# styleKit
Build your own art style out of stackable look blocks: a look is a stack of small blocks (a wash, an ink line, a paper, a grain) that runs over the whole frame after antialiasing, lives in the scene and looks the same for every player.
Version 1.7.8.
## Install and use
Install **styleKit** from ZeroMind (search "styleKit" or "stackable art style blocks"). Installs with several dependencies can be slow (#21420), so give an install time to finish and check `/zero/source` for the package folder. Then, from the console or any script, stack one of each of the kit's four example blocks:
```luau
local StyleKit = require("~styleKit.StyleKit")
StyleKit.add("posterize") -- Colour: a few flat tone steps
StyleKit.add("duotone", { strength = 0.5 }) -- Colour: two inks, laid half over the posterize
StyleKit.add("depthLine") -- Line: a clean ink outline from depth
StyleKit.add("sketchLine", { strength = 0.6 }) -- Line: soft pencil strokes laid over the outline
print(#StyleKit.describe().entries) -- 4
StyleKit.clear() -- the frame is exactly as rendered again
```
No material is edited. The first `add` makes a synced `StyleStack` entity in the scene, so every player draws the same stack; put a `StyleStack` component in your scene's build to keep a look with the scene.
## The blocks in this package
| Block | Slot | What it does |
|---|---|---|
| `posterize` | Colour | Cuts the picture into a few flat tone steps, like a screen print. |
| `duotone` | Colour | Maps the picture onto a shadow ink and a light ink. |
| `roleGlaze` | Colour | Roles glaze: a transparent glaze of one colour over the objects of one group (a variant), keeping their light and shade. `StyleRoles.use` adds one per group from a role set's `glazes`, so a hero, the enemies and the pickups that share a material still read apart under any look. |
| `depthLine` | Line | A clean, one-sided ink outline from depth plus creases from normals. |
| `sketchLine` | Line | Soft, broken pencil strokes over shape and colour edges, redrawn on the held clock. |
| `sharedView` | Medium | A diagnostic for block authors: shows one shared input in place of the picture. |
Every block lists its slot and its tags (the words a person would search for: `watercolor`, `pencil`, `halftone`, `pixel art`, `chalk`), so you can pick blocks by tag with `StyleKit.blocks({ tag = "pencil" })` or `zero styleKit list --tag pencil`. A block's tags come from its `block.yaml` `tags:` list when it has one, otherwise from its `.metadata` tags.
More blocks come in their own packages and plug straight in: `starterBlocks` has a block for every slot (card surface, light bands, flat wash, surface dots, plain paper, dark board, line boil, chunky pixels, soft grain, vignette), `watercolorBlocks` has ten chained watercolor Fill blocks and a cold-press paper, with pigment sets, whole looks and a hero material, `inkBlocks` has five ink lines, four ink fills, ink colours and whole looks, and `paperBlocks` has seventeen papers (cold press, rough, hot press, rice, newsprint, kraft, parchment, canvas, board, blackboard, green chalkboard, black paper, aged sketchbook, toned grey and tan, ruled notebook, vellum) that publish the paper field and layer by region. `StyleKit.blocks()` lists every block installed. A package can also ship pigment sets (`pigmentSet` assets: `StyleKit.pigmentSets()`) and whole looks (`styleLook` assets: `StyleKit.looks()`) next to its blocks.
## How a stack works
A stack is an ordered list of entries `{ id, block, on, strength, mask, settings }` on a `StyleStack` component. Blocks sit in ten slots that always run in this order:
1. **Surface**: what things are made of.
2. **Light**: how surfaces take light.
3. **Fill**: how areas of colour are laid.
4. **Tone**: marks that carry shade.
5. **Line**: edges.
6. **Colour**: palette and grade.
7. **Ground**: what it is drawn on. The first Ground block publishes the shared ground (and the paper field, when it has one); a later Ground block whose field says `layers: true` lays its ground over the one before by its strength and mask.
8. **Time**: how motion is held. The first Time block publishes the shared wobble.
9. **Raster**: how the image is sampled.
10. **Medium**: what it is seen through.
Any slot can be empty. Any block swaps for any other block of its slot with no other change (`StyleKit.swap(entry, otherBlock)`); strength and mask stay. Inside a slot the list order decides: the first block reads the slot's input, and every later block in the same slot reads the same input and is laid over it by its strength and mask (`stacking: layer`, the default). So two Line blocks with weights 1 and 0.4 blend 60 / 40, and two Fill blocks with complementary masks split the frame. A block whose `block.yaml` says `stacking: chain` reads what the blocks before it in its slot made instead, so modifier blocks build on each other (a wash, then its granulation, then its darkened edges). A block's `rank: <number>` decides where `StyleKit.add` places it: before the first block of its slot with a higher rank, so a chain lands in its painting order whatever order its blocks are added in.
### Keeping a stack
A stack made with these calls is a live edit of the scene, like moving a prop. Save the scene (`layers.active:save()` or the editor's save) and the stack is in its `scene.json`: it comes back on every load and after a restart, the same for every player. `StyleKit.clear()` takes every block off the frame; to get the saved stack back, revert the scene (`tools.use("scene", "revert")`). `scene.reload` keeps live edits by saving them first, so a clear followed by a reload is a saved clear. Keep one StyleStack per scene.
### Shared by every block in a stack
- **One clock.** `sk_time()` and the held clock `sk_held()` / `sk_step()`, stepped at the stack's `holdRate` (8 per second by default). Every wobble and boil runs on the held clock.
- **One set of coordinates that stay on things.** `sk_surface_coord(uv, perMetre)` lays marks on surfaces from world position along each surface's main axis, and on the sky from the view direction, so a pan or a dolly does not drag them.
- **Surface data copied once per frame**: view distance, world position, normal and base colour (`sk_dist`, `sk_world`, `sk_normal`, `sk_albedo`).
- **One ground** (`sk_ground_at`): paper colour and grain height, published by the first Ground block (plain white paper with flat grain when there is none).
- **Layered grounds** (since 1.2.0): a later Ground block whose `block.yaml` field says `layers: true` lays its ground, and its paper field, over the one before by its strength and mask, so regions of a world sit on different papers: `StyleKit.add("paperParchment", { id = "mapRoom", mask = { kind = "sphere", center = { 30, 2, 10 }, radius = 6 } })`. A Ground block without `layers: true` publishes only when it is the first.
- **One paper field** (`sk_paper`, read through `skPaper`, since 1.2.0): how the sheet under each pixel takes paint and ink, for a block that lists `paper` under `reads`: `sk_thirst(uv)` (-1 hard-sized, 0 the standard sheet blocks are tuned on, +1 thirsty), `sk_sheen(uv)`, `sk_fibre(uv)` and `sk_dark_ground(uv)` (1 where the sheet is dark and the picture is made of light marks). A Ground block publishes it with a `field.paper` shader; where none does every value is 0, so a block reading it draws exactly as before. It is only made when some live block reads it. paperBlocks' papers publish it; watercolorBlocks' bleed and edge darkening and every inkBlocks block read it.
- **The scene's sun** (`sk_sun`, a float4, since 1.2.0): for a block that declares it and lists `sun` under `reads`, the direction toward the sun the renderer shades the scene with (`lighting.sun()`), `w` 1 when it gives light (0, and a default direction, when there is none). Read every frame as a table the engine keeps current (nothing in the scene is surveyed), and written to blocks only when its direction (to three decimals) or whether it lights actually changes, so a still sun costs nothing and a moving one reaches blocks on the next frame. Raking light on paper tooth and sheen use it.
- **One wobble** (`sk_wobble_uv`): an offset published by the first Time block (still when there is none). A wash and an ink line that both read it tremble together.
- **One pattern library** (`skPatterns`): noise, fractal noise, cells, dots, hatching, ordered threshold, triplanar noise for materials (`skp_triplanar_fbm`) and brush strokes (`skp_stroke`).
- **One soft field** (`sk_soft`, read through `skSoft`): a half-resolution, depth-aware blur of what a block reads, for a block that lists `soft` under `reads`. Each tap is tested against the centre surface's plane rebuilt from world positions, so a nearer thing is never smeared over a farther one. The stack makes it for the first block that reads it and makes it again whenever a later reader reads a different image; its width is the stack option `softRadius` (reference pixels, 10 by default).
- **One pigment set** (`sk_pigments`, read through `skPigmentSet`): the scene's palette of paints as a 16 x 1 texture, for a block that lists `pigments` under `reads`. `StyleKit.setPigments(id)` changes it for every block and every player at once; with none picked blocks keep the scene's own colours.
- **One rendered frame** (`sk_rendered`, read through `skRendered`): the frame as it entered the stack, for a block that lists `rendered` under `reads`. Light estimates read it against the base colour, so surface texture cancels instead of printing back as shade. With `sk_geo_normal` (a normal rebuilt from world positions, every tier) it lets a block judge light and shape on the low tier and the forward path too.
- **One way to mix paint** (`skPigment`): pigment mixing in transmittance and optical density (log space), two pigments in one wash, glazes that multiply, dilution toward the paper. Pure math, so a full-frame block and a material mix paint the same way.
- **Noise that stays on things and keeps its size** (`skAnchored`): triplanar noise on world position that stays on still things (`sk_anchored_fbm`, `sk_anchored_cells`), and noise that also keeps its size in the picture at any distance by blending two octaves by distance (`sk_scaled_fbm`, `sk_scaled_level`, `sk_surface_plane`, `sk_ref_metres`), for paper tooth, flecks and strokes. `sk_tooth(uv)` reads the shared paper's valleys (0 on a peak, 1 in a valley).
- **Chain stacking** inside a slot (`stacking: chain` and `rank` in `block.yaml`): later blocks build on what the ones before them made.
- **Many blocks, one shader** (`defaults:` in `block.yaml`, since 1.2.0): a block's own defaults over its shader's, so a family of blocks (paperBlocks' seventeen papers) shares one shader and each block is its own starting point. `StyleKit.settingsOf(id)` reports the block's defaults.
- **Swap memory** (since 1.2.0): each stack entry remembers the settings of every block it has held, so `StyleKit.swap("paper", "paperKraft")` then `StyleKit.swap("paper", "inkPaper")` puts back exactly the settings `inkPaper` had, and the frame returns exactly. The memory is kept on the entry in the synced stack, so it is shared and saved with the scene; `load` and `applyLook` write new entries without it.
- **One exit** (`sk_finish`): strength, mask, layer mode, and black stays black: a block never makes light where the frame had none.
- **Strokes for shading blocks** (`skStrokes`, since 1.3.0): hand-laid strokes that build up by tone, for any block that shades with marks (pencil, charcoal, hatching, scribbles). Since 1.4.0 every layer of strokes at one place shares the same patch grid and the same random slant, so layers keep exact angles to each other; outside crossed mode the first two layers run parallel (the second fills between the first) and darker layers cross at clear angles (60 degrees each way, then 30, then square), so soft strokes never beat into broad zebra bands. `sks_layer_d(p, n, tone, dpx, dpy, px_per_ref, st, k)` sizes the strokes across their own width on screen from the position's screen derivatives, so the spacing is the same number of pixels whichever way strokes run and however steeply a surface is seen; prefer it over `sks_layer` wherever derivatives are at hand. Since 1.4.1, `sks_edge_on(facing)` says how much of its stroke pattern a surface keeps from how squarely it faces the eye, and `sks_layer_mean(tone, st, k)` is the average coverage a layer gives: on a floor or a field seen at a low angle, patches of strokes squeeze into rows whose ends read as zebra bands, so mix each layer toward its average there and the ground keeps its value as even rubbed tone.
- **The dim guard** (since 1.4.0): keeps a dark scene dark. See the next section.
### Keep a dark scene dark: the dim guard
A shading block lays marks and a sheet over the picture, and in a dim room the sheet and the light between the marks can end up brighter than the room was. The dim guard stops that, so no shading block (hatching, stipple, dots, pencil, charcoal) brightens a dim or unlit room.
- **When it runs.** When any live block in the stack declares `keepDim`. Every Tone slot block does unless its `block.yaml` says `keepDim: false`; a block in another slot opts in with `keepDim: true`. So a new shading block in the Tone slot gets the guard for free.
- **What it does.** Two small passes after the chain. A meter at an eighth of the chain's resolution records the mean brightness of the stack's result and of the frame as rendered under each of its texels, plus how bright the rendered picture's brightest parts are (a power mean over a fixed 5 by 5 grid) and its overall mean. Then the guard scales the result down wherever, smoothed over about 36 pixels, it is brighter than the rendered frame there (with no allowance since 1.4.4; it was 0.006, about 1.5 of 255). Then it holds each pixel to at most a fifth above the rendered brightness on its darker side: the darkest of four small patches ten pixels off (left, right, above, below), never below the pixel's own rendered brightness. Bare sheet beside a pale prop in a dark-walled room is held to the wall it lies on instead of keeping the prop's brightness and glowing around it, the prop's own edge to the prop, and a patterned surface to its mean, so no texture prints back. It scales rather than clips, so every mark keeps its shape.
- **Only on a dim or unlit picture.** It lets go fully once the picture's mean is 0.42 or above, its brightest parts are at 0.55 or above and some light other than an ambient one is on (daylight, an overcast day, a well lit room), and is fully on below a mean of 0.30 or brightest parts of 0.40. A scene with no light switched on (the engine falls back to a default ambient) counts as unlit however bright that fallback draws it. In a lit frame it hands the result on exactly unchanged. A night street, or a room with dark walls under one lamp, has bright spots but a low mean, so it is guarded.
- **Cost.** Two small passes, only while a block that keeps dim is live. `StyleKit._status().dimGuard` reports whether it ran last frame.
Blocks registered before 1.4.0 (the registry survives hot reloads) get their `keepDim` filled in by the kit, so an upgrade needs no restart.
### Masks
`StyleKit.setMask(entry, mask)` limits a block:
- `{ kind = "depth", near = 0, far = 12, soft = 0.2 }`: by view distance in metres. The soft edges lie outside the band and scale with their own distance: the block fades in over `soft` x `near` metres before `near` and out over `soft` x `far` metres after `far`, so nothing nearer than `near` x (1 - `soft`) is touched, whatever `far` is. The sky lies at the viewing camera's far distance (its `Camera.far`, 8000 m by default): a band whose `far` reaches that distance takes the sky in, a shorter one leaves it out.
- `{ kind = "rect", rect = { x0, y0, x1, y1 } }`: a screen region in 0..1 (top left is 0, 0).
- `{ kind = "sphere", center = { x, y, z }, radius = 5 }`: a region of the world.
- `{ kind = "group", group = 1 }`: the objects in mask group 1, 2 or 3 (`StyleKit.options({ groups = { { id, id }, { id } } })`).
- `invert = true` flips any of them.
`StyleKit.options({ exclude = { "SignBoard" } })` leaves objects (and their children) exactly as rendered: a prop, a world label. Name them: an entry is an entity name or an id, and a scene should store names, because an id written into the stack does not survive the scene being saved and loaded again (the entity comes back under a new id). Entities that arrive later are picked up within half a second. Put labels on a card of their own (a plain white quad behind the words), so the area left out around them is card rather than the raw scene.
### Players, enemies and pickups under one look: style channels
A look treats every object the same unless the object says otherwise. A `StyleChannel` on an object (synced, saved with the scene) says how the look treats it and everything under it:
- **variant**: which variant of the look it gets. A stack entry that names a `variant` draws only on objects of it, and an entry of the same slot that names none leaves those objects to it, so two blocks share one slot by group: one ink for the player, a different ink for the enemies. `"asRendered"` leaves the object exactly as rendered (labels, in-world HUD, a screen inside the world).
- **line**: how heavy its lines and strokes are (1 as the look draws them, 2 twice as heavy, 0 none). Read by styleKit's `depthLine` and `sketchLine`, every inkBlocks line, pencilBlocks' lines (width) and shading (stroke density), and toonBlocks' ink (width).
- **marks**: how much of each kind of mark reaches it, 0 to 1: `line` (Line blocks), `tone` (Tone and Light blocks: hatching, dots, bands), `fill` (Fill, Surface and Colour blocks: washes, cel fills, posterize) and `paper` (Ground blocks: how much paper shows through). Every block honours this with no change of its own: the kit weighs each block by its kind of mark on each object. A block may name its kind in `block.yaml` (`marks: line | tone | fill | paper | none`); otherwise its slot decides, and the Time, Raster and Medium slots are screen-wide.
- **value**: one free value, 0 to 1, a block may use however it likes; 0 changes nothing. The shipped blocks read it as "how much of its own colour the object keeps": watercolorBlocks' wash keeps its colour brighter and purer, inkBlocks' wash keeps its colour through the ink, toonBlocks' ink takes its colour.
- **ink**: the colour of its line, for blocks that take one per object (toonBlocks' ink).
- **role**: a game role (`hero`, `enemy`, `pickup`, ...). The scene's role set fills in every field the object does not set itself. Role sets are `styleRoleSet` assets; the **styleRoles** package ships sixteen roles tuned for six kinds of look.
```luau
local StyleKit = require("~styleKit.StyleKit")
StyleKit.give(entity.findAll("Goblin")[1], "enemy") -- one thing, one change
StyleKit.setChannel(entity.findAll("Coin")[1], { marks = { tone = 0, paper = 0 }, value = 0.8 })
StyleKit.setChannel(entity.findAll("ScoreBoard")[1], { variant = "asRendered" })
StyleKit.add("depthLine", { id = "enemyInk", variant = "enemy", settings = { ink = "#6a0a0a", width = 2.4 } })
StyleKit.setRoles("toonGame") -- the whole game's roles, one change
print(StyleKit.channelOf(entity.findAll("Goblin")[1]).line)
```
An object with no channel, in a scene where none has one, draws exactly as before: the kit makes no object mask and gates no block. Every field is synced on the object's own entity, so every player sees the same roles on the same things, they survive a save and reload, and a change made in play (by any player; a write from a player who does not own the object goes to its owner) reaches everyone. A child with a channel of its own keeps its own; things attached under a role holder later are picked up within ten seconds (at once when its channel changes).
**How it runs.** Objects are grouped by what they ask for into at most eight distinct settings; each setting is one geometry pass into a small image of its own, and one pass packs them into the object mask every block reads (`sk_obj`). A block reads its object through `sk_chan_line(uv)`, `sk_chan_line_near(uv, r)` (the nearest surface within `r` reference pixels, so a line drawn just outside a silhouette belongs to the thing in front), `sk_chan_value(uv)`, `sk_chan_mark(uv, kind)` and `sk_chan_tagged(uv)` from `skCommon`; the gate by kind of mark and variant is applied inside `sk_weight`, so `sk_finish` honours it for every block. Measured on toonBlocks' busy bench (2,500 things, 250 moving, 25 a second spawned and despawned, Comic ink cel), look stack script time a frame:
| | The whole look (median) | average | Toon ink (median) | styleKit (median) |
|---|---|---|---|---|
| Roles off | 0.44 ms | 0.53 ms | 0.11 ms | 0.32 ms |
| 150 things with a role, toon game set | 0.63 ms | 0.78 ms | 0.17 ms | 0.44 ms |
| 600 things with a role, toon game set | 0.93 ms | 1.25 ms | 0.26 ms | 0.64 ms |
| 600 things with a role, before this release (toonBlocks 1.3.5, styleKit 1.6.3) | 1.33 ms | 2.29 ms | 0.60 ms | 0.55 ms |
Measured on 11 November 2026 with toonBlocks' busy bench (`design/roles/bench_roles.luau` in the Style Cookbook world): 2,500 things, 250 of them moving, 25 a second appearing and 25 disappearing, 10 recoloured a second, under the Comic ink cel look; script time the look spends each frame. It ran in a test client of its own drawing 14 to 17 frames a second on a shared, busy desktop, which overstates every per-frame number; your game on its own will measure lower. What changed: toonBlocks 1.3.6 draws every role ink through one shared flood (104 render passes a frame for 600 role holders in nine inks before, 22 now), and styleKit 1.6.4 rebuilds the role mask at most four times a second while roles keep changing. So the whole look stays under a millisecond on the median frame with 600 role holders, and well under one (about 0.6 ms) with the 150 or so a typical level gives roles to. What is left is mostly styleKit queuing its eight mask passes and their selection lists again every frame (#24548): with hundreds of role holders and a slow client the average frame can pass 1 ms. The GPU time of the look did not change measurably.
### Quality
The stack follows the device's quality tier (`renderer.qualityTier()`), and the tiers really change the cost:
- `high` and `medium`: every block pass and every shared input at full resolution. `medium` also applies each block's `quality.medium` settings where it states them.
- `low`: the shared inputs (surface data, ground, wobble) and the block chain itself run at half resolution, a quarter of the pixels for every pass, and blocks take their `quality.low` settings (fewer samples, fewer strokes). The copy of base colour and normals is made on `low` too, at that half resolution (since 1.1.4), so a block judges light and lays marks by the same rules on every tier and a look stays the same look; only the forward render path, which keeps no g-buffer, goes without it. The result is scaled up once at the end, before the UI; excluded objects are put back from the full-size frame, so they stay crisp. Measured on the Lookbook's textured wall at 2097 by 1409 with a wash, a sketch line, paper, boil and dots: 1.64 ms of block passes at full resolution, 0.30 ms on `low` (before the half-size copy, which adds one small compute pass).
On `low`, marks are drawn at half resolution and so are softer: the honest trade for running on a small device. `StyleKit.setQuality("low")` forces a tier on this machine; `StyleKit.setQuality(nil)` follows the device again. The Lookbook measures the cost per pass at each tier.
### Whole frame or inside a material
Every block works on the whole frame with no material edits. Where it means something, a block also ships a shader module a material includes for a chosen set of objects (`include:` in its `block.yaml`), for example starterBlocks' Stepped light, whose `bandedToon` shader bands a material by the same rule. Use a group mask to keep the full-frame pass off those objects.
## API
| Call | What it does |
|---|---|
| `StyleKit.blocks(filter?)` | Every block installed: `{ id, name, slot, tags, summary, reads, writes, settings, package, path }`. Filter by `slot`, `package`, `tag` (one tag or a list, all must match) or `text` (which also searches tags). In the shell: `zero styleKit list`, `zero styleKit list line`, `zero styleKit list --tag pencil`. |
| `StyleKit.block(id)` / `settingsOf(id)` | One block / its settings with docs, defaults (the block's own `defaults:` where it states them) and ranges. |
| `StyleKit.add(id, opts?)` | Add a block (`strength`, `mask`, `settings`, `id`, `index`). Returns `ok, entryId`. |
| `StyleKit.remove(entry)` | Remove an entry (by entry id, block id or position). |
| `StyleKit.swap(entry, id, opts?)` | Swap an entry's block for another of its slot; strength and mask stay. The entry remembers the settings of every block it held, so a block swapped back in gets exactly the settings it had (a block it never held starts from its defaults). `{ keepSettings = true }` carries the current settings across instead; `{ anySlot = true }` allows a swap across slots. |
| `StyleKit.move(entry, index)` | Reorder (orders blocks inside a slot). |
| `StyleKit.setStrength(entry, 0..1)` / `setOn(entry, bool)` | Weight / switch an entry. |
| `StyleKit.setMask(entry, mask?)` | Limit an entry by depth, region or group. |
| `StyleKit.set(entry, { name = value })` | Change settings. Colours take `"#rrggbb"`, `{ r, g, b }` or `{ r, g, b, a }`. |
| `StyleKit.load(entries, opts?)` / `clear()` | Replace the stack's blocks / empty the stack. `opts`: `holdRate`, `markScale`, `pigments`, `softRadius`, `look`; an option left out keeps the stack's current value, so to reproduce a look exactly pass all of its options or use `applyLook`. |
| `StyleKit.options({ holdRate, markScale, exclude, groups, active, pigments, softRadius })` | Stack-wide options. `pigments` is a pigment set id (`""` for none); `softRadius` is the soft field's width in reference pixels (2 to 40, default 10). |
| `StyleKit.pigmentSets(filter?)` / `pigmentSet(id)` | Every pigment set installed: `{ id, name, summary, tags, pigments = { { name, colour, granulation } }, shadow, light, air, package, path }`, filtered by `package`, `tag` or `text` / one set by id. |
| `StyleKit.setPigments(id)` | Pick the pigment set every block that reads pigments mixes from, for the whole stack and every player. `""` picks none, so blocks keep the scene's own colours. |
| `StyleKit.looks(filter?)` / `look(id)` | Every whole look installed (`styleLook` assets): `{ id, name, summary, tags, entries, pigments, holdRate, markScale, package, path }` / one look by id. |
| `StyleKit.applyLook(id, opts?)` | Apply a whole look in one change: its blocks replace the stack and its pigment set, hold rate, mark scale and soft radius are set with them. `{ keep = { "line" } }` keeps the entries of the slots named from the current stack. A look naming a block that is not installed is refused and nothing changes. |
| `StyleKit.describe()` | What is in the stack, in run order, its `pigments`, `look` and `softRadius`, and what the runtime drew last frame. |
| `StyleKit.checkBlock(id)` | Problems with a block against the block rules (empty when it passes). |
| `StyleKit.newBlock(name, { slot, folder })` | Write a working block of your own to start from. |
| `StyleKit.quality()` / `setQuality(tier?)` | The tier the stack draws at / force one here. |
| `StyleKit.setLocal(entry, settings?)` / `localSettings()` | Lay settings over one entry on this machine only, over its synced settings (a tint or a ramp for the level this player is on); nothing is synced or saved, `nil` takes them off / list what is laid here. |
| `StyleKit.setLocalEntry(id, entry?)` / `localEntries()` | Lay a whole entry on this machine only (`{ block, strength, on, mask, settings, variant }`), after the synced stack and ordered by slot like any other: for short local effects such as an impact frame on a hit, which a game keeps in step across players with a message of its own. Never synced or saved; works with no StyleStack in the scene; `nil` removes it. (Contributed by another account as pull request #10, ported onto 1.7.0.) |
| `StyleKit.publishInput(name, texture?)` / `inputs()` / `inputWanted(name)` | Publish a texture another package computes (a heat map, a sonar wave, a thickness map) on this machine for blocks that list `input:<name>` under reads and declare an `in_<name>` texture; `nil` withdraws it (blocks read black). `inputWanted` says whether a live block reads it now, so the producer works only then. Never synced or saved (1.8.0). |
| `StyleKit.setClock(fn?)` | Drive the stack's one clock (boil, grain, wobble) from your own time in seconds, for a cutscene, a recording or the Lookbook; `nil` returns to the shared clock (wall-clock UTC, the same held step on every player's machine). |
| `StyleKit.give(entity, role?)` | Give one object (and everything under it) a game role, for every player and saved with the scene; `nil` takes it away. |
| `StyleKit.setChannel(entity, fields)` / `channelOf(entity)` / `clearChannel(entity)` | Set how the look treats one object: `{ role, variant, line, marks = { line, tone, fill, paper }, value, ink, children }` (-1 or "" gives a field back to the role) / what it gets now, read through the scene's role set, and which object's channel reaches it / take its channel away. |
| `StyleKit.channels()` | Every object a channel reaches with what it gets, and a revision that moves when any of it changes (for runtimes that draw per object). |
| `StyleKit.setVariant(entry, variant?)` | Share a slot by group: the entry draws only on objects of that variant, and the entry of its slot naming none leaves them to it. `add` takes `variant` too. |
| `StyleKit.setRoles(id)` / `roles()` / `roleSets(filter?)` / `roleSet(id)` | Pick the role set for the whole scene and every player ("" follows the look, "none" reads no roles) / the set in use / every installed set / one set, with what it extends laid under it. |
| `StyleKit.lookRoles()` | The set the scene's look suits by itself: what a set written with `extends: look` (a set for a kind of game) is laid over. |
Writes go through the stack's owner, so in play any player can change the shared stack and every player sees the change. The `styleKit` toolbox offers the same from the shell (`zero styleKit list`, `zero styleKit stack`, `zero styleKit add posterize`).
## Writing a block
A block is a plain folder holding `<name>.styleBlock/` (block.yaml, README.md, .metadata) and its `<name>.shader/` (a post-process shader whose `properties.yaml` starts with the styleKit shared header). `newBlock` writes one that works and registers it at once:
```luau
local StyleKit = require("~styleKit.StyleKit")
local path = StyleKit.newBlock("warmTint", { slot = "medium", title = "Warm tint" })
print(path) -- /zero/source/myBlocks/warmTint/warmTint.styleBlock
StyleKit.add("warmTint", { strength = 0.6 }) -- a warm filter over the whole look
StyleKit.set("warmTint1", { tint = "#ffcf8a" }) -- its setting, like any block's
```
Then edit `/zero/source/myBlocks/warmTint/warmTint.shader/shader.wgsl` to make it do what you want; the stack picks up the change as you save. To share blocks, move their folders into the `blocks/` folder of a package of your own (a `<name>.package` folder with its `package.yaml`): a package that holds blocks adds them to every world that installs it, with no code. The guide "Writing your own style block" walks the whole shape every block follows.
A block that shades the picture (hatching, stipple, dots, a pencil shade) belongs in the Tone slot, where it is kept from brightening a dim room by the dim guard with no work of its own. A block in another slot that lays a light sheet or light marks over the picture can ask for the same with `keepDim: true` in its `block.yaml`; a Tone block that should be left alone in the dark says `keepDim: false`.
## Rules every block follows
- It states its slot, what it reads and what it writes, and takes over nothing it does not own (a Ground block alone writes the ground, a Time block alone the wobble).
- Every setting is named, documented in `block.yaml` and has a good default in its shader.
- Marks stay on things unless the medium really is a screen (pixels, grain, a vignette), and they run after antialiasing and upscaling.
- Wobble runs on the held clock.
- Its README says what it gives up in motion and how it treats opaque things, transparent things, particles, sky, fog, skinned characters, foliage and world text.
- With every light off the frame stays black (`sk_finish` holds every block to it), and a dim room stays dim (the dim guard holds every block that keeps dim to it).
- The UI is drawn after the stack, so it stays readable.
- No full-screen flashing above 3 Hz and no fine high-contrast stripes filling the view.
`StyleKit.checkBlock` and the `styleKit` test suite check the rules that can be read from files.
## Guides and workflow
This package carries its guides (search them with the `guides` tool or ZeroMind): combining blocks into your own art style, how to use styleKit, writing your own style block, and a worked example from an empty world to a small stylized game. The `restyleWorld` workflow gives an existing world, scene or kit a new look non-destructively, and the `styleKit` agent skill tells an agent what it can do with blocks.
## What a look costs
GPU milliseconds per frame of the style passes, 3D view drawing at 1893 x 1072 (98% of a 1920 x 1080 frame's pixels), measured with the Lookbook (`zero lookbook cost 1920 1080 true`) over its pan on a desktop GPU. The rest of the frame (scene, shadows, antialiasing) is not counted. Low quality runs the whole chain at half resolution.
| Look | High | Low |
|---|---|---|
| README example: posterize, duotone 0.5, depthLine, sketchLine 0.6 | 0.60 | 0.11 |
| Cookbook: flatWash, depthLine, plainPaper, lineBoil, softGrain 0.6 | 1.03 | 0.21 |
| Comic: lightBands, surfaceDots, depthLine, posterize | 0.41 | 0.10 |
| Chalk: depthLine, darkBoard | 0.43 | 0.10 |
| Retro: posterize, chunkyPixels, softGrain, vignette | 0.22 | 0.07 |
Each block alone, high / low: flatWash 0.45 / 0.07, sketchLine 0.32 / 0.06, darkBoard 0.24 / 0.07, plainPaper 0.23 / 0.07, depthLine 0.19 / 0.04, cardSurface 0.11 / 0.03, lightBands 0.09 / 0.03, surfaceDots 0.09 / 0.03, softGrain 0.09 / 0.03, lineBoil 0.06 / 0.02, chunkyPixels 0.06 / 0.02, posterize 0.05 / 0.02, duotone 0.04 / 0.02, vignette 0.04 / 0.02. A look costs a little less than its blocks added up, because the shared inputs (depth, surface coordinates, ground) are made once per frame for the whole stack.
**Frame rate and the web.** On the web the Cookbook look was measured at about 29 to 30 fps in the Cookbook scene (one sample 25), against about 43 fps with no look, and the low tier did not change it: the look cost about 10 ms a frame there while its passes took about 1 ms of GPU time. That measurement is of styleKit 1.0.6. The cost was not the shading but the work around it, which a browser pays several times over: 15 screen-sized images made up front whatever the look, 11 passes, and engine calls every frame for every effect. 1.0.7 cuts it, measured on desktop: the Cookbook look now makes 7 images on high and 5 on low (was 15), runs 9 passes on high and 8 on low (was 11), and spends 0.19 ms of Luau a frame on high and 0.12 ms on low (was 0.40 ms on both), so low now saves passes and Luau, not only pixels. **Whether 1.0.7 holds 30 fps on the web is not measured yet:** an agent cannot start a browser engine (#20946), and in the desktop editor the whole-frame rate swings more from the editor's own load than the look costs, so no desktop fps is quoted here. `zero lookbook report` now starts with the frame rate in the scene you are in (no look, the look on high, on low, measured in turns), so one run in the web build gives the answer. The web limits it may still meet are open engine issues: #20750 (web frame rate), #17441 (cost of scene-colour features) and #17478 (post tiers on mobile).
## How it runs
The `styleStack` render feature starts on a machine as soon as the scene's stack holds a live block (one that is on, has strength above 0 and is installed), on every player's machine alike. Each frame it copies the base colour and normal from the deferred g-buffer at `afterLighting`, then at `afterScenePost` (after antialiasing and the render-scale upscale, before the UI) it rebuilds world position and distance from depth, publishes the ground and wobble (only when a block of the look publishes them; otherwise blocks read a tiny constant image written every few seconds), runs each slot and hands the result to the screen, blending it over the frame as rendered so excluded objects stay as they were (one pass, no copy of the frame). When a live block keeps dim, the dim guard's two small passes (a meter at an eighth of the chain's resolution, then the guard) run between the last slot and the hand-off. A frame of the Cookbook look runs 9 passes: the copy of base colour and normals (only when a live block reads them, and never on the low tier), world position, the two shared fields, one pass per block, and the hand-off to the screen; the blocks read the scene directly, so there is no copy of the frame first. Only five screen-sized images are made up front (the chain's two, world position and the two fields); the rest are made the first time a look uses them (group masks, excluded objects, the base colour and normal copies). A frame in which nothing the stack draws from has moved (the stack, its options, the tier, the images, the camera anchor, the view, the channels) hands the engine the passes of the frame before as they were and skips the rest of its work; a held clock step writes only the clock. Otherwise the Luau work per frame is a few calls into the engine: uniforms, settings and mask groups are written only when the stack, the tier or the camera's frame inputs change, and the clock only when the held step moves. When the stack holds no live block (after `StyleKit.clear()`, or with every entry off) the kit lets everything of its own go within two frames: every block effect and its internal effects (the dim guard's two among them) are removed from the screen effect list, its screen targets, buffer and mask materials are freed, and the render feature itself ends. `StyleKit.holdsAnything()` then reads false, and the first block added brings it all back.
## Known limits
- **Offscreen captures and render-to-texture cameras do not draw the stack.** The stack runs at `afterScenePost`, which only the presented frame reaches, so `capture` with `source = "main"`, `"entity"` or `"position"` shows the unstyled scene; use `source = "presented"` or `renderer.presentedFrames`. Per-camera post: engine issues #66 and #17427.
- **Depth in a feature's fragment pass reads 0 after post.** `zero_scene_depth` is not bound to a render feature's fragment pass at `afterScenePost`, so the kit copies depth with a compute pass once per frame and blocks read that copy (#17427).
- **G-buffer channels are refused after post**, so base colour and normal are copied at `afterLighting`. On the forward path there is no base colour: Surface and Light blocks fall back to working from brightness alone.
- **Marks on moving objects.** Surface coordinates come from world position, so marks stay put on still things under any camera move, but slide across a moving skinned character or a moving prop. Object-space coordinates need per-object data a full-frame pass does not have, so marks that stay on a moving object can only be drawn inside its material: watercolorBlocks' `watercolorSurface` shader shows how, carrying the object's own position from `vertex_clip` (`local_position`) to the fragment in the vertex colour (applied with `WatercolorHero.paint`).
- **The soft field costs two half-resolution passes each time a reader reads a new image.** In a chain, a later block that reads `soft` reads a different picture from the first one, so the field is made again for it; keep soft readers few, or next to each other on the same input, on small devices.
- **Post effects get one pass each and no history** (#17427): a block cannot read the previous frame or its own last output, so nothing in a block can build up or settle over time.
- **Effect uniforms reset on recompile** without notice; the runtime renews its writes twice a second, so after a shader edit a setting can take up to half a second to come back.
- **One held clock for every player, as close as their clocks are set.** Since 1.1.3 the held clock runs on wall-clock UTC (caught to the frame from `os.time()` ticking over), so every machine shows the same boil step at the same moment; machines whose clocks are set by the network agree to a few hundredths of a second, so two screens differ only in that sliver around each step. A machine whose clock is set wrong boils out of step with the others. `StyleKit.setClock` drives it from a time of your choosing instead, so a cutscene or the Lookbook repeats exactly.
- **Transparent things and particles have no depth**, so lines and surface marks treat them as the opaque surface behind them; each block's README says how.
- **Stage order** (#17426, #17439): blocks rely on `afterScenePost` running after the TAA resolve and the upscale, as the engine documents today.
- **A full-screen pass cannot tell which object a pixel belongs to (#22178), and a geometry pass carries no per-object value (#24549).** Style channels are drawn as at most eight distinct settings, one geometry pass and one small image each, packed by one more pass into the object mask. Objects asking for the same thing share a setting, so a game's sixteen roles fit; a ninth distinct setting is left out (the objects it reaches draw as if they had no channel), and `StyleKit._status().parts` says how many were drawn. Colours per object (an enemy's ink) cannot ride the mask; blocks take them through a variant entry of their own (styleRoles adds one per coloured variant), and toonBlocks' ink, which draws per object, takes the channel's `ink` directly.
- **Every pass is queued again every frame (#24548).** Each mask setting's selection list costs a little script time per selected object every frame; with 600 role holders that is most of the 0.2 to 0.3 ms the channels add.
- **A pass reads no depth unless it also writes the scene picture (#24550)**, so the mask parts test against the scene's depth in their own shader, and a line block finds the object a line belongs to by looking at the nearest surface around it (`sk_chan_line_near`) rather than at a depth the mask pass kept.
- **An unknown blend draws as replace (#24551)**: the mask parts use `opaque` (each part is an image of its own), and nothing in the kit relies on another blend name.
- **A running editor that had styleKit 1.4 open keeps the StyleStack's old fields until it restarts.** It cannot read or write the stack's new `roles` field, so the role set lives on a `StyleRoleSet` component beside the stack (new in 1.5.1), which every client has at once; objects' channels (a component new in 1.5.0) replicate at once too.
- **Edit mode does not share live changes between two editors.** Roles given in one editor reach another through the saved scene; in play every change reaches every player at once.
- **One stack per scene.** With two StyleStack entities in a scene the kit draws the last one it found; keep one, and change it rather than adding a second.
- **After a scene switch** the engine can free the stack's screen targets; the runtime notices within about two seconds and rebuilds them, so the look can be missing for that moment.
- **Switching looks can show picture memory a little above where it began.** Two seconds after the last live block goes, the kit frees every screen image it made (checked with 1.4.4: four looks applied and cleared, five rounds). The engine keeps freed render scratch for reuse until its next collection, so between collections the texture memory reads one or a few screen images above the start (112, 136, 160, then back to 88 MB in one session); `renderer.collect()` brings it back to the same figure every round (64 MB there). It does not grow without end.
- **The dim guard keeps a dim room's level, not every pixel's.** It matches brightness smoothed over about 36 pixels and holds each pixel to a fifth above the render around it, so one light mark can still sit a little lighter than the render right under it while the area around it stays as dark as rendered. In a lit picture (brightest parts at 0.7 or above and a mean of 0.4 or above) it does nothing, so how a block treats daylight is up to the block.
- Post-process shaders have no preview image, so the shader previews in this package are absent.
- **Other engine issues that apply here:** the camera gives no near and far planes (#24514); a two-client editor sometimes refuses to load two different scenes at once (#24666, #24667); a world boot can wait about 30 s on a network handshake before content starts (#24669); installs with several dependencies are slow (#21420); the ZeroMind description can lag one release (#19258); the web build was not checked from here (#20946); two engines on one account collide as one player, so a two-player check needs two accounts (#14448).
## Version history
**1.8.3 (2026-11-13)**: a block added with no `rank` lands before the first block of its slot ranked 100 or more, so a finishing block (a printed margin, a frame, weather laid over the picture) that states `rank: 100` and up with `stacking: chain` stays on top when grain or a vignette is added after it. Before, the later block was appended last and, as a plain layer reading the slot's input, painted the margin out. Blocks with a rank, and every stack already saved, are placed as before.
**1.8.2 (2026-11-13)**: the sky distance a depth mask uses costs the render path nothing. The active camera, its entity and its Camera component were read from inside the render feature's frame (about twice a second: forced on the replay check and again on the full path), where those engine calls can cost far more than from a script: a profile of a 2-3 block look in a large world (about 2,900 entities) put about 24 ms of every second there. The render path now reads only the cached far distance; the lookup runs in a deferred task off the frame, at most every 2 seconds, and a replayed frame takes the full path as soon as the cached value differs from the one it was built with. A camera far-plane change is picked up within about 2 seconds; the frame drawn is unchanged.
**1.8.1 (2026-11-13)**: a still frame costs less script time. (1) A frame in which nothing moved now replays every frame instead of taking the full path every 15th frame: the periodic checks run on cadences of their own, spread over the frames, and send a frame down the full path only when one finds something moved. The renewal backstop (one effect every 30 frames) and the dim guard's rewrite are a few property writes on a replayed frame; each effect's shader-reload count is read in turn (about every 30 frames each); the lighting watch every 15 frames (a change takes the full path and reads the sun again); the sky distance every 60 frames; the neutral fields' backstop write keeps its 240-frame cadence. Target residency, the quality tier and a rebuilt device are read before the replay check as before, and a new effect settling, a driven clock, objects still awaited, a tier change, lost targets and any stack, option, view or channel change still take the full path. (2) A replayed frame allocates nothing: the live-entry list is kept per stack (ordered list, revision and registry), and the anchor and clock are made only when used. (3) The stack's 60-frame backstop copy of an unchanged stack replays like the list it copies (compared by content) instead of forcing a full frame. Measured on the Cookbook's main scene (5 entries, 9 passes, static camera, 12 s): forced full frames went from 58 to 0, mean styleKit render time from 0.20 to 0.17 ms a frame in a headless client; what is left on a replayed frame is almost all the engine's own per-pass enqueue (about 4 µs warm per pass). The frame drawn is unchanged.
**1.8.0 (2026-11-13)**: published inputs. A block can now read a texture another package computes on each machine: the producer calls `StyleKit.publishInput(name, texture)` and every block that lists `input:<name>` under `reads` gets it bound to its `in_<name>` texture (black while nobody publishes it); `StyleKit.inputs()` lists what is published and `StyleKit.inputWanted(name)` says whether a live block (synced stack or local entry) reads it now, so the producer only works while one does. `checkBlock` pairs each `input:` read with its `in_` texture and does not ask for a description or default for `in_` textures. Inputs are never synced or saved. Asked for by sensor-look blocks (thermal, sonar, x-ray) that read a heat map, a sonar wave and a thickness map made by other packages; before this a block could only read styleKit's own fields. The test suite gains "published inputs reach the blocks that read them"; the writing-a-style-block guide and the styleBlock type README describe it. Nothing changes for existing blocks.
**1.7.9 (2026-11-12)**: a look costs the engine less per frame. Measured with the engine's own CPU and GPU time (look off against on, 1920 x 1080, 2,500 things) on animeBlocks' one-pass TV anime look, which went from 10 passes to 5 and from about 0.33 to 0.19 ms of styleKit script a frame. (1) A frame in which nothing the stack draws from has moved (the stack, its local entries and options, the tier, the images, the camera's anchor, the view, the channels) hands the engine the previous frame's passes again and skips the rest of its work; a held clock step writes only `sk_clock`; every 15th frame, a frame with a new effect still settling, a driven clock and objects still awaited take the full path, so every periodic check keeps running. (2) The ground and wobble a look reads when no block of it publishes them are tiny constant images written every few seconds instead of two full-screen passes every frame. (3) Excluded objects: the hand-off blends the styled result over the screen, which still holds the frame as rendered, transparent where an object is kept, so there is no copy of the original frame and no copy before the hand-off (three passes become one; kept objects stay full size and crisp on every tier). (4) When the object mask only keeps objects as rendered (no mask group, no channel), the restore reads the exclusion's own part and the merge pass is skipped; when the merge does run it reads only the parts in use. (5) The stack's own effects are registered the first time a look uses them (the engine surveys every registered effect each frame): the TV anime look holds 4 instead of 10. Checked: excluded objects stay exactly as rendered (a pedestal and a ball excluded on animeBlocks' plaza), and the TV anime look looks as before.
**1.7.8 (2026-11-12)**: quieter on a scene saved before styleKit 1.5. Its StyleStack has no `roles` field (the role set lives on the `StyleRoleSet` component beside it), and setting a role set tried to write the old field as well, logging a refusal each time; once a read has found the field missing, the write is skipped. Found following the Summit Sprint example guide on a fresh fork. Nothing else changed.
**1.7.7 (2026-11-12)**: a role set or look that another player changes (or a pull brings) now reaches this machine at once. The asset type's change hook runs only on the machine that made the change, so every other player kept the old roles and looks until they reloaded; each registered role set and look is now watched, and a change arriving from elsewhere is read again. Checked with two clients: a role set edited in one was read in the other within the same session, with no restart. Nothing else changed.
**1.7.6 (2026-11-12)**: a style channel reports `inkByHand` when its ink was set on the object by hand rather than by its role, so runtimes can draw it as set (toonBlocks 1.4.0 mixes none of the object's own colour into it). Role sets take `noLook:`, the set one written with `extends: look` sits over while no look is on: styleRoles' sets for kinds of game name the toon game set, so a game's hero, enemies and pickups carry their groups (player, enemy, pickup, and with them the variant lines and glazes) from the start rather than only once a look is picked; before, with no look on, a game set gave its hero the variant `hero`, which no line or glaze targets. Nothing else changed.
**1.7.5 (2026-11-12)**: styleKit's checks pass in a world that holds only styleKit (or only styleRoles, which brings it). "Clearing puts the frame back as rendered" loaded starterBlocks' `plainPaper`, so where starterBlocks is not installed the load was refused, nothing drew and the check failed (since 1.7.4 a plain styleKit install no longer brings starterBlocks along); it now stacks styleKit's own `posterize`, `duotone` and `depthLine`. "Block defaults stand over the shader's" skipped itself unless another package's block stated defaults (paperBlocks); it now writes a throwaway block of its own whose block.yaml states `amount: 0.37` over the template shader's 0.5, checks it, and removes it, then checks every other installed block that states defaults. In a fresh world with only styleRoles installed the suite passed 13 of 13 twice, with nothing skipped (it was 11 passed, 1 failed, 1 skipped). No block, shader or runtime changed.
**1.7.4 (2026-11-12)**: styleKit (and everything that brings it, styleRoles first) installs into an empty world again. The styleKit agent skill listed the Lookbook (`lookbookKit.Lookbook`, `lookbookKit.lookbook`) among its dependencies, so every install of styleKit brought lookbookKit along under deps, and lookbookKit's own pin on starterBlocks could not be met there: a fresh world that installed only styleRoles refused it ("is not installed, because it does not run"). The Lookbook stays optional, as the skill and the restyle workflow already said (they use it only where a world has it); it is no longer a dependency. Nothing else changed.
**1.7.3 (2026-11-12)**: fixes on top of 1.7.2, which came in as pull request #12 (another account's, merged). The stand-in search 1.7.2 added ran one name search over the whole scene per object with a channel on every sweep, and every second for 15 s after roles arrive: 0.023 ms each in a 158-thing scene, so 14 ms a sweep with 600 roles, growing with the scene. It now builds one index of stand-ins per prefix with a single glob search (`pk_vis_*`), kept until anything spawns or despawns: 0.0035 ms. And `look(id)` / `pigmentSet(id)` scan again on a miss at most once every 2 s: the role set picker and the roles test ask for looks that are not installed on purpose, and each miss read every look and pigment set the world holds. Checked here: 40 coin stand-ins and their child meshes all take the pickup role.
**1.7.2 (2026-11-12)**: a role reaches pickupsKit coins (from the first second of play), and a look installed after the first scan is found.
- A style channel now also reaches an object's **stand-ins**: root entities named `<prefix><owner id>` that a kit draws in the object's place. pickupsKit pools every pickup's model as its own root, `pk_vis_<pickup id>`, so `StyleRoles.give(coin, "pickup")` (the guide's own example) reached nothing the coin draws: on a pencil level the coins stayed grey, with no gold glaze and no ochre line, while the hero and enemies took theirs. `pk_vis_` is registered by default; `StyleChannels.addStandInPrefix(prefix)` registers another kit's. For 15 seconds after channel owners arrive (a scene load, the start of play), subtrees are swept every second instead of every 10, since pickupsKit builds its visuals a frame or more after the pickups; without that, coins stayed grey for the first 10 seconds of play. On a 16-coin, 2-gem level, 52 of 52 coin and gem models took their role (0 before), and ReadApart's least distance read 19.3 with roles on and 2.5 with roles off.
- `look(id)` and `pigmentSet(id)` scan again on a miss. Looks were scanned once a session and then kept up by per-file registration, so a block package installed into its own folder after another install had first placed it under `deps/` (inkBlocks, brought by pencilBlocks, then installed itself) lost its looks in the move: `applyLook("comicInk")` answered "no look called comicInk" and the RoleSetSwitcher's Ink comic button switched the roles alone, until the editor restarted.
**1.7.0 (2026-11-12)**: groups that share a material read apart under every look, untagged things keep the plain line, and a look switched on never flashes black.
- New block **Roles glaze** (`roleGlaze`, Colour): a glaze of one colour over one variant's objects, keeping each surface's own value (a white surface takes a light tint of it, never paper white). Role sets gain `glazes:` beside `lines:` (merged through `extends` the same way), and styleRoles 1.3.0 adds one glaze per group. On a plain stage where one player, five enemies and five pickups share one white material, the least distance between the groups with roles on went from 4.0 to 18.9 under Picture book, 4.2 to 22.6 under Watercolor, 10.7 to 26.5 under Pencil sketch, 10.1 to 17.4 under Ink drawing, 11.5 to 19.4 under Comic ink and 20.2 to 43.0 under Paper cutout game (styleRoles' README has the table).
- `describe()` reports each entry's `variant`. It did not, and toonBlocks' ink reads its entries from there, so under a toon look every variant line a role set adds was drawn on every object, things with no role included, and the last one won: lamp posts nobody tagged came out in the arena shooter's hazard orange, and in enemy red once the hazard line was removed. Untagged things now keep the look's own line.
- A look's passes run only once every shader they use has compiled; until then the frame is shown as rendered (at most 4 s per new effect). A pass whose shader has not compiled draws nothing, and the next pass then read an image that was still black, so in a fresh engine, where a look's first compile takes seconds, switching to a look could flash the screen black.
- The dim guard lets an object whose style channel asks for a lighter line weight (under 1) lighten in proportion, so a light pencil on one object reads lighter than the rendered frame even in a dim picture (pencilBlocks' per-object weight test: weight 0.3 read 245 against 204 with no channel, where before the guard held it at the rendered 90).
- `StyleKit.setLocalEntry(id, entry?)` and `localEntries()`: a whole entry on this machine only, from pull request #10 (another account's, made on 1.5.1 with seven conflicts; ported by hand, the rest of that pull request was older copies of files this world has since moved on).
**1.6.5 (2026-11-11)**: A new effect's settings, frame values and sun are written every frame for its first three seconds. A value written before an effect's shader had compiled read as the shader's default (#22741) and was written again only by the every-30-frames backstop, one effect at a time, so a look switched on from nothing (a stack with a paper most of all) drew a lighter, half-set picture for two to four seconds and then snapped to the real one. It now draws the settled picture from its first frames (measured: FLIP 0.049 from the settled frame for the first 2.5 to 3.5 s before, 0.000 from the first sample after). Found by running every suite twice in a freshly installed world: paperBlocks' unpaint test and pencilBlocks' shadow-band test read the picture inside that window.
**1.6.4 (2026-11-11)**: the role mask is rebuilt at most four times a second while channels keep changing (enemies spawned and killed every frame in a busy level), where it was rebuilt on every frame something changed; a read for one object (`channelOf`) always sees the latest. With toonBlocks 1.3.6 this takes the look with 600 role holders from 1.33 to 0.93 ms on the median frame (table under Style channels). Every README in the package (each block, look, shader and guide) now ends with a Known limits section naming the engine issues that apply there. It also carries what was prepared as 1.6.3: a role or channel given to a player template (an entity holding a `PlayerPrototype`, such as the `Player` of the engine's player scene) now goes on the template's body, which is what the engine copies into the world for each player who joins. Before, `StyleRoles.give(entity.find("Player"), "hero")` reported success and read back as hero in edit, but the template itself is never in the world, so no actual player, late joiners included, carried the role. `setChannel`, `give`, `clearChannel` and `channelOf` all follow the template to its body. Found by installing the look into a new world and building a small scene from the engine's player template.
**1.6.2 (2026-11-11)**: a look, a block or a role set change works after a play restart. The kit held a handle to the scene's StyleStack across the restart; the old component still answered reads, so the kit kept using it, and every write failed with "the component this proxy names was removed". A write that fails that way now drops every held handle, finds the scene's stack again and runs once more. Found while checking styleRoles 1.1.0 on a fork of Summit Sprint.
**1.6.1 (2026-11-11)**: (released as 1.6.1 because the number 1.6.0 was already taken on ZeroMind) role sets for a kind of game. A set may write `extends: look` to sit over whichever set the scene's look suits (its line weights and inks over that set's, which keeps deciding tone, fill, paper and value), so one set works under every look and follows a look change by itself; variant lines merge the same way, variant by variant. `StyleKit.roleSet(id)` lays such a set over the current look's set, `StyleKit.lookRoles()` names that set, `StyleKit.EXTENDS_LOOK` is the word. New `ChannelProbe` module, a test helper that measures what one object's channel does to the presented frame (used by every block package's tests). A scene's script cost is unchanged (see What it costs).
**1.5.1 (2026-11-11)**: the scene's role set is kept on a `StyleRoleSet` component beside the StyleStack (and in the stack's `roles` field where a client has it). An editor that had styleKit 1.4 open when the update arrived keeps the stack's old fields until it restarts; there `StyleKit.setRoles` used to report success while the write failed, and reading the set logged an error every frame. Both work now. Found by installing the release into a fork of Summit Sprint.
**1.5.0 (2026-11-11)**: one look, and players, enemies and pickups read apart under it. A `StyleChannel` on any object (synced, saved, reaching everything under it) says which variant of the look it gets or `asRendered`, how heavy its line is, how much of the look's line, tone, fill and paper reaches it, one free value and an ink colour, or just names a game role the scene's role set fills in. A new asset type, `styleRoleSet`, holds role sets (with `extends:` for an author's own); `StyleKit.give`, `setChannel`, `channelOf`, `clearChannel`, `channels`, `setRoles`, `roles`, `roleSets`, `roleSet` and `setVariant`, and `variant` on `add`. The object mask (`sk_obj`) now packs exclusion, mask groups and each object's channel; every block is weighed by its kind of mark and its entry's variant inside `sk_weight`, so every installed block honours channels with no change; `sk_chan_*` helpers let a block read line weight and value; `depthLine` and `sketchLine` take the line weight. A scene with no channel draws exactly as before and makes no mask. Writes to another player's object go to its owner. `StyleKit.VERSION` says 1.5.0.
**1.4.24 (2026-11-10)**: less work every frame. The stack's entries are copied and put in order once per change (any property change on any machine moves the stack's revision) instead of every frame, with a fresh copy every 60 frames as a backstop; each effect's shader-reload counter (a name lookup) is read every 30 frames instead of every frame; the quality tier and the render targets' residency are checked every 15 frames (a rebuilt render device is still caught the frame it happens). Measured under toonBlocks' bench with 2,500 things, the style stack takes about 0.23 ms on the median frame on one shared line (0.33 ms with eight per-object ink styles, which was 0.33 ms with more variance on 1.4.23).
**1.4.23 (2026-11-10)**: nothing on a timer looks over the scene. The sun and the light rows are read only when a light or the sky changed (a native watch over the engine's Light and Sky components; before, the "is anything lit" check read the light rows once a second); a scene stack and the objects named in its groups are looked for again only when the entity set changed (before, twice a second while one was missing). The cached writes to every effect are renewed when the render device is
# paperBlocks
Put your game on paper: seventeen papers a stylized look is drawn on, from cold-press watercolor paper and thirsty rice paper to brown kraft, parchment, canvas, a dusty blackboard and black paper for white ink. Each paper is a styleKit Ground block, so it swaps for any other paper in one change, stays on the surfaces it covers while the camera moves, and lets the blocks above it feel it: washes granulate into its tooth and run further on a thirsty sheet, ink lines break on rough grain and feather on soft fibres, and on toned and dark sheets the picture's lights are put on in white gouache, chalk or white ink. Parts of one world can sit on different papers (a village on cold press, a map room on parchment), and every player sees the same paper.
Version 1.2.3. Built on styleKit 1.7.8 (style channels: how much paper shows through on each object).
## Install and use
Install **paperBlocks** from ZeroMind (search "paperBlocks" or "paper texture for a watercolor game"). It brings what its own files use, and nothing needs installing by hand: **styleKit** (the look stack its shaders build on), and **lookbookKit** with **starterBlocks** (the package's test suite measures at the Lookbook's every-light-off moment, and the Lookbook uses starterBlocks' blocks). It does not bring watercolorBlocks or inkBlocks: those are separate installs, and the papers work without them. Checked on 2026-11-01 by installing paperBlocks from ZeroMind into a brand new private world: 617 files in about 3 s, and the example below ran there line by line exactly as written (the frame changed on load, kraft read clearly differently, swapping back brought cold press back with its deeper tooth, and `clear()` returned the frame to exactly its unstyled brightness). Installs with several dependencies can be slow (#21420): give the install time to finish and check `/zero/source` for the `paperBlocks.package` folder.
Then, in a fresh world with only paperBlocks installed, from the console or any script:
```luau
local StyleKit = require("~styleKit.StyleKit")
StyleKit.load({
{ id = "line", block = "depthLine" }, -- styleKit's own clean ink outline
{ id = "paper", block = "paperColdPress" }, -- the picture laid on cold-press paper
})
StyleKit.set("paper", { toothHeight = 1 }) -- a deeper tooth: one change
StyleKit.swap("paper", "paperKraft") -- the same picture on brown kraft, lights in white gouache
StyleKit.swap("paper", "paperColdPress") -- back on cold press, with the deeper tooth it had
StyleKit.clear() -- the frame exactly as rendered again
```
Swapping back puts back exactly the settings the paper had (styleKit 1.2.0 remembers them on the entry), so the frame returns as it was. The stack lives on the scene's synced StyleStack, so every player sees the same paper, and it is saved with the scene.
To see what is installed: `StyleKit.blocks({ package = "paperBlocks" })`, or in the shell `zero styleKit list ground` and `zero styleKit list --tag paper`.
## The papers
Every paper is the same paper model with its own defaults, so a setting means the same thing on every paper and a paper you tuned on one block carries its meaning to another. "Lights" says what the paper needs from the picture's light: on a white sheet the lights are bare paper; on a toned or dark sheet they are put on in an opaque light, which only shows where the scene really is lit.
| Paper | Block | Character | Best for | What lights need |
|---|---|---|---|---|
| Cold press | `paperColdPress` | Warm white, a rough felt tooth, the standard thirst the watercolor and ink blocks were tuned on. | Watercolor, line and wash, picture books: the default sheet. | Nothing special: lights are bare paper. |
| Rough | `paperRough` | The deepest, largest tooth, warm white, a little thirstier. | Loose washes full of granulation, broken dry brush, scratchy lines. | A low sun across surfaces shows the tooth most. |
| Hot press | `paperHotPress` | Smooth, almost no tooth, crisp white, hard-sized. | Crisp washes with hard edges and dark dried rims, exact ink lines, botanical and technical illustration. | Nothing special. |
| Rice paper | `paperRice` | Thin and soft with long visible fibres, very thirsty. | Ink wash painting, soft washes that feather, lines that bleed along the fibres. | Nothing special. |
| Newsprint | `paperNewsprint` | Grey-cream, a little yellowed, a fine flat grain and darker fibres, thirsty. | Zines, cheap comics, fuzzy print-like ink. | The sheet's grey-cream is the lightest light. |
| Kraft | `paperKraft` | Brown, fibrous, a mid-tone ground. | Toned sketches, crafts and cardboard, gouache highlights. | A clear lit side: lights are put on in white gouache only where the picture is brighter than about halfway from the brown to white. A flat, dim scene stays brown. |
| Parchment | `paperParchment` | Warm cream, strongly blotchy, a few age spots, a soft sheen, fairly hard-sized. | Old maps, documents, scrolls, a map room. | The sheen shows only where light reaches a surface, most under raking light. |
| Canvas | `paperCanvas` | Primed canvas: a plain weave of threads under everything. | Painterly looks; a weave showing through paint and line. | Raking light shows the weave. |
| Illustration board | `paperBoard` | Hard, smooth, bright and even. | Clean comic ink, controlled washes, graphic looks. | Nothing special. |
| Blackboard | `paperBlackboard` | Dark slate with sharp grit, chalk dust and the ghosts of old wiping, very hard-sized, dead matte. | Chalk drawings, classrooms, puzzle rooms, night sketches. | The picture is made of chalk: only lit things are drawn. A dark scene is bare board. |
| Green chalkboard | `paperGreenChalkboard` | A green school board, dustier than the slate. | The same as the blackboard, friendlier and more schoolroom. | As the blackboard. |
| Black paper | `paperBlack` | Near-black with a soft tooth. | White ink drawings, silhouettes, night. | The picture is drawn in white ink: only lit things are drawn. |
| Aged sketchbook | `paperSketchbook` | Yellowed, mottled, foxed with brown age spots that show through the paint. | Travel journals, old sketchbooks, memories. | Nothing special: lights are the yellowed sheet. |
| Toned grey | `paperTonedGrey` | A mid-grey drawing paper. | Toned drawings: darks laid in, highlights in white chalk and gouache. | Lights are put on in white from about 40% of the way from the grey to white; the values between stay bare grey, the middle tone. |
| Toned tan | `paperTonedTan` | A warm tan drawing paper. | Warm toned drawings with white highlights. | As toned grey. |
| Ruled notebook | `paperNotebook` | Smooth bright paper with pale blue rules, a little thirsty. | Doodles, school-book and diary looks. | Nothing special; the rules show through every wash. |
| Vellum | `paperVellum` | Cool cream, a cloudy formation, a satin sheen, very hard-sized. | Drafting and technical looks, washes that pool on top with hard edges. | The sheen needs light on the surface. |
How the papers change watercolor and ink blocks, paper by paper, is in the guide "How paper changes watercolor and ink"; how to pick and tune one is in "How to choose a paper for my art style".
## Whole looks
Three looks made only of paperBlocks' papers and the blocks it brings (styleKit's and starterBlocks'), each on in one call:
| Look | What it is |
|---|---|
| `paperCutoutGame` (Paper cutout game) | Card on kraft: matte card surfaces, two flat light bands, flat colour and a white die-cut border round every object. styleRoles' paper cutout set is made for it. |
| `chalkboardLesson` (Chalkboard lesson) | A lesson in chalk: a few flat tones, a bold white chalk line, lights put on in chalk, on a dusty blackboard. |
| `parchmentMap` (Parchment map) | An old map or a quest journal: sepia and cream inks, soft light bands and a brown ink line, on parchment. |
```luau
require("~styleKit.StyleKit").applyLook("chalkboardLesson")
```
## The shared settings
Every paper takes the same settings (each paper starts from its own defaults; `StyleKit.settingsOf("paperKraft")` lists them with that paper's defaults and the ranges). Change them with `StyleKit.set(entry, { name = value })`; colours take `"#rrggbb"`.
| Setting | Range | What it does |
|---|---|---|
| `tint` | colour | The sheet's colour. On a white paper the bright areas become this colour; on a toned or dark sheet it is the ground the darks sink into and the lights are put on over. |
| `toothHeight` | 0 to 1 | How tall the tooth is: how much it shades the sheet, how deep paint settles into it, and how strongly granulation, dry brush and ink lines feel it. |
| `toothSharpness` | 0 to 1 | 0 soft rounded felt bumps, 1 sharp grit. Sharp grit makes put-on lights catch the peaks first (chalk on slate). |
| `toothSize` | 1 to 40 | How large one tooth looks, in reference pixels (1080p). It keeps about this size at any distance. |
| `fibre`, `fibreLength`, `fibreTone` | 0 to 1, 1 to 30, -1 to 1 | How many long fibres show, how long they are (in teeth), and whether they read lighter (above 0) or darker (below 0) than the sheet. |
| `weave` | 0 to 1 | How much canvas weave replaces the tooth. |
| `mottle` | 0 to 1 | How blotchy and uneven the sheet's tone is (warmer where darker). |
| `foxing` | 0 to 1 | How many brown age spots the sheet carries. |
| `rule`, `ruleColour`, `ruleSpacing` | 0 to 1, colour, 8 to 200 | Ruled lines: how strongly they show, their colour, and their spacing in reference pixels. |
| `thirst` | 0 to 1 | How thirsty the sheet is: 0 hard-sized (crisp wash edges, dark dried rims, exact lines), 0.5 the standard sheet the watercolor and ink blocks were tuned on (nothing changes), 1 thirsty (washes spread and feather, lines bleed into the fibres). Felt only by blocks that read the paper field. |
| `emboss` | 0 to 1 | How strongly the tooth catches light, strongest where the scene's sun rakes low across a surface. |
| `sheen` | 0 to 1 | How much the sheet shines where light reaches it. Sheen is light the scene sends: none where there is none. |
| `lights` | 0 to 1 | How much of the picture's light can be put on in an opaque light (gouache, chalk, white ink) instead of being left as bare sheet. 0 for white papers. Any amount above about 0.25 makes the sheet a toned ground (the darks sink into it) and the value is then the most light put on: the boards use 0.6, so a chalk drawing stays marks over slate even where the scene is brightest. |
| `lightColour`, `lightTint` | colour, 0 to 1 | The colour of the put-on lights, and how much of the picture's own colour they keep. |
| `lightsFrom` | 0 to 1 | Where lights start being put on, as a share of the way from the sheet's value to white: 0 everything lighter than the sheet, toward 1 only the brightest lights. |
| `dust` | 0 to 1 | Chalk dust and the ghosts of old wiping on a board. |
## How a paper draws
- **Darks go into the sheet** as transparent media: multiplied onto the sheet's colour and settling into the valleys of its tooth, thinner on its peaks.
- **Lights are put on** on toned and dark sheets (`lights` above 0), from `lightsFrom` of the way between the sheet's value and white; on sharp grain they catch the tooth's peaks first, on soft grain they lie more evenly.
- **The tooth is lit by the scene's sun** (the first directional light, read every few frames and written again whenever the paper's shader compiles), so a low raking sun shows it most, plus a little relief from the upper left as a sheet reads on a table.
- **Everything sits on surfaces** and keeps its size in the picture: the tooth, fibres, blotches, spots and rules are laid on world position, so they stay on things under a pan, a dolly or a zoom, and any part finer than about two drawn pixels fades to its average, so nothing shimmers in the distance or on the low tier. They slide across moving characters, as every full-frame mark does.
- **Black stays black.** Every pixel goes through styleKit's shared exit, so with every light off the frame is black on every paper, white paper and chalkboard alike.
Each paper publishes the stack's shared ground (the sheet's colour and tooth height, which granulation, dry brush and ink lines read) and the **paper field** (thirst, sheen, fibre and whether the sheet is dark), which the watercolor and ink blocks read. The paper field is made only when a block in the stack reads it, so a paper under styleKit's own blocks costs nothing for it.
## Regions on different papers
A later paper in the stack lays its ground and paper field over the one before by its strength and mask, so regions of one world sit on different papers:
```luau
StyleKit.add("paperColdPress", { id = "paper" }) -- the whole world on cold press
StyleKit.add("paperParchment", { id = "mapRoom",
mask = { kind = "sphere", center = { 40, 2, -12 }, radius = 6 } }) -- the map room on parchment
```
Every mask kind works (depth, rect, sphere, group; `invert = true` flips it). The older papers (`plainPaper`, `darkBoard`, `inkPaper`, `wcColdPress`) do not layer: one of them can be the first paper under a paperBlocks region, but not a region laid over another paper.
## A paper on an object
The papers above lay the sheet over the whole frame from world position, so a prop that moves, or a character that walks, slides under the tooth. To put an object on the same paper with the tooth on its own surface, give it the paper as a material:
```luau
local PaperMaterial = require("~paperBlocks.PaperMaterial")
PaperMaterial.paint(heroId) -- the paper the stack is on, with its tuned settings
StyleKit.options({ groups = { { "Hero" } } }) -- name the object (or its parent) as mask group 1
StyleKit.setMask("paper", { kind = "group", group = 1, invert = true }) -- the full-frame paper leaves it to its own material
```
`paint` gives every Model at or under the entity a material on paperBlocks' `paperSurface` shader, derived from the one it wears: its base colour and texture are kept (the original collection is never rewritten), and it takes every setting of the paper the stack's first paperBlocks Ground entry uses, your tuning included. `PaperMaterial.paint(heroId, "paperCanvas", { emboss = 0.8 })` names another paper and overrides any setting. What the material does:
- **The tooth stays on the object.** Tooth, fibres, weave, blotches, foxing and rules are laid on the object's own position, so they ride with it as it moves (on a skinned character, the bind pose, so it stays on the body as it animates). They keep about the full-frame paper's size in the picture: the object's `PaperSurface` component keeps each machine's screen height in the material, so it matches on every screen.
- **Every light rakes it.** The tooth bends the surface normal and the engine's own lighting does the rest, so the sun, point lights and shadows show the tooth the way raking light shows it on the frame; `sheen` makes it shine where light reaches it.
- **The same picture rules.** Darks go into the sheet and its valleys; on toned and dark papers the lights are put on in `lightColour` from `lightsFrom` up; nothing is brighter than the light that reached it, so an unlit object stays black.
- **One paper, not two.** The group mask keeps the full-frame paper off the object, so the tooth is not laid twice (once sliding, once riding). If you would rather keep the full-frame paper on it, paint with `{ sheet = 0 }`: the material then adds only the tooth's relief and tone, and the frame lays the sheet.
The blocks above the paper (washes, granulation, lines) still run over the object as they run over everything; granulation and dry brush read the frame's tooth there, so on a moving object keep those off it with a group mask too. `PaperMaterial.unpaint(heroId, { removeFiles = true })` puts every original material back and deletes the materials it wrote; `PaperMaterial.list()` names every painted entity. The materials are world assets (in `/zero/source/paperMaterials` by default) and the swap is a synced change, so every player sees the paper on the object.
Limits: see-through materials (glass, water) are left as they are. A material whose own shader lays its texture in its own way (a pixel-art atlas, world-space tiling) keeps its colour and texture but the texture is then laid on the mesh's own UVs, so check it after painting. The material reads no style stack: if you retune the stack's paper, paint the object again to match.
## Mixing with watercolorBlocks and inkBlocks
The papers are Ground blocks, which run after Fill and Line, so washes and lines are laid first and the paper takes them. With watercolorBlocks or inkBlocks installed beside paperBlocks:
```luau
StyleKit.applyLook("watercolor") -- watercolorBlocks' whole look, on its own cold press
StyleKit.swap("paper", "paperRice") -- on rice paper: washes run further and rims soften
StyleKit.swap("paper", "paperHotPress") -- on hot press: crisp washes, darker narrower rims
StyleKit.applyLook("brushAndWash") -- inkBlocks' ink wash painting
StyleKit.swap("paper", "paperRice") -- the brush line feathers into the fibres
StyleKit.swap("paper", "paperBlackboard") -- the ink is drawn light, the picture in chalk
```
`wcBleed` and `wcEdgeDarken` (watercolorBlocks 1.1.0) and every inkBlocks line and fill (inkBlocks 1.0.4) read the paper field. Granulation, dry brush and every ink line's `paperBreak` read the tooth through the shared ground, so they feel how much tooth each paper really has (almost none on hot press and board). On cold press with its standard thirst every block draws exactly as it did on the paper it was tuned on.
## Relation to the other papers
| Paper | Package | What it is | Next to paperBlocks |
|---|---|---|---|
| `plainPaper` | starterBlocks | Plain warm white paper with a fine grain. | Swaps with any paperBlocks paper in one change. Publishes no paper field, so the blocks that read it draw as on the standard sheet. |
| `darkBoard` | starterBlocks | A simple chalk board with a ceiling on chalk coverage. | `paperBlackboard` is the fuller board: slate grain, dust, wiping ghosts, layering, and a paper field that tells ink blocks to draw light. |
| `inkPaper` | inkBlocks | A copy of `plainPaper`, shipped so inkBlocks' looks need nothing else. | Swaps in one change; no paper field. |
| `wcColdPress` | watercolorBlocks | watercolorBlocks' rough cold-press paper. | Swaps in one change; no paper field. `paperColdPress` is the same kind of sheet in the shared paper model. |
None of the older papers publishes the paper field, so the frame under them stays exactly as before.
## Quality
The papers follow the stack's quality tier. On `low` the chain runs at half resolution: the paper keeps its tooth at the same size in the picture, and any part too fine to draw at half resolution fades to its average instead of shimmering. `StyleKit.setQuality("low")` forces the low tier on this machine; `StyleKit.setQuality(nil)` follows the device again.
A paper costs one pass for the picture and one for its ground; the paper field adds one more pass, made only when a block in the stack reads it. Each extra region on another paper adds its own picture pass, ground pass and (when read) paper field pass.
## Multiplayer
The paper is an entry on the scene's synced StyleStack, so every player sees the same paper, settings and regions, and any player's change reaches everyone through the stack's owner. Swap memory is kept on the entry, so it is shared too. The papers keep no state of their own. Only the quality tier is per machine.
## What is in the package
| Asset | What it is |
|---|---|
| `blocks/` | The seventeen papers (`<name>.styleBlock`) and the three shaders they all share: `paper.shader` (the picture on the sheet), `paperField.shader` (the shared ground) and `paperTakes.shader` (the paper field). |
| `shaders/paperTooth.shaderModule` | The pure tooth math (felt tooth and grit, fibres, weave, rules, foxing) shared by the full-frame papers and the paper material. |
| `shaders/paperCore.shaderModule` | The full-frame paper model built on it: tooth, fibres, weave, blotches, foxing, rules, chalk dust laid on surfaces. |
| `shaders/paperSurface.shader` | The paper as a material for an object: the same tooth on the object's own position, lit by the scene's lights. |
| `PaperMaterial`, `PaperSurface` | Put an object on paper (`paint`, `unpaint`, `list`) and the component that records its original materials and keeps the tooth's size right on each screen. |
| `Gallery.scene`, `PaperGallery`, `PaperGalleryLook` | Every paper side by side under one stack, each around its own stand as a layered region; the look over them is fitted to what your world has installed every time the scene loads. |
| `guides/` | How to choose a paper for my art style, how paper changes watercolor and ink, and a worked example level on paper. |
| `workflows/paperWorld.workflow` | Put an existing world on paper, check it, and leave it undoable. |
| `tests/` | The package's test suite. |
## How much paper shows through each object
Every paper is a Ground block, so styleKit's style channels decide how much of it shows through on each object: an object whose channel (or game role) says `marks.paper = 0` stays clean printed card on any sheet, 0.33 lets a little grain through, 1 lays it fully on the paper. The styleRoles package's paper cutout set uses it: the hero printed clean, enemies on the paper's grain, coins as bright clean card.
## Known limits
- **Installs with several dependencies can be slow (#21420).** paperBlocks brings styleKit, lookbookKit and starterBlocks; the check above took about 3 s, but give a busy engine time to finish before using it.
- **The web build was not checked from inside** (#20946: an agent cannot start a web engine). `zero lookbook report` is the one command to run there.
- **A second machine was not tried** (#14448: a second engine on the same account joins as the same player). The stack's fields are replicated, the papers keep no state, and the held clock is wall-clock UTC.
- **Frame rate on the web and in the web gallery** is lower than on desktop; start slower machines on the low tier with `StyleKit.setQuality("low")`.
- **The full-frame sheet slides across moving characters.** It is laid on world position. Put the character on paper inside its own material instead (`PaperMaterial.paint`, "A paper on an object" above), and keep the full-frame paper off it with a group mask.
- **Glass, water and particles have no depth**, so the sheet under them is the sheet of the surface behind.
- **Offscreen captures and render-to-texture cameras do not draw the stack** (styleKit's Known limits).
- **Older papers do not layer.** `plainPaper`, `darkBoard`, `inkPaper` and `wcColdPress` can only be the first paper of a stack with regions.
- **Other engine issues that apply here:** every render pass is queued again every frame, so script cost grows with the passes a look queues (#24548); a per-object value cannot reach a geometry pass, so per-object looks go through styleKit's object mask (at most eight distinct settings) or need passes of their own (#24549); a two-client editor sometimes refuses to load two different scenes at once (#24666, #24667); a world boot can wait about 30 s on a network handshake before content starts (#24669); the ZeroMind description can lag one release (#19258).
## Removing it
`StyleKit.swap("paper", "<the paper you had>")` returns the old paper with exactly its settings; `StyleKit.remove("paper")` takes the paper off; `StyleKit.clear()` takes the whole look off. paperBlocks edits no material and nothing outside the stack, so nothing else needs putting back. Uninstalling the package leaves a stack that names its papers drawing without them until they are swapped out.
## Version history
**1.2.2 (2026-11-12)**: pins starterBlocks 1.0.49, styleKit 1.7.8 (styleKit 1.7.8 logs no second refusal for the old roles field on a scene saved before styleKit 1.5). Nothing else changed.
**1.2.1 (2026-11-12)**: Chalkboard lesson reads as a board: on a light scene it came out flat grey, because the blackboard paper draws everything in light marks; a two-ink step now maps the mid tones onto the dark board and only the lit parts onto chalk (the squint test with roles on went from 23.9 to 27.1). The test suite names the Lookbook indirectly, so installing paperBlocks no longer brings lookbookKit along. Pins styleKit 1.7.7 and starterBlocks 1.0.48.
**1.2.0 (2026-11-12)**: paper gets whole looks of its own, three of them, each made only of paperBlocks' papers and the blocks paperBlocks brings (styleKit's and starterBlocks', now a dependency): **Paper cutout game** (`paperCutoutGame`, card on kraft with a white die-cut border; it lived in toonBlocks made of toon blocks, and toonBlocks brought paperBlocks just for it), **Chalkboard lesson** (`chalkboardLesson`, a lesson in white chalk on a dusty blackboard) and **Parchment map** (`parchmentMap`, sepia and cream inks with a brown line on parchment, for a fantasy map or a quest journal). styleRoles' paper cutout role set follows all three. Pins styleKit 1.7.6 and starterBlocks 1.0.47.
**1.1.32 (2026-11-12)**: pins styleKit 1.7.5 (styleKit 1.7.5: its checks pass where styleKit is installed alone). Nothing else changed.
**1.1.31 (2026-11-12)**: pins styleKit 1.7.4 (styleKit 1.7.4: installs on its own into an empty world again). Nothing else changed.
**1.1.30 (2026-11-12)**: pins styleKit 1.7.3 (styleKit 1.7.3: a role reaches what a kit draws in an object's place (pickupsKit's coins), and looks installed after the first scan are found). Nothing else changed.
**1.1.29 (2026-11-12)**: pins styleKit 1.7.0 (styleKit 1.7.0: role sets' glazes tell groups sharing a material apart, things with no role keep the plain line under a toon look, a look switched on never flashes black while its shaders compile). Known limits no longer cite #20750, which is closed. Nothing else changed.
**1.1.28 (2026-11-11)**: pins styleKit 1.6.5 (styleKit 1.6.5 draws a newly switched-on look's settled picture from its first frames instead of a lighter, half-set one for two to four seconds; lookbookKit 1.0.44 finds its Lookbook scene where it is installed). Nothing else changed.
**1.1.27 (2026-11-11)**: pins styleKit 1.6.4 (styleKit 1.6.4 puts a role given to a player template on the body every joining player is copied from and rebuilds the role mask at most four times a second, and toonBlocks 1.3.6 starts its toon ink by itself and draws every role ink through one shared flood). Every README in the package (each block, look, shader and guide) now ends with a Known limits section naming the engine issues that apply there.
**1.1.26 (2026-11-11)**: pins styleKit 1.6.2 (styleKit 1.6.2 finds the scene's stack again when the one it held was removed (after a play restart a look or role set change failed with "the component this proxy names was removed")). Nothing else changed.
**1.1.25 (2026-11-11)**: a style channel test on styleKit's ChannelProbe (the paper mark reaches the ground on the presented frame); pins styleKit 1.6.1.
**1.1.24 (2026-11-11)**: pins styleKit 1.5.1 (styleKit 1.5.1 keeps the scene's role set working in an editor that had an older styleKit open). Nothing else changed.
**1.1.23 (2026-11-11)**: documents how much paper shows through each object with styleKit 1.5.0's style channels. No block changed. Pins styleKit 1.5.0.
**1.1.22 (2026-11-08)**: moves to the engine's current built-in library (@builtin e89b9db), whose update rewrote small reference files inside the package (a world on an older built-in library is hidden from players). The package description now says what it gives you and how to install and use it. Nothing else changed.
**1.1.21 (2026-11-08)**: moves to the engine's current built-in library (@builtin edb50c2), whose update rewrote small reference files inside the package (a world on an older built-in library is hidden from players). Nothing else changed.
**1.1.20 (2026-11-08)**: moves to the engine's current built-in library (@builtin f4425a5), whose update rewrote small reference files inside the package (a world on an older built-in library is hidden from players). Nothing else changed.
**1.1.19 (2026-11-08)**: moves to the engine's current built-in library (@builtin f39f044), whose update rewrote small reference files inside the package (a world on an older built-in library is hidden from players). Nothing else changed.
**1.1.18 (2026-11-08)**: the package's description now reads as a store listing (what you get and how to start), without a version number or a list of changes. No paper changed.
**1.1.17 (2026-11-08)**: moves to the engine's current built-in library (@builtin f110cb3), whose update rewrote small reference files inside the package. Nothing else changed.
**1.1.16 (2026-11-08)**: moves to the engine's current built-in library (@builtin a29b227), whose update rewrote small reference files inside the package. Nothing else changed.
**1.1.15 (2026-11-08)**: moves to the engine's current built-in library (@builtin c61f473), whose update rewrote small reference files inside the package. Nothing else changed.
**1.1.14 (2026-11-07)**: moves to the engine's current built-in library (@builtin 7ca1f19), whose update rewrote small reference files inside the package. Nothing else changed.
**1.1.13 (2026-11-07)**: moves to the engine's current built-in library (@builtin fdc9a99), whose update rewrote small reference files inside the package. Nothing else changed.
**1.1.12 (2026-11-07)**: moves to the engine's current built-in library (@builtin 7e13a48), whose update rewrote small reference files inside the package. Nothing else changed.
**1.1.11 (2026-11-07)**: moves to the engine's current built-in library (@builtin 2809239), whose update rewrote small reference files inside the package. Nothing else changed.
**1.1.10 (2026-11-07)**: moves to the engine's current built-in library (@builtin 278f02f), whose update rewrote small reference files inside the package. No block, look or shader changed.
**1.1.9 (2026-11-07)**: moves to the engine's current built-in library (@builtin 6656162), whose update rewrote small reference files inside the package. No block, look or shader changed.
**1.1.8 (2026-11-07)**: moves to the engine's current built-in library (@builtin 808b4a1), whose update rewrote small reference files inside the package. No block, look or shader changed.
**1.1.7 (2026-11-02)**: "Lookbook numbers" records the check made on a fresh install of 1.1.6: cold press, kraft and blackboard measured twice each in one unattended batch, every run finished on its own and matched the table within 0.01, the unstyled frame 2.30 every time. No paper changed.
**1.1.6 (2026-11-02)**: the measurement in this README stalled partway and its numbers drifted because the Lookbook stepped once per drawn frame while the engine stopped drawing a still picture; lookbookKit 1.0.14, which a fresh install now brings, holds every-frame drawing during a run. "Lookbook numbers" says so, says how much a run can still move (a first, warm-up run by about 0.05), and withdraws 1.1.5's explanation (#19849), which was wrong. No paper changed.
**1.1.5 (2026-11-02)**: needs styleKit 1.2.4. Documentation: a reviewer's two new findings, a stall partway through a long measurement (#19849) and a mismatch beyond 0.05 on cold press and blackboard not seen here, are written up under "Lookbook numbers" with what is known and not known about them. No block changed.
**1.1.4 (2026-11-02)**: needs styleKit 1.2.3, and the Lookbook table is new. A reviewer could not reproduce the 1.1.3 table on a fresh install. The cause was styleKit 1.2.2 losing the scene's sun on a paper whenever its shader compiled, so each run drew the tooth with or without sun relief depending on timing (see "Why the earlier table did not reproduce"). styleKit 1.2.3 fixes it, and every paper now measures the same on both runs of a fresh install: all seventeen papers measured twice with exactly the call the README names. No block changed.
**1.1.3 (2026-11-01)**: documentation only: the Gallery's README quotes a single paper's cost as a fresh install measures it (0.52 to 0.74 ms high, 0.15 to 0.20 ms low at 1920 x 1080).
**1.1.2 (2026-11-01)**: documentation only. The Lookbook numbers are now what a fresh install measures: every paper twice in a brand new world with only paperBlocks installed, right after the engine started, at 1920 x 1080. Where a paper also showed a second repeatable outcome (kraft on a fresh install; several papers in a long session) the table lists it, since which outcome a run lands on depends on what ran before it in the engine. The 1.1.1 table came from a long session and held some of those second outcomes (cold press 3.13 / +0.18 / 1.86 where a fresh install gives 3.64 / -0.05 / 1.57). No block changed.
**1.1.1 (2026-11-01)**: the Gallery fits its look to your world every time it loads (in edit, in play, on every peer), so the first open in a paperBlocks-only world already draws styleKit's pencil line over every paper and its title card says so; before, only a rebuild did. Lookbook numbers measured again on styleKit 1.2.2 and lookbookKit 1.0.11: the earlier numbers had been taken in one long batch in which each paper's first second of path was embossed under the sun left over from the run before (styleKit read the sun about once a second), so they did not match a run on its own. The section now states the exact call and view size, and how the numbers move with the view size.
**1.1.0 (2026-11-01)**: a paper on an object: the `paperSurface` material lays any paper's tooth, fibres, weave and tone on an object's own position, lit by the scene's own lights, and `PaperMaterial.paint(entityId)` puts an object on the stack's paper (or a named one) and `unpaint` puts it back (see "A paper on an object"); the pure tooth math moved into `paperTooth`, shared by the frame and the material, with no change to how any paper draws; a test for painting and unpainting.
**1.0.1 (2026-11-01)**: the README says exactly what an install brings (styleKit, lookbookKit and starterBlocks), as a fresh-world install showed; no block changed.
**1.0.0 (2026-11-01)**: first release: seventeen papers sharing one paper model and one set of settings, the Gallery, guides, the paperWorld workflow and the test suite.
## Lookbook numbers
**What a fresh install measures.** Every paper alone, measured twice on 2026-11-02 in a brand new private world: paperBlocks installed from ZeroMind (which brought styleKit, lookbookKit 1.0.11 and starterBlocks 1.0.11) and nothing else, the stack empty before and after every run (no entries, hold rate 8, soft radius 10, no pigment set), desktop GPU, the Lookbook's 3D view at exactly **1920 x 1080** (checked before every run), one discarded warm-up run first, no missed frames. Each paper was measured with exactly this call and nothing else on the stack:
```luau
local Lookbook = require("~lookbookKit.Lookbook")
local r = Lookbook.measure({ stack = { { id = "paper", block = "paperKraft" } } })
print(r.size[1], r.size[2], r.stylized.all.warpedDiff, r.stylized.markStuck, r.stylized.screenStuck, r.stylized.zoomMarks.growth, r.passMs.high.total, r.passMs.low.total)
```
`zero viewport size` reports the 3D view as `sceneWidth` and `sceneHeight`; the window is larger than the view by the editor's panels, and the Lookbook refuses a run whose view changes size. The styleKit these numbers were taken on carries the sun fix styleKit 1.2.3 ships (below), which paperBlocks 1.1.4 needs: the published styleKit was still 1.2.2 that night, so the one-line fix was put into the fresh install's styleKit by hand. **Every paper's two runs agree** (to 0.01 where they differ at all), and the unstyled frame the same call measures was identical in all 34 runs: 2.30 (whole path), 6.57 (pan), 0.00, -0.53, 7.88. The runs are in `design/lookbook/paper_fresh_runs.txt` in the Style Cookbook world.
**Why the earlier table did not reproduce.** The Gallery's stack settings were not the cause: the Lookbook replaces the Gallery scene before it measures, the measured stack was empty with the default hold rate and soft radius before and after every run, and those defaults are the Gallery's own values anyway. The cause was a styleKit bug. styleKit 1.2.2 writes the scene's sun to a block only when the sun changes, and whenever a block's shader is compiled (the first time a paper is used in a fresh install) or recompiled, the engine puts the block's values back to their defaults, which for the sun means "no sun". So a run drew the paper with or without the sun raking its tooth depending on timing. With paperBlocks 1.1.3 and styleKit 1.2.2 freshly installed, 11 of the 17 papers gave two different results on their two runs (kraft 2.26 / +0.41 / 1.81, then 1.93 / +0.50 / 2.59; cold press 3.13 / +0.18 / 1.86 on both runs right after a warm-up of 3.64 / -0.05 / 1.57), while the unstyled frame never changed. The sunless result is the one with higher mark growth and a little less swimming; the earlier table's last column ("other repeatable outcome") was that result, and its two kraft rows were one of each. With the fix put in, the same fresh install gave every paper the same numbers twice. A mark growth of about 1.0 was not seen in any of 87 runs on this machine (the lowest is 1.57, cold press): if your runs show it, `zero lookbook report` and your GPU would help find out why.
**A run used to stall partway and drift; it no longer does (lookbookKit 1.0.14).** A reviewer running the call above in a fresh world saw every run stop partway (usually around step 90 of 780) and sit there until the engine was poked every few seconds; when the runs did finish, kraft matched this table but cold press and blackboard did not (cold press 3.73 / growth 1.29 on one run, blackboard 2.95 and growth 1.35 on one run and 2.43 on the other), and the unstyled frame read 2.32 to 2.41 instead of 2.30. All of it had one cause, and it was in the Lookbook, not in the papers. The engine draws only frames that show something new unless it is told to draw every frame; the Lookbook advances one camera pose per drawn frame, so once the renderer judged the picture still it stopped drawing and the run waited, and each poke let it on for a few frames at uneven moments, so the frames it measured were not the frames an unbroken run measures. These numbers were taken by runs that never stalled, because the tool calls of the agent taking them kept the engine drawing. From lookbookKit 1.0.14 on (a fresh paperBlocks install brings it), `Lookbook.measure` and `Lookbook.run` hold every-frame drawing for their whole length and put the setting back afterwards, so a run finishes on its own, with no one touching the engine, and plays the same frames every time. An earlier version of this README blamed a task-scheduling issue (#19849): that was wrong.
**How much a run can still move.** The first run after a paper is first used in a session (its shader compiles while the path plays) can differ from the later ones by about 0.05 in the warped difference and the marks score; that is the warm-up run discarded above. After it, two runs of the same paper agree to 0.01, and the unstyled frame reads 2.30 every time. A difference beyond that, with lookbookKit 1.0.14 or later, is worth reporting with the paper and `zero lookbook report`.
**Checked on a fresh install after the fix.** On 2026-11-02 (18:50 to 18:57 UTC), in a brand new private world with only paperBlocks 1.1.6 installed from ZeroMind (it brought lookbookKit 1.0.14, styleKit 1.2.4 and starterBlocks), the 3D view at 1920 x 1080, the call above was run for cold press (a warm-up, then twice), kraft and blackboard (twice each), one after another in a single unattended batch. Every run finished on its own; the renderer held every-frame drawing for 377 of 382 one-second samples taken during the batch (the other five fall between runs) and gave the setting back afterwards; no frame was missed. Cold press read 3.64, 3.64 and 3.63 / 7.30 / -0.05 / -0.54 / 1.57, kraft 2.26 / 4.90 / +0.41 / -0.53 / 1.81 twice, blackboard 2.87 and 2.88 / 4.51 / +0.37 / -0.60 / 2.43, and the unstyled frame 2.30 / 6.57 / 0.00 / -0.53 / 7.88 in all seven runs: the table's rows, within 0.01.
How to read them, against the unstyled frame rather than as pass marks:
- **Warped difference** (0..255, lower is steadier; "how much the picture swims"): how much each frame differs from the one before once the camera's motion is taken out. Unstyled: 2.30 (whole path), 6.57 (pan).
- **Marks on things** (the shower-door score): negative means the stack's own marks line up with the surfaces when the camera pans, positive that they line up with the screen. Read the warning below.
- **Screen-stuck in the pan**: the whole picture's detail against the same spot on the screen (more negative is less stuck). Unstyled: -0.53.
- **Mark growth over the 10x zoom**: about 1 means marks keep their size in the picture, about 10 that they grow with the things they sit on (unstyled 7.88). Papers with almost no tooth (vellum, board, hot press, the notebook's rules) mostly follow the scene.
- **GPU ms high / low**: the style passes alone at the pan, on the high and the low quality tier.
- **Every light off**: the frame's mean brightness at the Lookbook's every-light-off moment, one frame per paper (unstyled 1.22): it must stay at the unstyled frame's black.
| Paper | Run | Warped difference, whole path | Warped difference, pan | Marks on things | Screen-stuck in the pan | Mark growth over the 10x zoom | GPU ms high | GPU ms low | Every light off (0..255) |
|---|---|---|---|---|---|---|---|---|---|
| Cold press (`paperColdPress`) | 1 | 3.63 | 7.30 | -0.05 | -0.54 | 1.57 | 0.64 | 0.17 | 1.15 |
| Cold press (`paperColdPress`) | 2 | 3.63 | 7.30 | -0.05 | -0.54 | 1.57 | 0.64 | 0.18 | |
| Rough (`paperRough`) | 1 | 4.12 | 7.96 | -0.06 | -0.52 | 1.58 | 0.64 | 0.17 | 1.17 |
| Rough (`paperRough`) | 2 | 4.11 | 7.96 | -0.06 | -0.52 | 1.58 | 0.64 | 0.18 | |
| Hot press (`paperHotPress`) | 1 | 2.35 | 6.42 | +0.45 | -0.53 | 3.86 | 0.64 | 0.17 | 1.19 |
| Hot press (`paperHotPress`) | 2 | 2.35 | 6.42 | +0.45 | -0.53 | 3.86 | 0.64 | 0.18 | |
| Rice paper (`paperRice`) | 1 | 2.87 | 6.56 | +0.15 | -0.53 | 2.02 | 0.64 | 0.18 | 1.18 |
| Rice paper (`paperRice`) | 2 | 2.87 | 6.56 | +0.15 | -0.53 | 2.02 | 0.64 | 0.17 | |
| Newsprint (`paperNewsprint`) | 1 | 2.19 | 5.71 | +0.84 | -0.53 | 3.26 | 0.64 | 0.17 | 1.06 |
| Newsprint (`paperNewsprint`) | 2 | 2.19 | 5.71 | +0.84 | -0.53 | 3.26 | 0.64 | 0.17 | |
| Kraft (`paperKraft`) | 1 | 2.26 | 4.90 | +0.41 | -0.53 | 1.81 | 0.64 | 0.18 | 1.14 |
| Kraft (`paperKraft`) | 2 | 2.26 | 4.90 | +0.41 | -0.53 | 1.81 | 0.64 | 0.18 | |
| Parchment (`paperParchment`) | 1 | 2.37 | 6.25 | +0.65 | -0.53 | 2.10 | 0.71 | 0.20 | 1.08 |
| Parchment (`paperParchment`) | 2 | 2.36 | 6.25 | +0.65 | -0.53 | 2.10 | 0.71 | 0.20 | |
| Canvas (`paperCanvas`) | 1 | 4.10 | 7.79 | +0.05 | -0.44 | 3.94 | 0.70 | 0.19 | 1.08 |
| Canvas (`paperCanvas`) | 2 | 4.09 | 7.79 | +0.05 | -0.44 | 3.94 | 0.70 | 0.19 | |
| Illustration board (`paperBoard`) | 1 | 2.28 | 6.35 | +0.78 | -0.53 | 6.07 | 0.54 | 0.15 | 1.15 |
| Illustration board (`paperBoard`) | 2 | 2.28 | 6.35 | +0.78 | -0.53 | 6.07 | 0.54 | 0.15 | |
| Blackboard (`paperBlackboard`) | 1 | 2.88 | 4.51 | +0.37 | -0.60 | 2.43 | 0.67 | 0.18 | 1.17 |
| Blackboard (`paperBlackboard`) | 2 | 2.87 | 4.51 | +0.37 | -0.60 | 2.43 | 0.67 | 0.18 | |
| Green chalkboard (`paperGreenChalkboard`) | 1 | 2.10 | 2.42 | +0.67 | -0.59 | 2.48 | 0.67 | 0.18 | 1.33 |
| Green chalkboard (`paperGreenChalkboard`) | 2 | 2.10 | 2.42 | +0.67 | -0.59 | 2.48 | 0.67 | 0.18 | |
| Black paper (`paperBlack`) | 1 | 3.45 | 8.74 | -0.56 | -0.55 | 2.09 | 0.64 | 0.17 | 1.10 |
| Black paper (`paperBlack`) | 2 | 3.45 | 8.74 | -0.56 | -0.55 | 2.09 | 0.64 | 0.18 | |
| Aged sketchbook (`paperSketchbook`) | 1 | 2.67 | 6.07 | +0.47 | -0.53 | 1.71 | 0.76 | 0.21 | 1.03 |
| Aged sketchbook (`paperSketchbook`) | 2 | 2.67 | 6.07 | +0.47 | -0.53 | 1.71 | 0.76 | 0.21 | |
| Toned grey (`paperTonedGrey`) | 1 | 2.72 | 6.62 | +0.07 | -0.53 | 1.77 | 0.64 | 0.17 | 1.20 |
| Toned grey (`paperTonedGrey`) | 2 | 2.72 | 6.62 | +0.07 | -0.53 | 1.77 | 0.64 | 0.18 | |
| Toned tan (`paperTonedTan`) | 1 | 2.86 | 7.08 | -0.10 | -0.53 | 1.77 | 0.64 | 0.17 | 1.17 |
| Toned tan (`paperTonedTan`) | 2 | 2.86 | 7.08 | -0.10 | -0.53 | 1.77 | 0.64 | 0.17 | |
| Ruled notebook (`paperNotebook`) | 1 | 2.54 | 6.39 | +0.20 | -0.52 | 4.48 | 0.66 | 0.18 | 1.19 |
| Ruled notebook (`paperNotebook`) | 2 | 2.54 | 6.39 | +0.20 | -0.52 | 4.48 | 0.66 | 0.18 | |
| Vellum (`paperVellum`) | 1 | 2.23 | 6.22 | +0.91 | -0.53 | 7.41 | 0.64 | 0.18 | 1.13 |
| Vellum (`paperVellum`) | 2 | 2.23 | 6.22 | +0.91 | -0.53 | 7.41 | 0.64 | 0.18 | |
What they say:
- **The shower-door score does not show the tooth staying on things at 1080p, for most papers.** At 1920 x 1080 most papers score positive (+0.05 to +0.91); cold press, rough, black paper and toned tan score negative. A paper's tooth is laid on world position, so this points at the Lookbook's mark comparison at this size rather than at the papers, but it is not proved. Check sliding by eye with a slow pan in your own scene. The screen-stuck column, which reads the whole picture, is -0.44 to -0.60 for every paper: nothing glues the picture to the screen.
- **Swimming.** Smooth papers add almost nothing to the unstyled path (hot press 2.35, board 2.28, vellum 2.23, newsprint 2.19, kraft 2.26 against 2.30). Deep tooth adds detail that moves with the surfaces: cold press 3.63, rough 4.12, canvas 4.10, black paper 3.45.
- **Black stays black.** With every light off every paper stays at the unstyled frame's black (1.03 to 1.33 against 1.22).
- **Cost at 1920 x 1080.** A paper alone is 0.54 to 0.76 ms on the high tier and 0.15 to 0.21 ms on the low tier, about four times cheaper; most of it is the paper's ground field (about 0.5 ms high), the picture pass about 0.14 ms. Under blocks that read the paper field (watercolorBlocks' bleed and edge darkening, every inkBlocks block) the paper field adds one more pass of about 0.5 ms high.
- **Line and wash on paper** (watercolorBlocks and inkBlocks installed; from an earlier long session on styleKit 1.2.2, so its two runs can differ by the sun bug above): 4.55 to 4.71 ms high and 0.93 to 1.00 ms low; calmer than unstyled on hot press (1.25) and 2.5 to 3.8 on cold press and canvas.
- **The Gallery and the Paper Swatches recipe** run seventeen papers at once (70 passes with the line and wash look): 14.7 ms of GPU time on high at 2097 x 1409. They are swatch books, not stacks to ship; a game picks one or two papers.
- **Web frame rate:** run `zero lookbook report` in the web build; it has not been measured from here (#20946).
Line and wash, and the unstyled baseline, from the earlier long session on styleKit 1.2.2 (`design/lookbook/paper_runs.json`; columns: warped whole path, warped pan, marks, screen-stuck, growth, GPU high, GPU low, every light off):
| Stack | Run | Warped difference, whole path | Warped difference, pan | Marks on things | Screen-stuck in the pan | Mark growth over the 10x zoom | GPU ms high | GPU ms low | Every light off (0..255) |
|---|---|---|---|---|---|---|---|---|---|
| baseline | 1 | 2.30 | 6.57 | 0.00 | -0.53 | 7.88 | - | - | 1.22 |
| baseline | 2 | 2.30 | 6.57 | 0.00 | -0.53 | 7.88 | - | - | |
| lineAndWash:paperHotPress | 1 | 1.25 | 1.43 | 0.73 | -0.61 | 3.81 | 4.55 | 0.93 | 1.04 |
| lineAndWash:paperHotPress | 2 | 1.25 | 1.43 | 0.73 | -0.61 | 3.81 | 4.56 | 0.93 | |
| lineAndWash:paperColdPress | 1 | 2.46 | 2.13 | 0.65 | -0.62 | 1.82 | 4.57 | 0.93 | 1.03 |
| lineAndWash:paperColdPress | 2 | 3.13 | 3.39 | 0.52 | -0.58 | 1.49 | 4.56 | 0.93 | |
| lineAndWash:paperCanvas | 1 | 2.97 | 3.11 | 0.62 | -0.21 | 4.84 | 4.71 | 1.00 | 1.02 |
| lineAndWash:paperCanvas | 2 | 3.79 | 4.72 | 0.49 | -0.18 | 3.27 | 4.71 | 1.00 | |
| lineAndWash:paperHotPress | bleed.wcBleed | 0.39 | 0.39 | 0.10 | 0.10 |
| lineAndWash:paperHotPress | dry.wcDryBrush | 0.27 | 0.27 | 0.07 | 0.07 |
| lineAndWash:paperHotPress | edge.wcEdgeDarken | 0.20 | 0.21 | 0.05 | 0.05 |
| lineAndWash:paperHotPress | grain.wcGranulation | 0.23 | 0.23 | 0.07 | 0.06 |
| lineAndWash:paperHotPress | line.inkPenLine | 1.19 | 1.19 | 0.15 | 0.15 |
| lineAndWash:paperHotPress | paper.paperHotPress | 0.14 | 0.14 | 0.03 | 0.03 |
| lineAndWash:paperHotPress | paper.paperHotPress.field | 0.50 | 0.50 | 0.13 | 0.13 |
| lineAndWash:paperHotPress | paper.paperHotPress.paper | 0.50 | 0.50 | 0.13 | 0.13 |
| lineAndWash:paperHotPress | wash.wcWash | 1.12 | 1.12 | 0.20 | 0.20 |
| lineAndWash:paperColdPress | bleed.wcBleed | 0.40 | 0.40 | 0.10 | 0.10 |
| lineAndWash:paperColdPress | dry.wcDryBrush | 0.27 | 0.27 | 0.07 | 0.07 |
| lineAndWash:paperColdPress | edge.wcEdgeDarken | 0.21 | 0.21 | 0.05 | 0.05 |
| lineAndWash:paperColdPress | grain.wcGranulation | 0.23 | 0.23 | 0.07 | 0.07 |
| lineAndWash:paperColdPress | line.inkPenLine | 1.19 | 1.19 | 0.15 | 0.15 |
| lineAndWash:paperColdPress | paper.paperColdPress | 0.14 | 0.14 | 0.03 | 0.03 |
| lineAndWash:paperColdPress | paper.paperColdPress.field | 0.50 | 0.50 | 0.13 | 0.13 |
| lineAndWash:paperColdPress | paper.paperColdPress.paper | 0.50 | 0.50 | 0.13 | 0.13 |
| lineAndWash:paperColdPress | wash.wcWash | 1.12 | 1.12 | 0.20 | 0.20 |
| lineAndWash:paperCanvas | bleed.wcBleed | 0.40 | 0.40 | 0.10 | 0.10 |
| lineAndWash:paperCanvas | dry.wcDryBrush | 0.28 | 0.28 | 0.08 | 0.07 |
| lineAndWash:paperCanvas | edge.wcEdgeDarken | 0.21 | 0.21 | 0.05 | 0.06 |
| lineAndWash:paperCanvas | grain.wcGranulation | 0.23 | 0.23 | 0.07 | 0.07 |
| lineAndWash:paperCanvas | line.inkPenLine | 1.20 | 1.20 | 0.16 | 0.16 |
| lineAndWash:paperCanvas | paper.paperCanvas | 0.14 | 0.14 | 0.03 | 0.04 |
| lineAndWash:paperCanvas | paper.paperCanvas.field | 0.57 | 0.57 | 0.15 | 0.15 |
| lineAndWash:paperCanvas | paper.paperCanvas.paper | 0.56 | 0.56 | 0.15 | 0.15 |
| lineAndWash:paperCanvas | wash.wcWash | 1.13 | 1.13 | 0.20 | 0.20 |
Each pass's GPU milliseconds in the long session, both runs, both tiers:
| Stack | Pass | High, run 1 | High, run 2 | Low, run 1 | Low, run 2 |
|---|---|---|---|---|---|
| paperColdPress | paper.paperColdPress | 0.13 | 0.13 | 0.04 | 0.04 |
| paperColdPress | paper.paperColdPress.field | 0.50 | 0.50 | 0.13 | 0.13 |
| paperRough | paper.paperRough | 0.13 | 0.13 | 0.04 | 0.04 |
| paperRough | paper.paperRough.field | 0.50 | 0.50 | 0.13 | 0.13 |
| paperHotPress | paper.paperHotPress | 0.13 | 0.13 | 0.04 | 0.04 |
| paperHotPress | paper.paperHotPress.field | 0.50 | 0.50 | 0.13 | 0.13 |
| paperRice | paper.paperRice | 0.13 | 0.13 | 0.04 | 0.04 |
| paperRice | paper.paperRice.field | 0.50 | 0.50 | 0.13 | 0.13 |
| paperNewsprint | paper.paperNewsprint | 0.13 | 0.13 | 0.04 | 0.04 |
| paperNewsprint | paper.paperNewsprint.field | 0.50 | 0.50 | 0.13 | 0.13 |
| paperKraft | paper.paperKraft | 0.13 | 0.14 | 0.04 | 0.04 |
| paperKraft | paper.paperKraft.field | 0.50 | 0.50 | 0.13 | 0.13 |
| paperParchment | paper.paperParchment | 0.14 | 0.14 | 0.04 | 0.04 |
| paperParchment | paper.paperParchment.field | 0.58 | 0.58 | 0.15 | 0.15 |
| paperCanvas | paper.paperCanvas | 0.14 | 0.13 | 0.04 | 0.04 |
| paperCanvas | paper.paperCanvas.field | 0.56 | 0.56 | 0.15 | 0.15 |
| paperBoard | paper.paperBoard | 0.13 | 0.13 | 0.04 | 0.04 |
| paperBoard | paper.paperBoard.field | 0.40 | 0.40 | 0.11 | 0.11 |
| paperBlackboard | paper.paperBlackboard | 0.14 | 0.14 | 0.04 | 0.04 |
| paperBlackboard | paper.paperBlackboard.field | 0.54 | 0.54 | 0.14 | 0.14 |
| paperGreenChalkboard | paper.paperGreenChalkboard | 0.14 | 0.14 | 0.04 | 0.04 |
| paperGreenChalkboard | paper.paperGreenChalkboard.field | 0.54 | 0.54 | 0.14 | 0.14 |
| paperBlack | paper.paperBlack | 0.14 | 0.13 | 0.04 | 0.04 |
| paperBlack | paper.paperBlack.field | 0.51 | 0.50 | 0.13 | 0.13 |
| paperSketchbook | paper.paperSketchbook | 0.13 | 0.13 | 0.04 | 0.04 |
| paperSketchbook | paper.paperSketchbook.field | 0.62 | 0.63 | 0.17 | 0.17 |
| paperTonedGrey | paper.paperTonedGrey | 0.14 | 0.14 | 0.04 | 0.04 |
| paperTonedGrey | paper.paperTonedGrey.field | 0.51 | 0.50 | 0.14 | 0.13 |
| paperTonedTan | paper.paperTonedTan | 0.13 | 0.14 | 0.04 | 0.04 |
| paperTonedTan | paper.paperTonedTan.field | 0.50 | 0.50 | 0.13 | 0.13 |
| paperNotebook | paper.paperNotebook | 0.13 | 0.13 | 0.04 | 0.04 |
| paperNotebook | paper.paperNotebook.field | 0.53 | 0.53 | 0.14 | 0.14 |
| paperVellum | paper.paperVellum | 0.13 | 0.13 | 0.04 | 0.04 |
| paperVellum | paper.paperVellum.field | 0.50 | 0.49 | 0.13 | 0.13 |
| lineAndWash:paperHotPress | bleed.wcBleed | 0.39 | 0.39 | 0.10 | 0.10 |
| lineAndWash:paperHotPress | dry.wcDryBrush | 0.27 | 0.27 | 0.07 | 0.07 |
| lineAndWash:paperHotPress | edge.wcEdgeDarken | 0.20 | 0.21 | 0.05 | 0.05 |
| lineAndWash:paperHotPress | grain.wcGranulation | 0.23 | 0.23 | 0.07 | 0.06 |
| lineAndWash:paperHotPress | line.inkPenLine | 1.19 | 1.19 | 0.15 | 0.15 |
| lineAndWash:paperHotPress | paper.paperHotPress | 0.14 | 0.14 | 0.03 | 0.03 |
| lineAndWash:paperHotPress | paper.paperHotPress.field | 0.50 | 0.50 | 0.13 | 0.13 |
| lineAndWash:paperHotPress | paper.paperHotPress.paper | 0.50 | 0.50 | 0.13 | 0.13 |
| lineAndWash:paperHotPress | wash.wcWash | 1.12 | 1.12 | 0.20 | 0.20 |
| lineAndWash:paperColdPress | bleed.wcBleed | 0.40 | 0.40 | 0.10 | 0.10 |
| lineAndWash:paperColdPress | dry.wcDryBrush | 0.27 | 0.27 | 0.07 | 0.07 |
| lineAndWash:paperColdPress | edge.wcEdgeDarken | 0.21 | 0.21 | 0.05 | 0.05 |
| lineAndWash:paperColdPress | grain.wcGranulation | 0.23 | 0.23 | 0.07 | 0.07 |
| lineAndWash:paperColdPress | line.inkPenLine | 1.19 | 1.19 | 0.15 | 0.15 |
| lineAndWash:paperColdPress | paper.paperColdPress | 0.14 | 0.14 | 0.03 | 0.03 |
| lineAndWash:paperColdPress | paper.paperColdPress.field | 0.50 | 0.50 | 0.13 | 0.13 |
| lineAndWash:paperColdPress | paper.paperColdPress.paper | 0.50 | 0.50 | 0.13 | 0.13 |
| lineAndWash:paperColdPress | wash.wcWash | 1.12 | 1.12 | 0.20 | 0.20 |
| lineAndWash:paperCanvas | bleed.wcBleed | 0.40 | 0.40 | 0.10 | 0.10 |
| lineAndWash:paperCanvas | dry.wcDryBrush | 0.28 | 0.28 | 0.08 | 0.07 |
| lineAndWash:paperCanvas | edge.wcEdgeDarken | 0.21 | 0.21 | 0.05 | 0.06 |
| lineAndWash:paperCanvas | grain.wcGranulation | 0.23 | 0.23 | 0.07 | 0.07 |
| lineAndWash:paperCanvas | line.inkPenLine | 1.20 | 1.20 | 0.16 | 0.16 |
| lineAndWash:paperCanvas | paper.paperCanvas | 0.14 | 0.14 | 0.03 | 0.04 |
| lineAndWash:paperCanvas | paper.paperCanvas.field | 0.57 | 0.57 | 0.15 | 0.15 |
| lineAndWash:paperCanvas | paper.paperCanvas.paper | 0.56 | 0.56 | 0.15 | 0.15 |
| lineAndWash:paperCanvas | wash.wcWash | 1.13 | 1.13 | 0.20 | 0.20 |
The fresh-install runs are in `design/lookbook/paper_fresh_runs.txt` in the Style Cookbook world.