.
• Atmospheric perspective: far layers fade toward the sky’s horizon color by 50–80%; only the heartfire seams keep full saturation. The Moth Queen is the exception: her wing glow reads cream until level 68 (dusted) and cold after.
• Titan-pass event. When a titan’s route crosses a hunting map (server-driven), the map darkens under its shadow for 30–60 s, a foot lands in the mid layer, dust falls across the gameplay layer, loose monsters flee, and the titan’s leitmotif plays. Its heartfire lights the map from above for the duration.

Boss telegraph language

Every telegraph is coded three ways: shape, color and sound.

• Lampghost telegraphs are cold (#7FE3F0), Cinder Eye and fire bosses warm-red (#E0533F); both use the same shapes. Warm-gold pools are always safe.
• The boss’s body poses a readable windup on the same beat as the telegraph (the wind-up pose is held the full lead time).
• Each telegraph type has a distinct rising sound that ends on the impact.
• Lead times shrink by at most 30% on hard modes; shapes never change meaning.

Environments

Layer stack (per map)

Foreground silhouettes cover at most 15% of the screen and never a walkable edge or an actor’s hitbox at standard zoom.

Map kits

Collision is footholds: line segments the designer draws (04-architecture.md §5.2). Art is never a grid:

• Surface strips. Platforms are painted as strips with a top lip, a body and left and right end caps. Body segments come in 128, 256 and 512 ref px lengths with 3 variants each; the top lip overhangs the collision edge by 8–16 ref px; strips can sag or bulge up to 24 px vertically between collision points without changing collision (the art floats over the foothold, as MapleStory does).
• Thin platforms (one-way) are visibly thinner (32 ref px body) with a wood, branch or shell-plate look, and always supported (posts, brackets, a trunk); solid ground has a deep painted body that fades to dark.
• Ropes and ladders. Repeating 64 ref px segments with top and bottom caps (a knot or hook at the top, a frayed end at the bottom; never a ball or arrow). Ropes are warm hemp, ladders warm wood, chains warm iron. A climbable 12 units or wider is a ladder. Every climbable has a lantern hook at its top in maps with ambient below 0.7.
• Decals and large props (128–1024 px) overlap freely: roots, rocks, moss mats, shell seams, hold buildings.
• Kit size per region: about 60–120 pieces (surfaces, caps, climbables, 30–60 props, 4–6 parallax layers, 1 LUT, 1 sky).
• Strata kits. Each region’s two strata (08-scale.md §3) reuse the surface kit’s strips and props under a darker LUT and ambient, plus a 20–30 piece deep extension (roots, cave lips, buried Concord ruin, Lampghost absences). The Deep’s map pieces are built from these kits; they add layout work, not art.

Standable surfaces

Anything drawn in the play plane that looks like a surface is standable on its visible top. A prop that is not standable must read as background or offer no flat top:

• Standable props are defs (props[].def) whose footholds run along every flat-enough visible top: porches, balconies, roof ridges, crate and barrel tops. A top too high to reach still gets its foothold; reach is the map’s business, honesty is the art’s.
• Background decor is an art prop marked "back": true: drawn behind every other prop, darker, desaturated and hazed toward the map’s low sky (back_prop.gdshader), so it reads as behind the play plane. Trees (a canopy is a round, inviting top), lamps and the hearth are back props.
• Front defs (front) draw over the ground and are not surfaces; nothing on them may look flat and standable.
• tests/test_map_file.gd holds every map to this: each prop has footholds, is back, or is a front def. If frames still read poorly, the next step is a lit lip along walkable prop tops, as the ground strips have.

Portals are drawn at every portal’s feet as a slowly turning, breathing swirl of lantern light (kit/portal, added onto the scene, game/render/portal_view.gd). The sim widens every trigger to at least Zone.PORTAL_REACH (40 units) around its centre, so maps need no edit, and Up takes a portal as E does (03 §2).

Authoring

1. A designer draws footholds, walls and climbables in the Godot map plugin (addons/mapkit).
2. The designer dresses it with the region kit: surface strips are placed automatically along footholds (an auto-dresser picks variants and caps, seeded per map so it is stable), then props and decals by hand.
3. Lights, parallax layers, sun side, ambient level and titan anchor points are set in a per-map resource.
4. The map saves to the data-only map format the platform uses (see docs/platform.md), so creators can later dress maps with the same kits.
5. Review: grayscale test, a phone-zoom screenshot, and a light-off screenshot (ambient 0.25) to check that walkable edges are still visible.

Target: a hunting map takes a designer 1–2 days to dress once its kit exists.

Built so far (look pass, town kit, Mossway maps). Lantern Town and the Mossway’s Mossy Field, Footprint Ponds and Old Tollbridge use this stack. The sky gradient, sun band and disc are drawn in code from the region palettes; the titan, far and mid layers, foreground silhouettes, props and kit pieces are painted locally with Z-Image Turbo by tools/art/world.py (prompts, seeds and crops in tools/art/world.json, only the finished PNGs are committed under assets/world/). game/render/ground.gd dresses the footholds with no extra map fields:

• Solid runs: a painted edge strip (kit/moss in holds and the Mossway’s fields) laid along the run at its own aspect, its walkable top on the foothold with the 2 ref px top line, overhanging and fading out at both ends; under it the body texture (kit/shell, kit/soil) fades to the region’s deep colour. A solid run with a foothold below it is a tapered shell-plate slab.
• One-way runs: a 32 ref px boardwalk deck (kit/deck) standing on timber posts (kit/plank) at most 90 units apart, nudged off even spacing so a row never reads as scaffolding, each reaching down to the nearest foothold below (or the map’s bottom); the end posts brace inward and rise above the deck as its caps, inner posts brace at most one side. No floating platforms.
• Climbables narrower than 12 units are hemp rope (kit/rope) with a knot on an iron hook at the top (kit/knot) and a plain frayed end of loose strands (no coil or tassel, which reads as a noose); 12 units and wider are wooden ladders. Every climbable hangs a lantern hook (kit/hook) at its top.
• Props for the Rim: town/house (a def, door at ground level, 74 units wide, its roof ridge walkable), town/crates (a def, 32 units, barrel and crate tops walkable), and the back props town/lamp (about 22), town/tree and the hearth; the Mossway has back props field/tree, field/lamp and ponds/lantern_lily.
• The Mossway’s defs: ponds/lily (a giant lily pad on its stem, its pad a one-way foothold), ponds/water (a translucent pond lens with over, drawn over the characters wading in a crater) and tollbridge/span (a two-arch Concord span, pier to pier, its deck a one-way foothold; three join into the bridge). Every surface in the play plane that looks walkable, deck, mound, pad or bridge, carries a foothold; trees, lamps and lantern-lilies are back decoration, hazed behind it.
• Buildings are prop defs (props[].def): a JSON beside the art (assets/world/<def>.json: ref size, anchor at the feet, footholds and climbables in ref px) whose walkable lines and ladders the loader adds to the map, mirrored when the prop is flipped, so a porch, balcony or roof is walkable where it is painted. game/render/props.gd draws props behind the ground, a def with front over it, and one with over over the characters; ground.gd dresses only the map’s own footholds but every climbable. The Rim’s four Lodges and storage hall are defs; the Stair head is a front def drawn over the ground slope that is its stair.

Layers fade into each other: a layer may carry a mist that hazes its base into the layer in front, so no band has a hard bottom edge, and every looping layer is wider than a 1920 ref px frame so it never repeats on screen. In a hold the near layers are the hold itself: the town’s titan layer is Mossback’s head far ahead, behind it the shell’s dome with the upper rings (town/dome, parallax 0.12, not looping), and nearer a terrace ring of upright houses on shell plates (town/mid). Fields keep the titan walking on the horizon. The town’s far layer drifts 4 ref px/s and the scene bobs 3 ref px every 10 s. The light overlay treats the sky around the sun as a light, so a Dusk map dims the ground but not the sun. Per-map values live in game/render/looks.gd until mapkit sets them. The tile rows in *.art.json are retired as art (still loaded, ids append-only). Still to do: strip variants and sag, chains as their own kind, more Rim props (stalls, NPCs), weather and motes, the actor rim shader and the LUT grade.

UI and UX art

Style

UI is the “lantern-glass” look: panels of deep dusk translucency (#1B1A33 at 88%) with a thin brass frame (#C99B4A), warm parchment text, rounded corners of 12 ref px, and a soft inner glow along the top edge as if lit from above by a lantern. Windows have painted brass corner pieces (9-slice). No Godot default theme anywhere, no flat gray, no hard drop shadows.

Typography (all SIL Open Font License)

All text has a 2 px outline in #120F1C at 70% or sits on a panel; nothing relies on a blur shadow. Fraunces and Nunito are variable fonts, whose overlapping outlines Godot’s MSDF breaks into stepped, notched edges, so they rasterize per size instead; the static Lilita One stays MSDF.

HUD layout, desktop

+--------------------------------------------------------------------+
| [minimap]                                     [quest tracker]      |
| [buffs]                                       [party frames]       |
|                                                                    |
|                         (gameplay)                                 |
|                                                                    |
| [chat: 6 lines, tabs]                                              |
| [Lv 34 Name ][ HP ======== ][ Resolve === ] (lantern) [hotbars     |
|  ================ XP bar (full width) ==================  2 x 10 ] |
+--------------------------------------------------------------------+

• HP bar heartfire-red to ember gradient. No MP: the second bar is the class resource (names and rules in 03-systems.md), each with its own fill: Warden Resolve brass, Kindler Heat ember with a band marking its bonus zone, Ranger Focus leaf-gold, Shade Gloom void-black with a warm edge. XP heartfire gold.
• The lantern gauge sits between the bars and the hotbars: a painted lantern that always glows. Below the surface (the strata and the Deep) its glass is the fuel meter (flame height = fuel; it gutters below 20%), with the number of carried motes orbiting it as small sparks and a count. In the fields and holds the lantern burns full with no fuel or load readout (fuel shows only while it is below full, after a run) and a line above it gives your banked motes and the Sunline. A hooded lantern shows a closed shutter. It is the most important gauge in the game and is never hidden.
• Hotbar count, key labels and layout are 03-systems.md’s; this sketch shows 2×10.
• Boss bar: top center, 960 px wide, the boss’s name in Fraunces, phase ticks.
• Built in game/ui/hud.gd and game/ui/hud_layout.gd on game/ui/lantern_glass.gd (the window theme is game/ui/lantern_glass.tres, set by game/main.gd), laid out in ref px and scaled by min(height / 1080, width / the bottom bar’s width). The level plate and XP bar read game/sim/progression.gd and the Resolve bar the sim’s Resolve; until the sim has it, the name is a placeholder (“Wanderer”). The lantern gauge is game/ui/lantern_gauge.gd, the guide line sits where the quest tracker will, and the hotbars are two rows of 10: skill bar 1 (game/ui/skill_bar.gd, keys 1–0, painted placeholder medallions with a cooldown sweep and a level lock) over the inventory’s first 10. The XP bar is 20 ref px tall with “EXP xp / next (pct%)” across it, so it reads at 1920x1080 (it was a 10 px line).
• Trip boons (03-systems.md §7.9), game/ui/boon_prompt.gd: the held boons are a row of six round placeholder icons (brass rim for Warden boons, heartfire for lantern boons, the initial in the middle; name and text as a tooltip) right-aligned over the lantern, with the motes toward the next offer beside them. An offer is three cards top center under the boss bar: icon, name, text and its key cap. It never pauses or takes focus; the cards and row take clicks and taps, so clicks on them never swing.
• Windows (inventory, equipment, skills, stats, guild, shop, mail) open as draggable panels that remember position.

HUD layout, mobile (landscape)

+--------------------------------------------------------------------+
| [HP/res/Lv] (lantern) [buffs]           [minimap] [menu]           |
| [quest: 1 line]                                                    |
|                                                                    |
|                           (gameplay)                               |
|  ( move   )                              [s1][s2][s3]  [ctx1]      |
|  ( stick  )                           [s4][s5][ ATTACK ][jump]     |
| [chat: 1 line, tap to open]           [s6]  [potion]  [ctx2]       |
+--------------------------------------------------------------------+

• Six skill buttons plus two contextual buttons (hood, oil, interact, beacon, as the situation needs). Counts and touch-target sizes are 03-systems.md’s; this layout draws them. Controls stay out of the bottom 16 dp and the notch sides (safe areas).
• The lantern gauge sits next to the HP and resource bars, as on desktop.
• Windows open as full-height sheets from the right, not floating panels. Inventory is a scrolling grid with 64 dp slots.
• UI scale defaults: desktop 1.0, tablet 1.15, phone 1.4; the player can set 0.75–2.0.

Iconography

• Item icons are 64 ref px at 1x (128 master), shown at 32–64. Three-quarter view, object centered, top-left key light, 2 px colored outline, no background. Must read as a silhouette at 32 px.
• Equipment icons are renders of the paper-doll parts (see above); consumables, materials and quest items are generated on the icon template.
• Rarity is shown by the slot frame, not the icon: frame color and frame shape (plain, notched, double, faceted, crowned) and a corner gem count. Never rarity by color alone.
• Skill icons are 64 px round medallions: a brass rim with a class motif (Warden shield, Kindler flame, Ranger leaf, Shade crescent), a painted action inside, and a warm or Lampghost color temperature matching the skill’s element.

Accessibility

• Color vision: telegraphs, damage types, rarity, party markers and map markers are coded by shape and pattern as well as color. Three colorblind presets (protan/deutan/tritan) remap UI and telegraph colors; the world palette stays.
• Text: size slider (independent of UI scale) 80–200%; dyslexia-friendly body font option (Atkinson Hyperlegible).
• Motion and flashes: screen-shake toggle and intensity, hit-stop toggle, “reduce flashes” mode that caps flashes at 3 per second (the WCAG 2.3.1 threshold) and replaces white flashes with tints.
• Clutter: sliders for other players’ skill effects (0–100%) and their damage numbers (on, party only, off). Crowded towns default to 40%.
• Contrast: a “high-contrast Lampghosts” option that adds a 3 px bright rim to Lampghosts and a light outline to walkable edges.
• Audio: subtitles for every bark and story line, visual sound cues for telegraphs.

VFX language

Light and dark families

Class motifs

Production rules

• Generated single-frame textures (smoke, sparks, petals, glyph rings, slashes) animated in Godot with shaders (dissolve, UV scroll, rim) and particles; flipbooks only if a video-model spike proves them (see the pipeline). Particles via CPUParticles2D (consistent on the Compatibility renderer used on web).
• Budget per skill: 64 particles and 3 flipbooks on mobile, 128 and 5 on desktop.
• Other players’ effects render at reduced opacity and particle count by default.

Damage numbers and hit feedback

• Numbers in Lilita One with a painted gradient fill and a dark outline: dealt heartfire gold, crit larger (1.4×) in ember with a burst behind, taken in Danger red with a downward drift, heals in Heal green with a “+”.
• Multi-hit numbers stack upward in a column (MapleStory’s readable pattern), 28 px spacing, rise 40 px over 0.8 s, then fade. At most 8 columns per target; extras merge.
• Damage number styles are a cosmetic slot (“damage skins”), so the art system must support swapping font, fill and burst per player.
• Hit feedback (presentation only, never in the sim): hit-stop 50 ms on hit and 100 ms on kill or crit; one-frame white flash then tint; knockback from the sim; screen shake 4 ref px for 120 ms on player hurt, 8 px on boss slams; impact flipbook at the contact point.
• Motes. Hitting a Lampghost knocks out a few sparks; putting one out releases its motes, which drift to the nearest lantern (00-canon.md, World rules). In a field they touch it and run up its thread as a brief bright streak, banked. Below the surface carried motes are visible on every character as a small swarm around the lantern (count-capped above 12), so a player loaded with cargo reads as a prize. On death they scatter to the ground as pickups. Banking them at a titan or beacon plays a rising stream into the heartfire.
• Natural monsters pop into leaf, spore or ember motes on death; no corpses.

Audio

Direction

Acoustic and warm in the holds, sparse and strange in the dark. Prompts ask for real acoustic instrument timbres, with ambient electronics only as texture below. No orchestral-epic default; the holds should sound like places people live.

Music by hold and region

Leitmotifs

• Heartfire / main theme. A rising three-note figure (scale degrees 1-5-6) that opens the title theme, plays on level-up and lantern relight, and appears in every hold theme in that hold’s instruments.
• Titans. One shared low, slow four-note call (horn or bowed contrabass, very low) with a variation per titan: Mossback in major, Greyhorn modal on brass, Ystra high on choir, the Moth Queen’s never played in full (only the first two notes, then silence, then wingbeats). It plays during titan-pass events and when a titan comes into view on the world map.
• The Lampghosts. The Lampghosts wear the shapes of what they consumed, so their music does too: Lampghost cues are the nearby region or hold theme reversed, detuned (−30 to −50 cents) and stripped of its bass. Near a Lampghost nest the region’s theme slowly “inverts” into its Lampghost version.
• The Cinder Eye. The heartfire motif played wrong: too fast, on distorted brass, never resolving.

Adaptive music

Godot 4.3+ has AudioStreamSynchronized (vertical layers) and AudioStreamInteractive (clips with a transition table), enough without middleware.

• Vertical layers per region cue: explore (always), combat (fades in within 1 s when the player or party engages, out over 4 s after), danger (Lampghosts nearby or HP < 30%, replaces the top layer with its Lampghost-inverted version).
• Horizontal resequencing for bosses: intro, phase loops, phase transitions on bar lines, victory sting.
• Darkness depth: ambient level of the map controls a low-pass and the share of warm instruments, so music literally dims in the dark.
• Titan proximity: titan motif enters as an overlay during passes, keyed to the region’s key.

SFX language

• Every hit has a sound; every skill has windup, release and impact.
• Footsteps per surface: moss, shell, wood, stone, metal, cloud (Ystra), silk rug (market).
• Telegraph sounds rise in pitch and land on impact; never reused for anything else.

Ambience

Each map has a 2-minute ambient bed plus 6–12 randomized one-shots (birds, creaks, distant Lampghost calls). Holds add the titan’s breathing: a 10–14 s rise and fall in the low end, synced to the visual footfall cycle. Crowd ambience is not faked in towns full of real players.

Voice

Barks and emotes only; no voiced dialogue at launch. - Players: 4 voice types (two higher, two lower), about 25 lines each (efforts, hurt, downed, level up, 8 emotes), in a short constructed gibberish language so nothing is localized. - NPCs: syllable “voice blips” per character, pitch and timbre set per NPC (Animal Crossing’s approach), plus named NPCs with 6–10 gibberish barks. - Bosses: 5–10 non-verbal vocalizations each.

Audio budget and tech

• Voice priority: own player, then party, bosses and telegraphs, then nearest. Other players’ skill sounds play at −9 dB and are culled beyond the nearest 8.
• Mix bus layout: Master > Music, SFX (Self, Party, Others, World), Ambience, UI, Voice; sidechain duck of Music and Others under telegraph sounds.
• Launch music total: about 90 minutes of finished music (≈ 35 cues with stems).

As built (placeholder, D81–D82). tools/audio/synth.py synthesizes everything (seed 81): mus_lanterntown_day_explore (48 s loop, D major, lilting 6/8 at a dotted quarter of 80: nylon guitar arpeggios, hammered dulcimer and fiddle on the heartfire figure, bassoon in the B section, frame drum and shaker) and mus_mossway_day_explore (60 s, G major, 4/4 at 96: marimba, guitar and flute, thinner, with Cm and E♭ for the minor turn), Ogg Vorbis 44.1 kHz stereo; 15 SFX as 16-bit mono WAV at 32 kHz. Level-up and lantern relight play the heartfire figure (1-5-6) on dulcimer. client/sound_cues.gd maps each tick’s change in one’s own character to a cue (a swing per class, hit, hurt, down, a monster’s pop, pickup, level-up, portal, relight, gutter below a fifth, snuff) and each map to its theme; client/audio.gd plays them on Music, SFX and UI buses, crossfades themes over 2.5 s, clicks on every button, plays nothing headless and nothing on the web until the first input.

Production pipeline: fully AI-generated

Canon fixes it: every shipped image and sound is AI-generated, directed and curated by us. There are no commissioned artists or composers and no paint-over requirement. People here are art directors, curators and toolsmiths, not painters: they write the style bible and prompts, pick and reject outputs, train the style models, build the tools that cut, align and check assets, and key animations on the rig.

The danger of this choice is the look Foster rejected: generic, inconsistent, obviously generated. Consistency therefore comes from four things, in this order: 1. Templates. Anything that must fit something else (character parts, platform strips, icons, UI frames) is generated on a fixed template image with control conditioning, never free-form. 2. Our own style models. LoRAs trained only on outputs we approved. 3. Automated QA. Scripts reject anything off palette, off silhouette, misaligned or seamed before a person sees it. 4. Curation. A person approves every asset against the style bible; nothing ships unreviewed.

Machine and toolchain

One Apple M5 Max with 128 GB of unified memory (this Mac) runs the whole image pipeline locally with mflux 0.20 (MLX-native). A second Mac can be added for throughput; nothing needs a cloud GPU.

Only models whose licenses allow commercial use are on the production allow-list (tools/art/models.toml); CI rejects an asset whose provenance names a model off the list.

Excluded from production (non-commercial weights): FLUX.1 [dev] and everything built on it (dev-fill, dev-kontext, dev-redux, dev-depth, dev-controlnet-*, CatVTON on dev-fill, in-context LoRAs on dev), FLUX.2 Klein 9B, OpenPose (CMU academic non-commercial license; commercial use needs a separate CMU license), Apple Depth Pro (weights under the Apple ML Research license, research use only, which covers mflux’s mflux-save-depth) and Depth Anything V2 Base/Large (CC BY-NC 4.0). They may be used for internal exploration only, never for a shipped asset or a LoRA training set. Krea 2, Ernie-Image, Fibo, Ideogram 4 and others mflux ships stay off the list until someone reads their licenses.

Style bible and reference sets

• The locked style bible is this document plus a reference board in art/refs/ (separate art repo, Git LFS): 40–80 approved images, grouped as characters, monsters, Lampghosts, environments by region, props, icons, UI. Every image is labeled with what it is a reference for (“palette”, “line”, “brushwork”, “scale”).
• Prompt templates per asset type in tools/art/prompts/, with fixed blocks (style block, lighting block, palette block, negative block) and a slot for the item. No artist names, game names or “in the style of” phrases, ever.
• Bootstrapping the style: the current key art in img/ has the right mood and the wrong finish (smooth CG). Round 0 generates 2,000–4,000 candidates from the templates, selects the 60–120 that best hit the painted-brushwork look and the palette rules, and trains style LoRA v1 on them. Round 1 generates with v1, curates again, trains v2. Expect the look to settle by v3. After that the LoRA only changes through a versioned retrain, and every asset records which version made it.
• LoRA set: wh-style-env (backgrounds, kits, props, titans), wh-style-char (characters, gear, monsters), wh-style-icon (icons and UI ornaments). Plus small per-subject LoRAs where identity must hold across hundreds of images: one per titan, one for the base body and face.

Per-asset-type workflows

Characters: the hard part

Characters must be consistent across ~1,000 gear items and every one must line up on the same cutout rig in every pose. Free generation cannot do that, so the pipeline generates gear on our own rig, then cuts it apart with the rig’s own part map.

1. The canonical mannequin. A base body built once (curated generations on a rig-drawn skeleton + depth control from a hand-built proportion template, then locked, face included). It must stand in a true open A-pose: spike A showed that a base with arms touching the torso tears at the shoulders and hips when posed. It is rigged and stored as a canonical part atlas: each part (head, torso, upper arm, forearm, hand poses, thigh, shin, foot, for the front three-quarter view and the back view) has a fixed canvas size, a pivot, joint sockets with overlap margins, and an id color in a part map (an image where each pixel’s color says which part it belongs to).

2. The reference pose. Gear is generated on the mannequin in one reference pose: an open A-pose, three-quarter view, arms clear of the body, legs apart, so almost no part hides another. The pose renders to control images (rig skeleton, depth, silhouette line) and to the part map. A matching back-view reference pose exists for climbing.

3. Generating an item. For a slot (say a hat), the base mannequin render on a flat key color goes into FLUX.2 Klein 4B edit with an instruction (“add a dented iron helmet with a lantern crest”) and wh-style-char. One item per pass, always onto the same base, never stacking items. Qwen-Image-Edit is the fallback when Klein will not draw a given item. Then: - Envelope: each slot declares its envelope (the region where that slot’s pixels may appear, with room for brims, plumes and capes) in the slot definition, not as fixed image rows. Pixels changed outside the envelope are either rejected or, if they belong to another slot (Klein added trousers with armor in spike A), routed to that slot or the generation is redone. - Extract: chroma key the background, then a color-matched difference against the base render, then morphological cleanup. Spike A found this alone gives clean layers; BiRefNet matting is the fallback for soft edges (fur, plumes). - Split: the part map assigns each item pixel to a part (a sleeve’s pixels split into upper-arm and forearm pieces at the elbow socket). Pixels near a joint are written to both neighboring parts so joints overlap. - Fill: each part’s area hidden in the reference pose (the inside of a sleeve behind the torso edge) is completed by a masked regeneration on the part’s own canvas. - Back view: the same item generated on the back-view pose with the front result as an image reference (Qwen-Image-Edit or FLUX.2 Klein 4B edit), then extracted and split the same way. - Dye regions: colors clustered by hue into up to three regions; the item is stored as value plus an RGB region mask, so one generation yields its tiers and dyes. - Variants: tier variants (trim, studs, glow) by low-strength img2img on the extracted item with the same mask, not a fresh generation, so tiers share shape.

4. Faces, hair, hands. Face sets: one neutral face generated on the head template, then each of the 12 expressions made by instruction editing (“eyes closed, smiling”) with the face mask locked. Hair: generated on the head template in full, then under_hat and back_only made by masked edits of the same hair. Hand poses: five fixed templates, generated once per glove item.

5. Limb angle variants. Each limb’s 2–3 angle swaps (extended, bent, foreshortened) are separate reference poses with their own part maps; gear parts for those limbs are generated on them too. This is the most expensive part of an item and is limited to arms.

6. Automated QA, in order (fail = regenerate, no human time spent):

Survivors go to the curator as a contact sheet: the item on the mannequin in the 8 poses, dyed three ways, beside its tier siblings. Approve, reject with a reason (the reason is logged and feeds prompt fixes), or send back for a variant.

7. Animation. Animations are keyed on the rig by us in AnimationPlayer (or written as data by a script from pose references). AI pose references help; AI does not key cutout animation. This is about 30 player animations, done once.

Spike A result (2026-10-03, this Mac): PARTIAL.

Passed: one-item-at-a-time editing onto a fixed body is the gear generator; Z-Image Turbo img2img is out for gear. Failed: the rig split. The base body was not in an open A-pose, so the walk and swing test poses tore at shoulders and hips; and fixed-row envelopes caught pixels from the wrong slot. Next, in order: (1) a true open A-pose mannequin with a locked face; (2) a drawn part map with joint sockets and overlap margins; (3) slot-defined envelopes; (4) 10 seeds each of a hat, a top and a cape through extraction, split and the 8-pose seam check. Pass: ≥ 1 of 10 per item survives the seam check with no hand-fixed pixels. If the split still tears, the fallback is generating each part on its own canvas with Klein, using the part’s silhouette as the reference, which aligns perfectly but risks style drift between parts.

Spike A3 result (2026-10-03, this Mac): PASS on the open mannequin. An open A-pose mannequin with a cutout rig, hidden socket discs under each joint and a separate pelvis part:

Next: a side or three-quarter view mannequin, a tunic skirt chain for long tops, and a readability check at 1x (112 px) in the client.

Audio pipeline

All music, SFX, ambience and barks are generated. The audio direction above stands; these are the tools that produce it.

Music workflow. 1. Per cue, a written brief from the music table above (instrumentation, tempo, meter, key, mood) becomes the prompt; leitmotifs are written as short melodies in MIDI by us and rendered with a CC0 soundfont as a guide track. 2. ACE-Step generates candidates conditioned on the brief and, where it supports audio reference, the guide track, so the heartfire and titan motifs actually appear. 3. Stems: generate the explore layer, then generate the combat and danger layers as extensions/variations of the same piece (ACE-Step’s repaint and edit modes), or split one full mix with Demucs into layers. Spike F decides which holds sync better. 4. Lampghost versions are made by script: reverse, detune −30 to −50 cents, low-cut. No generation needed, which also makes them recognizably “the same tune, wrong”. 5. A loop tool finds bar-aligned loop points (beat tracking), crossfades, and checks seam loudness; cues are normalized to −16 LUFS (music) and fit the per-region download budget.

SFX workflow. A sound list generated from game data (every skill × windup, release, impact; every monster × attack, hurt, death; every surface × footsteps), batched prompts to Stable Audio, 6–10 candidates each, automatic trim, peak normalize, onset alignment (the transient lands at sample 0 for hits), duplicates rejected by spectral similarity. Telegraph sounds get a fixed rising template so they stay recognizable.

Spikes: F, can ACE-Step hold a leitmotif and produce layerable stems; G, are Stable Audio SFX punchy enough for hits (short transients are where generative audio is weakest); H, gibberish barks with emotional range.

Legal and store policy

Cleanup and assembly (all automated, curator approves)

1. Generate N candidates from the template prompt and conditioning (N by asset type: 8 for icons, 20 for gear, 40 for key backgrounds).
2. Automatic QA (the per-type checks above) drops failures; the reasons are logged.
3. Cut, split, fill, dye-mask, downscale (Lanczos from the generation size to 2x, 1x, 0.5x), trim, pivot.
4. Curator reviews contact sheets in a small web tool (approve / reject with reason / variant).
5. Export with a provenance sidecar; approved outputs go into the next LoRA training pool if tagged “exemplary”.

Files, naming and atlasing

• Source: generation records, not layered paintings: the prompt template id, filled prompt, seed, model and LoRA versions, control images and masks, plus the raw output. Any approved asset can be regenerated or re-varied exactly. Stored in a separate art repository (Git LFS), not in the game repo.
• Export: tools/art/export.py writes 2x PNGs, 1x and 0.5x downsamples, trims, records pivots from the part atlas, and writes the sidecar.
• Naming (lowercase, underscores):

• Provenance sidecar <name>.prov.json: model and version, LoRAs and versions, prompt template and filled prompt, seed, control inputs, QA results, curator, approval date. CI fails a build that ships an asset without one or with a model not on the allow-list.
• Atlasing: our own packer in tools/art/ (MaxRects, 2048² pages for 1x, 4096² for 2x) writing PNG plus JSON rects, so the same format serves the data-only map and creator packages. Atlases are grouped by what loads together: one per region kit, one per monster family, gear by tier-set, UI by window. Characters are re-baked at runtime anyway.
• Import: VRAM-compressed (BPTC/S3TC desktop, ETC2 or ASTC mobile), mipmaps on, linear filtering.

Launch asset count, effort and throughput

Counts are unchanged by the production method. Curation minutes include reviewing contact sheets, rejecting, re-prompting and approving; rig and animation work is keyed by us. Generation compute is local and nearly free; it runs overnight.

Tooling build (once, before and during S1):

Cost:

Previously, with commissioned art, this was about $510k–$990k.

Throughput per curator-week (30 focused hours, tooling done):

With tooling done, the full 10-region art and audio is about 106 curator-weeks: about 12 months for two curators, 8 for three, and more than two years for Foster with agents alone. D1 launches with regions 1–5 (roughly 55–65 curator-weeks, shared tooling and characters included), curated by Foster with agents on the 14–16 month path in 06-roadmap.md; regions 6–10 are made for the Act updates. Region count stays the biggest lever if the schedule slips.

Art milestone plan

S0: pipeline spikes and style lock (4–6 weeks)

Also in S0: generate round 0, train style LoRA v1 and v2, lock the reference board, and redo the three style frames (Lantern Town on Mossback at true titan scale, Mossy Field with Mossback’s leg in the mid layer, a deep-land map with three light pools and Lampghosts) with the LoRA. Exit: Foster signs off that the frames are not generic, and the S0 blind test passes (10 people, ours next to MapleStory, Terraria and two cozy indie games; 8 of 10 describe ours without naming another game).

S1: pipeline proof (8–10 weeks)

The first deliverable proves the AI pipeline produces a playable slice at consistent quality, in the real engine, on all three platforms.

What it must prove, with numbers:

S2 onward

• S2: Kindler and Ranger, Greyhorn’s hold kit, the first Lampghost-dense region, LoRA v3 trained on S1’s approved assets, this document updated with what S1 taught.
• S3: production: one region per 2–3 curator-weeks (Foster with agents, D1), gear in tier-set batches, audio per region alongside.
• S4: titans and bosses in parallel with regions; key art and store assets two months before launch.

Phase 1 art to make

What 06-roadmap.md’s Phase 1 build order still needs drawn, from the gaps in the asset catalog (generated from the repo by tools/art/catalog.py, published beside this bible; it shows what exists, its size, its footholds and who uses it, and flags anything unused). Generation is not part of this list. Sizes are sim units with ref px in brackets (1 unit = 4 ref px); masters are 2x, so a 28-unit character is a 224 px master cut from a 768×1024 generation, as today.

Prompt families. A prompt below is the subject slot only. The generator adds its family’s fixed blocks, which are the ones that worked in the look pass:

Rules that every row follows: side view profile, warm light from the sun side (left) on environments and neutral top-front on actors, the dusk palette of its region, no text, no cyan except absence (a cyan key would eat a coldlight edge, hence V’s magenta). Anything marked (proposed) is open in 07-decisions.md (D50–D55).

Classes: Ranger, Kindler, Shade

As built for the Kindler (build.py --outfit kindler, tools/art/vfx.py): parts_kindler.png and rig_kindler.json re-painted on the base bones (a rod extra held like the sword, upright at rest, and a pelvis region of its own), with Spark petal, glyph ring and ember puff in assets/vfx/kindler/ as premultiplied alpha (premultiplied in vfx.json). The painted robe already carries a book pouch at the hip, and every glyph book seed put letter-like runes on its cover, so the glyph_book part is cut but not drawn (TODO: a re-roll without marks). The skill bar keeps drawn glyphs until the skill icons above are generated.

As built for the Ranger (tools/art/player/player.json outfits, build.py --outfit ranger, tools/art/vfx.json): parts_ranger.png and rig_ranger.json cut on the base bones, the bow re-edited with its lantern and the arrow as an outfit extra; the three V textures in assets/vfx/, drawn additively (game/render/enemy_view.gd), the leaf spark also marking a Waymarked monster. The hood stays up (owner, 2026-10-04: the “hood and bow” silhouette, as long as the face reads). The cloak hides the back upper arm, so that part is absent and skipped.

Lodge yard and its dummies

The Lodge yard (10-operations.md §4) is a small hold map built from the town kit and Lantern Town’s palette and layers; it needs only these.

The Mossway: Dewline Hedgerows, Shellshed Slopes, the Soak

The Glasslands to the Cinder Eye

05 has no Glasslands palette yet; the rows assume the proposed one in D50: smoky amber lit from cracks below, glass warm amber and rose, never cold (glass must not read cyan).

The prologue: the Pilgrim’s Stair and the Lighting

UI: party, finder and class select

Decisions for Foster

Open art and audio decisions live in 07-decisions.md: D8 (native Godot cutout vs Spine), D9 (Wend’s key art as a beast), D10 (Moth Queen key art), D19 (music fallback), D20 (fonts), D22 (the Stability revenue trigger) and the Phase 1 art proposals D50–D55. Settled: launch scope and curation (D1, Acts I–II curated by Foster with agents), the Lampghosts name (D3), and in 00-canon.md the reference scale (1920×1080, 1 unit = 4 ref px, 8-unit tiles), “cyan means absence”, and Wend as a beast.

Sources

• US Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (2025-01-29): summaries at Jones Day, Jenner & Block.
• Steam AI disclosure rules, revised 2026-01-16: Notebookcheck, GosuGamers.
• Google Play AI-Generated Content policy: Play Console Help.
• Apple guideline 5.1.2(i) on personal data and third-party AI (Nov 2025): AppleMagazine.
• Clair Obscur’s Indie Game Awards withdrawn over generative AI, Dec 2025: Engadget.
• Z-Image Turbo, Apache 2.0, released 2025-11-26: invideo overview, Hugging Face README.
• FLUX.2 Klein licenses (4B Apache 2.0, 9B non-commercial): black-forest-labs/flux2, Flami overview.
• SeedVR2, Apache 2.0: ByteDance-Seed/SeedVR.
• Licenses checked 2026-10-03: Depth Anything V2 Small (Apache 2.0), Depth Anything V2 Base (CC BY-NC 4.0), Apple Depth Pro (apple-amlr, research only), OpenPose (CMU non-commercial), DWPose (Apache 2.0), rtmlib (Apache 2.0), Z-Image Turbo Fun ControlNet Union 2.1 (Apache 2.0), Qwen-Image-Edit (Apache 2.0), ACE-Step 1.5 (MIT), Stable Audio 3.0 license summary.
• mflux (0.20 installed here; LoRA training via mflux-train): mflux; Z-Image LoRA training in AI-Toolkit: guide.
• ACE-Step 1.5: Hugging Face, overview.
• Stable Audio 3.0 and the Stability Community License: The Decoder, license terms summary.
• Suno ownership terms after the Warner deal: Music In Africa, eesel review.
• ElevenLabs music and SFX licensing: Eleven Music API, invideo overview.
• Godot 4.3 interactive music and QOA: Godot 4.3 release.
• Godot 4.4 2D batching across renderers: Godot 4.4 release.
• WCAG 2.3.1, three flashes or below threshold: W3C.