Skip to content

v0.12.0

Choose a tag to compare

@github-actions github-actions released this 29 Oct 20:01
· 99 commits to main since this release
6927378

Upgrading from 0.11.1.

  • A font path written with \ between folders is refused, on Windows as everywhere else:
    write /. So is one holding a control character, which used to crash build and check.
  • conflicting-setting covers a picture too: an image beside a place, masses or a
    horizon. Its SARIF description and help say so; a place beside masses is reported exactly
    as before. A panel that inherits a place and names an image is refused the same way; it
    writes place: null beside the image.
  • A comic script's prose is now read, as its panel's description. A line of it holding a
    control character an SVG cannot hold, which used to pass unnoticed, is reported at that
    line as invalid-field. description in a script's front matter is refused.
  • Every Panel Core gains description, and each balloon, caption and sound effect a
    label.
    Nothing else in it moves, and format_version stays 1: a Core without them is
    still valid and writes back byte for byte, but golden Core files of your own will need
    regenerating once. 0.11 refuses the new keys.
  • Every Panel Core gains camera; each actor joints, features and, when it looks at
    somebody, looking_at; a backdrop of masses time and, for a named place, place
    (#72).
    Nothing else in it moves, no SVG changes, and format_version stays 1: a Core without them is
    still valid and writes back byte for byte, but golden Core files of your own will need
    regenerating once. Every gallery Core was compared with its 0.11.1 golden and differs by these
    keys alone. A panel whose cast a picture draws has no camera, and 0.11 refuses the new keys.
  • A balloon over its own speaker's body costs what it always should have when the actor id
    is not the puppet's name (ALICE: {reference: alice}, carol: {reference: alice}). Such a
    speaker was charged as a bystander, so balloons drifted away from them. Three gallery
    panels letter differently, each balloon now beside its speaker: regenerate golden Cores of
    your own that cast an actor under another name.
  • CaptionEvent.prefer is None when it was not written, where it used to read top_left:
    an unwritten zone is where reading starts, which the solver decides from direction. A caption
    is placed exactly where it was. Code reading it from the Python API takes None as top_left
    for a left-to-right panel.
  • CastMember.reference may be None, for a member that is a figure a picture draws, and
    CompileResult.camera is None for a panel whose cast is all figures. Code reading either
    from the Python API narrows them first; PanelIR.puppets() gives every puppet's reference.
    A member that names neither a puppet nor a figure is reported exactly as before.
  • Every SVG's bytes change once, to carry what a screen reader reads: a role and a <title>
    per panel, and aria-hidden on a page's own shapes. Nothing is drawn
    differently and lettering stays outlines. --no-aria (aria=False) gives back exactly the
    SVG written before, for a pipeline that post-processes it; debug overlays are unchanged.

Added

  • Export a panel's staging for image models (#72): scenet build --conditioning writes five
    files beside each panel's SVG — the pose as OpenPose COCO-18 JSON, a phrase and a box per
    figure as GLIGEN takes them, a depth map (stepped grey planes in painter's order, white near),
    an instance mask under the panel's own ids, where a character keeps its colour across panels
    unless another in the same panel has the same hue, and the semantics: camera, setting, cast
    with each character's mask colour, and lettering, as structure and never as a prompt. No model runs in Scenet: each is one more
    emitter reading Panel Core, text, JSON or SVG, byte-identical like the rest, and nothing is
    rasterised. The pose names left and right as OpenPose does, by the subject's side; a point
    outside the panel, an ear, or a joint a puppet does not name is 0, 0, 0. Figures a picture
    draws have boxes and masks and no pose; a picture is the far plane; a panel read right to left
    exports the same pose, boxes, depth and mask. render_conditioning and Conditioning join
    scenet.__all__; MCP render takes conditioning (still five tools); the playground offers
    the five downloads for a one-panel document. See the how-to, "Export a panel's staging for
    image models".
  • direction: ltr | rtl, which way a comic is read (#68): rtl for manga, read from the
    top right. It is the comic's progression, not its text's, and EPUB's word for it. Written once
    for the whole document — at the top of a panel, beside panels:, or in a comic script's front
    matter — and inherited by every panel; a panel's own, and a script's @direction:, are the new
    rule panel-direction. The editor's schema offers it. Panel Core and Page Core record it when it
    is rtl; a left-to-right Core is byte-identical and format_version stays 1.
  • A panel read right to left is lettered from the top right. A balloon or caption may never sit
    above and right of one before it; a caption with no prefer, and a voice from off panel with no
    side named, start top right; every tiebreak between equal placements is mirrored, so the mirror
    image of a panel is lettered as the exact mirror of the original; and a generated description
    names the cast from the right. Nothing in the picture moves: figures, gaze, tails and sound
    effects are byte-identical, and prefer:, at: and named sides keep meaning places on the panel.
    The debug overlay says "read right to left".
  • Pages and strips read right to left. Under direction: rtl a tier's first panel, or first
    column, is at its right; each frame is the one the tier reversed would have, so every panel
    keeps its size and its art, an inset keeps its corner and a slant leans as written. The pages of
    a document side by side (render_pages, the playground) and a strip start at the right, and every
    one is still written in reading order. A strip or a row of pages read two ways is refused.
  • Two gallery pages, the same manga page read right to left and left to right.
  • setting: {image: ..., fit: cover | contain}: a picture drawn elsewhere, behind the
    cast (#70). It is a backdrop of its own, so it is written instead of a place or
    masses, never beside them or beside a horizon (conflicting-setting); a fit with
    no image is invalid-field. A panel that inherits a place writes place: null beside
    its image, and the editor's schema now accepts that null.
  • The picture is read from its PNG or JPEG header, beside the document as a font is, and
    only its size is taken. scenet check reports a picture that is absolute, missing, not a PNG
    or JPEG, or turned by its EXIF orientation as the new rule image-file, without the solver.
    Text compiled with no folder may be handed the files it names instead: base takes a
    mapping from the path the document writes to the file's bytes, for fonts as for pictures.
  • The picture is placed, and Panel Core records it. Centred, filling the panel (cover)
    or shown whole inside it (contain), never stretched. backdrop.image holds the path as
    written, the media type, a digest of the bytes, the pixel size, the fit in SVG's words and
    the rectangle it is drawn in; over a picture backdrop.horizon is left out. A Core of masses
    is byte-identical, and format_version stays 1. The picture costs a balloon nothing and does
    not move the camera, and a generated description names no place and no hour over it — only
    the weather drawn on top.
  • The picture is drawn, as an <image> on the paper, under every figure and the weather,
    in the rectangle Panel Core placed it in. It is linked by xlink:href — Inkscape does not
    load plain href — to the path exactly as written, which resolves when the SVG is beside
    the document; scenet build --embed-images (embed= on every renderer) writes its bytes in
    as a data: URI instead, and build says so in a note when a link will not resolve. The
    debug overlay shows where it was placed. A new how-to, Letter over your own art, and a
    gallery example drawn over a picture the repository paints itself, under its own licence.
  • The MCP server's validate, compile and render take files: the files a document
    names, by the path it writes, as base64. They stand in for the document's folder, read in
    memory, so a model can letter over a picture — or in a font — it sends; render embeds
    the picture. Still five tools.
  • The playground's Image… button adds a .png or .jpg to the document's art/ folder,
    for as long as the tab is open, and the preview and every download embed it. The VS Code
    preview embeds pictures too, since it is not drawn beside the document.
  • An artwork contract declares the figures a supplied picture draws (#71): beside
    art/p01.png, art/p01.art.yaml gives each figure's face (a disc lettering may not
    cover), mouth (where a tail ends) and body, and keep_out regions lettering should stay
    off, all in the picture's own pixels, with the size they were measured on.
    scenet schema --artwork emits its JSON Schema, which the VS Code extension associates with
    *.art.yaml and the documentation site publishes beside the others. Panel Core records what a
    contract declares, mapped into the panel through the picture's fit: figures, beside
    actors, and the picture's keep_out. Both are left out when empty, so every Core reads and
    writes as before and format_version stays 1; the debug overlay draws each figure's face,
    mouth and body, and each keep-out.
  • Figures a picture draws, in the cast (#71): cast: {ana: {figure: ana}} in place of
    reference:. The panel's picture's artwork contract says where each face, mouth and body is,
    and lettering is placed by the same rules as around a puppet: never over a declared face —
    anybody's in the picture, cast or not — tails aimed at the mouth, bodies and keep-out regions a
    cost. The camera and staging do not run. A cast figure a panel's picture does not draw is not in
    that panel. Four new rules, each from the cheap pass: unknown-figure, artwork-contract
    (reported in the contract when the document is checked as a file), conflicting-cast (a puppet
    and a figure at once, a puppet's key on a figure, a mixed cast, one figure played twice) and
    figure-staging (staging naming a figure, a camera over figures). A cover crop that cuts a
    speaking figure's mouth out of the panel ends its tail on the frame, with a note. scenet check
    and build refuse a *.art.yaml given as a document. The panel and scene schemas accept
    figure.
  • A gallery example of figures a picture draws, 35-figures.scene.yaml: a station
    platform painted by scripts/make_gallery_figures.py, with its artwork contract written from
    the same constants, lettered under cover and contain. Two of its three people are cast,
    the third is kept clear of all the same, and its timetable board is a keep-out.
  • Contracts travel with their picture. The playground's Image… button adds a picture's
    .art.yaml beside it, cleaned to the same name so the two pair; the MCP server takes it in
    files, its instructions and get_spec's characters part say so, and the spec pack prints
    the gallery's contract. Still five tools. A how-to, Letter the people in your own art.
  • A panel's description: what it shows, in words, for a reader who cannot see it.
    Written as description: in YAML; in a comic script, the panel's prose is it — its lines
    one paragraph, Markdown's markers taken off — unless an @description: directive, kept
    exactly as written, says something else. It is carried and never interpreted, for
    accessible output to read (#76). It describes one panel, so it is never inherited: beside
    panels: or in a script's front matter it is refused as unknown-key, and a panel drawn
    over: another does not take its parent's.
  • Panel Core carries the words a panel is read in, built in the pipeline so the emitter
    copies them. description is the panel's own, or generated from what it holds — Medium shot. A street at night, in the rain. Alice, left, pointing, angry, looking at Bob; Bob, right, arms crossed, sweating. — and every balloon, caption and sound effect has a
    label: who says it, how, and the words, Bob, whispering: I know. Read in order, they
    are the panel's text as its reader takes it. Generated text is templated English.
  • A screen reader can read the panel (#76). Each panel is one image, role="img", named
    by one <title>: the description, then every balloon, caption and sound effect in reading
    order — in the name, since a description alone is often not spoken, and nowhere else, since
    NVDA read a <desc> a second time. It needs no id, so panels inlined together in one page
    cannot name each other. A strip or a page is a named role="group" of its panels, in reading order.
    scenet build --no-aria, aria=False on every renderer and aria: false on the MCP server's
    render leave it all out. Checked with NVDA 2026.2 and Edge 155, and against Edge's UI Automation
    tree; results, and how to check with VoiceOver, are in the new how-to, Make a panel a
    screen reader can read
    .

Changed

  • A page's SVG writes its panels in reading order, insets included, since document order is
    what a screen reader follows. An inset used to be painted after every other panel to stay on
    top; now its parent is cut around the inset's ring instead, so an inset read: before its
    parent comes first and is still on top. The page looks the same, to within anti-aliasing along
    the rings; only a page with insets changes, and only its bytes.
  • The strip and the page place a panel's body, built for them, instead of cutting it out of
    the panel's serialised SVG, which made the panel's first two lines something they silently
    depended on. Byte for byte the same.

Fixed

  • A backdrop of more than ten masses is drawn back to front. Mass ids are numbered back
    to front, and the painter's sort compared them as text, so m10 was drawn before m2: ten
    walls and a window in one plane painted the window, one rung paler, out under the walls. Masses
    in one plane are now drawn in the order Panel Core lists them. No gallery panel has more than
    ten, and none changes.
  • A speaker is forgiven a balloon over their own body whatever the actor is called. The
    placer half-weights the speaker's silhouette, and looked the speaker up by the puppet's name
    while the silhouettes are keyed by actor id. An actor called anything but its puppet, which
    every comic script's capitals are, was charged as a bystander: in the gallery's comic script
    a line moved to the far corner of the panel behind a tail bent outside it.
  • A font path with a backslash is refused on every system, as font-file. Windows read
    fonts\Comic.ttf as a folder and a file, and every other system as one file name, so the same
    document lettered in its font on one machine and was refused on the next. Write / between
    folders.
  • A control character in a font path is a finding, not a crash. lettering: {font: "fonts/a\0.ttf"} reached the file system, whose ValueError was not one the compiler
    reports, so scenet build and scenet check both ended in a traceback. Any control
    character in a declared path is now font-file, naming the character.
  • Mutation testing runs again. mutation.yml installed only the mutation group, and the
    documentation site's tests need docs, so mutmut's clean run failed and no mutant was tried.
    It now installs every group, as CI's test jobs do.
  • The VS Code preview draws each figure with its own clip. It inlines every SVG a build
    wrote into one page, and a page has one set of ids, so a second panel in a slanted tier was
    clipped by the first one's outline, and page 2's panels by page 1's frames. Each figure's
    ids, and every reference to one, now carry a prefix of their own.

Install

pip install scenet==0.12.0

The VS Code extension is the attached scenet.vsix:
Extensions -> ... -> Install from VSIX...

Documentation |
Playground