Panel Core#

The resolved intermediate format, between the authored source and emitted SVG.

*.panel.yaml   →   IR   →   *.core.json   →   *.svg
authored           validated   resolved         rendered

Why this tier exists#

Vega-Lite compiles a high-level grammar into a lower-level one before emitting SVG, rather than going straight to pixels. The same split pays for itself here, in four ways:

  1. Tests target stable numbers, not markup. Golden-file tests compare Panel Core JSON, which changes only when layout genuinely changes. Diffing SVG text is brittle: reordered attributes or a changed path-rounding convention produce enormous diffs that mean nothing.

  2. Layouts can be adjusted by hand. An artist who wants one balloon two millimetres left can edit the Core file, without touching the source or re-running the solver.

  3. One input, several emitters. SVG, the debug overlay, and later formats all consume the same resolved tier.

  4. The high-level grammar stays a strict subset of the low-level one, so nothing expressible in the source is inexpressible after resolution.

Properties#

  • Fully numeric. Every position, size and scale is an absolute number in panel units. No left_third, no medium_shot — those are resolved away. The camera records the shot it was asked for beside what it made of it, but nothing is drawn from the name.

  • Still named. Actors, balloons, captions and anchors keep their source identifiers, so a Core file is readable and diffable. This is what separates it from SVG.

  • Deterministic. Floats are rounded to a fixed precision on emission, so identical input yields byte-identical Core output on any platform.

Shape#

{
  "panel": { "width": 1000, "height": 1000 },
  "actors": [
    {
      "id": "alice",
      "reference": "alice",
      "pose": "pointing",
      "transform": { "x": 208.33, "y": 140.0, "scale": 0.72, "mirrored": false },
      "anchors": { "mouth": [263.1, 268.4], "eyes": [259.0, 251.2], "feet": [270.0, 680.0] },
      "face_exclusion": { "cx": 259.0, "cy": 255.0, "r": 46.8 },
      "gaze": [1.0, 0.0],
      "hull": [[180.2, 140.0], [340.5, 140.0], "..."]
    }
  ],
  "balloons": [
    {
      "id": "b0",
      "speaker": "alice",
      "order": 1,
      "kind": "speech",
      "box": { "x": 60.0, "y": 55.0, "width": 330.0, "height": 118.0 },
      "lines": ["You forgot your", "umbrella!"],
      "font_size": 28.0,
      "line_height": 35.0,
      "tail": { "start": [225.0, 173.0], "end": [263.1, 268.4], "control": null, "width": 14.0 }
    }
  ],
  "captions": [
    {
      "id": "c0",
      "order": 0,
      "kind": "locale",
      "box": { "x": 19.0, "y": 19.0, "width": 147.0, "height": 96.0 },
      "lines": ["Midnight.", "The docks."],
      "italic": true,
      "fill": "#ffffff",
      "ink": "#111111"
    }
  ],
  "backdrop": {
    "horizon": 341.0,
    "seed": 3062117294,
    "masses": [
      {
        "id": "m0",
        "kind": "sky",
        "plane": "far",
        "depth": -3,
        "tone": "#4d4d4d",
        "polygon": [[0.0, 0.0], [1000.0, 0.0], [1000.0, 341.0], [0.0, 341.0]]
      }
    ],
    "atmosphere": {
      "time": "night",
      "weather": "rain",
      "tone": "#4d4d4d",
      "veil": { "tone": "#222222", "opacity": 0.3, "frequency": 0.0016, "octaves": 3, "seed": 52939 },
      "streaks": [{ "start": [88.4, 12.0], "end": [99.7, 55.4] }],
      "flecks": [],
      "streak_width": 1.4,
      "fall_tone": "#4d4d4d"
    }
  }
}

lines is the resolved line breaking, not the source string. Wrapping is decided during compilation using real font metrics, so the emitter never re-measures and never disagrees with the solver about how wide a balloon needs to be. For a spoken caption the quotation marks are part of those lines for the same reason: marks added afterwards would not fit the box drawn for them.

A line with emphasis in it is still plain text in lines, the text as it reads. Beside lines, the balloon or caption then carries runs: every line cut wherever its face changes, so the emitter draws each run in the face it was measured in. The runs of a line must spell it exactly, and a Core whose runs do not is refused.

"lines": ["I told you it", "would rain."],
"runs": [
  [{ "text": "I ", "face": "regular" }, { "text": "told", "face": "bold_italic" }, { "text": " you it", "face": "regular" }],
  [{ "text": "would rain.", "face": "regular" }]
]

face is regular, italic or bold_italic. A balloon or caption with no emphasis has no runs key at all, so every Core written before emphasis existed reads, and is written, byte for byte as before — which is why it needed no format_version bump.

A panel lettered in an author’s face carries fonts: each face set in a font other than the shipped one, by its family and style as the font names them, and the SHA-256 of the file. Never a path — where the file was is a fact about one machine.

"fonts": {
  "regular": { "family": "Comic Neue", "style": "Regular", "sha256": "4c1f…" },
  "italic": { "family": "Comic Neue", "style": "Regular", "sha256": "4c1f…" }
}

The renderer is handed fonts by digest and refuses a Core naming one it was not given, so a Core is never drawn in a face it was not measured in. A face not listed is the shipped one; a panel in the shipped family has no fonts key at all, so no format_version bump either.

italic is recorded rather than re-derived from kind, so the emitter cannot draw a box in a face the solver did not measure it in. fill and ink are there for the same reason: a caption’s tone resolves to a value here, and the ink is chosen against it by contrast — so a dark box arrives at the emitter already lettered in paper, and the emitter is left with no decision about which mark reads. Both are defaulted to the white and near-black every caption had before tones existed, which is why adding them needed no format_version bump either.

order is one sequence across both lists. Balloons and captions are placed in a single pass in script order, so a caption between two lines of dialogue takes the number between theirs and a balloon list can have gaps in it. A panel with no captions is unaffected, which is why this did not need a format_version bump.

A balloon spoken from off panel carries off_panel, the side of the frame its speaker is on, as the solver settled it: left, right, top or bottom. When it is set, speaker names somebody who is not an actor, so it is not looked up among actors, and the tail’s end lies on the frame on that side — on outline, when the panel has one. A balloon whose speaker is drawn has no off_panel key, so no existing Core changes.

backdrop is absent on a panel that says nothing about where it is, which is why adding it did not need a bump either. Its masses carry fully resolved numeric polygons and tone values, the same discipline as capsules, blobs and face_marks: the emitter is left with no layout decision, and a backdrop can be read, diffed and hand-adjusted like everything else in this tier. depth is the same painter’s order the actors use, so masses and figures sort into one sequence.

A backdrop that is a supplied picture has an image in place of masses, and no horizon: the compiler does not know where the picture’s horizon is, and does not pretend to. The solver was told the picture’s size and nothing else; what it resolved is box, the rectangle the picture is drawn in — centred, overhanging the panel under cover, inside it under contain. Change box and the picture moves.

"backdrop": {
  "seed": 3062117294,
  "image": {
    "href": "art/p01-street.png",
    "media_type": "image/png",
    "sha256": "9f2c…",
    "pixels": [960, 640],
    "fit": "cover",
    "preserve_aspect_ratio": "xMidYMid slice",
    "box": { "x": 0.0, "y": -100.0, "width": 1200.0, "height": 800.0 }
  }
}

href is the path exactly as the document wrote it, relative to the document: the one fact about where the file was that reaches any output, because a linked SVG needs it. sha256 is what the picture is known by, as a font is — a renderer asked to embed it is handed bytes by digest and refuses a Core naming bytes it was not given. preserve_aspect_ratio is the fit in SVG’s own words, recorded so the emitter copies it. A backdrop of masses has no image key, and every Core written before pictures existed reads and writes byte for byte as it did, so neither image nor a missing horizon needed a format_version bump.

A picture that draws its own people has an artwork contract beside it, and what the contract declares is recorded here in the panel’s units, mapped from the picture’s pixels through the same fit as box. The people are figures, beside actors; the regions to stay off are the picture’s keep_out.

"figures": [
  {
    "figure": "ana",
    "actor": "ana",
    "face_exclusion": { "cx": 413.33, "cy": 103.33, "r": 50.0 },
    "mouth": [423.33, 133.33],
    "body": [[350.0, 73.33], [483.33, 73.33], [500.0, 566.67], [333.33, 566.67]]
  },
  { "figure": "ben", "face_exclusion": { "cx": 736.67, "cy": 93.33, "r": 46.67 } }
],
"backdrop": {
  "image": {
    "...": "...",
    "keep_out": [[[133.33, 300.0], [300.0, 300.0], [300.0, 466.67], [133.33, 466.67]]]
  }
}

A figure is not an actor: nobody posed it, so it has no transform, no skeleton and nothing to draw — the picture draws it, and an emitter draws nothing for it. What it has is what lettering worked around. figure is its name in the contract; actor is the cast member it plays, which a balloon’s speaker and a sound effect’s source name, and is left out for a figure nobody cast, whose face lettering kept clear of just the same. mouth is where a tail to it aims, left out when the contract declares none, so a tail aims at the face’s centre; body is left out when it declares none. keep_out holds what balloons, captions and sound effects paid to cover and were never forbidden to. They are resolved numbers, like everything here: change box by hand and the picture moves, and the figures stay where they were resolved.

A panel with no figures has no figures key, and a picture with no keep-out regions no keep_out, so every Core written before them reads and writes byte for byte as it did, and neither needed a format_version bump.

An actor with marks carries three more fields. marks names what was asked for; emanata is the drawing, in the same stroke and disc primitives as face_marks, with ids numbered per mark (plewd_0, plewd_1, …); and emanata_zones holds one polygon per mark, which is what balloon and caption placement paid to avoid. The zones are recorded so a Core document says why a balloon is where it is. All three default to empty, so an actor without marks serialises with [] in each and nothing else changes — which is why they needed no format_version bump either.

"marks": ["plewds"],
"emanata": [{ "mark": "stroke", "id": "plewd_0", "points": [[171.2, 98.6], "..."], "width": 5.9, "closed": true }],
"emanata_zones": [[[150.3, 61.0], [205.8, 70.4], "..."]]

A panel with sound effects carries sfx, one entry per effect in script order: its id (s0, s1, …), its order in the one sequence it shares with balloons and captions, its text, its source, its font_size, and every letter as a glyph — the character, the origin its baseline starts at, and the angle it is turned by. The arc and the rotation are already in those numbers, so the emitter draws each letter’s outline where it is told. zone is the footprint balloons and captions paid to avoid and no face was allowed inside; fill, stroke and stroke_width are resolved, with fill none for a hollow effect.

"sfx": [{
  "id": "s0", "order": 0, "text": "KRAK", "source": "bob", "font_size": 106.4,
  "glyphs": [{ "char": "K", "origin": [716.07, 486.47], "angle": -14.0 }, "..."],
  "zone": [[688.18, 392.21], "..."], "fill": "#111111", "stroke": "#ffffff", "stroke_width": 8.51, "hollow": false
}]

A panel with no effects has no sfx key, so no format_version bump.

seed is kept so that a Core document explains itself: two panels with the same masses and different skylines differ here, and here is where to look.

A panel read right to left carries "direction": "rtl", the reason its lettering is where it is: see direction. A panel read left to right has no direction key, so every Core written before it reads, and is written, byte for byte as before — no format_version bump.

The staging#

What the solver used and once threw away is recorded too, so that an export for image models, reading only this tier, can say how the panel is staged (#72). From examples/duel.panel.yaml, and a backdrop set in a street at night:

"camera": { "shot": "medium_shot", "angle": "eye_level", "reference": "alice", "scale": 1.94,
            "headroom": 0.1, "footroom": 0.0, "pullback": 0.71 },
"actors": [{
  "id": "alice",
  "joints": { "neck": [96.87, 355.74], "shoulder_l": [31.0, 367.37], "...": "..." },
  "features": { "eye_l": [58.12, 173.62], "eye_r": [135.62, 173.62], "nose": [96.87, 204.62], "...": "..." },
  "looking_at": "bob",
  "...": "..."
}],
"backdrop": { "place": "street", "time": "night", "...": "..." }
  • camera is the camera’s verdict: the shot asked for and the angle, the puppet it was composed on, the one scale every actor shares, the headroom and footroom as fractions of the panel’s height, and pullback, below 1 when the camera retreated to fit the cast. A panel whose cast a picture draws has none: the artist framed it.

  • joints is where every joint of the skeleton landed, by the puppet’s name for it. _l and _r are the sides of the screen in the puppet’s rest pose, so on a figure facing left shoulder_l is on the right.

  • features is where each face feature point sits: the anatomy, not the marks, so the same whatever the expression and present when the face is too small to draw.

  • looking_at is whom the actor looks at — an actor id, or somebody heard from off panel — where gaze_aim is the direction.

  • place names the place the masses are, whether place: wrote them or an author wrote the same list out; time is the hour the values were chosen for, which atmosphere records too but only when something is in the air. Over a picture neither is known, and both are left out.

None is read to draw the panel. Each is left out of a document that does not use it, so a Core written before them reads and writes back byte for byte — no format_version bump.

The words#

A panel is also read, by somebody who cannot see it, and the words it is read in are resolved here like everything else, so the emitter copies them and decides nothing:

"description": "Medium shot. A street at night, in the rain. Alice, left, pointing, angry, looking at Bob; Bob, right, arms crossed, sweating.",
"balloons": [{ "id": "b0", "order": 1, "label": "Alice: You forgot your umbrella!", "...": "..." }],
"captions": [{ "id": "c0", "order": 0, "label": "Caption: Midnight.", "...": "..." }],
"sfx": [{ "id": "s0", "order": 2, "label": "Sound effect, from Bob: BLAM.", "...": "..." }]
  • description is the panel’s description when it has one — written, or a comic script’s prose — and otherwise generated from what it holds: the shot and angle, the place, the hour and the weather, then each figure in reading order (from the right in a panel read right to left), by thirds of the panel at the centre of its face, with its pose, expression and marks, whom it looks at, and whom it stands in front of or behind. A panel with nothing in it is An empty panel.

  • label, on every balloon, caption and sound effect, is who says it, how, and the words: Bob, whispering, quietly: I know., Caption, inner voice: …, Carol, off panel: …. The words are the author’s with emphasis markers off — never lines, which are broken to fit and, on a spoken caption, carry quotation marks that are lettering, not speech. Each is a sentence: a line that does not end gets a full stop, so two lines are heard as two.

Read in order, the labels are the panel’s text in the order its reader takes it, sound effects included. Generated text is templated English; a French script’s prose stays French, but the words round its dialogue do not. Both are left out of a document that has none, so a Core written before them is still valid and writes back byte for byte — no format_version bump — and neither may hold a character an SVG cannot, since Core is hand-edited and both go into the SVG.

Determinism is on this text, and on the SVG text — not on pixels#

The atmosphere’s veil is a set of feTurbulence parameters rather than an image, and the filter is reproducible by specification. What browsers paint from it agrees only approximately. See the language reference, where the boundary is stated in full; golden-file tests target this tier and the SVG text, never a raster.

Page Core#

A scene with pages: gets one more document per page, alongside each panel’s Panel Core. It holds where each panel’s frame is, and nothing drawn:

{
  "format_version": 1,
  "width": 1500.0,
  "height": 2250.0,
  "lettering_height": 700.0,
  "frames": [
    { "panel": "establishing", "x": 75.0, "y": 75.0, "width": 880.0, "height": 628.12 },
    { "panel": "reply", "x": 985.0, "y": 75.0, "width": 440.0, "height": 628.12 }
  ]
}

frames are in reading order: tier by tier, left to right — right to left, for a page whose direction is rtl — and down each column of a tier of columns, with an inset right after the panel it is set into or, for read: before, right before it. An inset’s frame also carries inset_of, the panel it sits in, and clearance, the area it covers with its ring of white; the page paints that white under it. Neither field is written for a panel that is no inset.

The parent’s own Core then carries exclusions: each inset’s clearance, moved into the parent’s coordinates, which no balloon or caption was allowed to touch. Nothing else in the parent moved for it, so the art under an inset is exactly the art of the panel compiled alone. A panel with no inset has no exclusions key at all, so every Core written before insets existed reads, and is written, byte for byte as before.

A panel in a slanted tier carries outline, its four corners clockwise from the top left, on its frame in page units and in its own Core in panel units; the frame is the outline’s bounding box. The panel is drawn clipped to it, and bordered by it. A rectangle has no outline key. Each panel’s own Core was compiled at exactly its frame’s width and height, so a panel’s Core reads the same on a page as it would alone at that size, and a page adds no layout decision to a panel.

A page read right to left carries "direction": "rtl" as its panels do; one read left to right has no direction key.

lettering_height is the one thing every panel on the page shares: the height their type sizes are a fraction of, in place of each panel’s own. It is recorded so that a Page Core says why two panels of different heights letter at the same size.

scenet build --core writes it as stem.page-<n>.core.json; compile_book returns it as Book.pages.