Design 3d space

A simplified 3D space to draw in: planes, parallax first, orbit second

Status: designed, nothing built. Decided with the owner on 2026-08-12 (Q72), the same day the infinite canvas was removed (Q71) — this is the direction that removal made room for. Nothing here exists in the tree yet; the roadmap items under Camera and scene carry the evidence anchors that will say when it does.

What it is

A document whose drawings live on planes in a 3D space: each drawing is a flat canvas with a position and (eventually) an orientation, the artist can rotate and zoom around the arrangement while working, and a camera renders it. There are no meshes and no 3D strokes — the unit of art stays a 2D drawing, exactly as the tool understands it today.

The decided constraint, and the reason the whole thing is cheap relative to what it looks like:

Strokes never learn 3D exists. A stroke is today's record, in plane-local 2D coordinates. The brush engine (the one pixel path), replay, undo, fills, selections, and the AI inbetweener are untouched. 3D is an arrangement of drawings, not a new kind of drawing.

The rejected alternative — stroke points carrying x,y,z — was priced and declined: it rewrites the stroke record (invariant 1's format), BrushEngine (invariant 1's only door), hit-testing, and the AI payload, to buy lines that bend through space, which the reference workflow (Grease Pencil) mostly does not use either.

The two stages, in order

Stage 1 — multiplane. Layers gain a depth: parallel planes stacked away from the picture plane, classic Disney multiplane. The camera stays the 2D record it is today (CameraFraming: x, y, zoom, roll); parallax falls out of depth-dependent response to camera moves — a pan moves a distant layer less, a zoom scales it less, by the perspective factor f / (f + depth). No orbiting in stage 1: that is the cost of shipping the small half first, chosen knowingly (Q72 — the recommendation was free planes first, and the owner picked the smaller, sooner ship).

Stage 2 — free planes and orbit. A layer's depth generalises to a pose (position, orientation, scale); the working view gains an orbit — rotate and zoom around the arrangement while drawing; the camera gains an authored 3D pose of its own. Stage 1's record is designed as the degenerate case of stage 2's so there is no migration cliff: depth: d reads as a pose translated d along the view axis, and a document authored in stage 1 opens unchanged in stage 2.

The rules it must keep

1. Optional means absent. A layer at its default depth writes no key; a document that never uses depth or pose serialises byte-identically to today. No 3D UI exists until the artist asks for it. This is the camera's precedent, applied again — and it is testable the cheap way: Assert.DoesNotContain("\"depth\"", json) on a fresh document.
2. Orbit is view-only (invariant 5). Rotating and zooming while working is navigation: never serialised, never in an export. What renders and exports is the authored camera. A document with no camera shows its planes head-on and behaves exactly like today — which also means depth without a camera does nothing, so multiplane never taxes an asset document and never conflicts with sprite export. No feature conflict is needed anywhere in this design.
3. The stroke record is the document (invariant 1). A plane's pixels are derived by exactly today's pipeline; the pose is document data about the plane, never about the strokes on it.
4. Render bigger by scaling the surface, never the geometry (invariant 7). Projection happens to the plane's rasterised surface, not to stroke coordinates — a plane is rasterised flat at a resolution chosen from its screen footprint, then drawn through a projective matrix (Skia's 3×3 carries perspective). Hash01-seeded dab dynamics are therefore identical at every viewing angle, which is what keeps a turntable of the scene from boiling.
5. No randomness (invariant 2). Projection is a pure function of pose and camera; there is nothing stochastic to seed.

How drawing works when the view is tilted (stage 2)

The pointer ray is unprojected against the active layer's plane; the intersection is a plane-local 2D point; from there it is today's input pipeline, unchanged. Drawing at a grazing angle is imprecise by geometry, not by bug — the input maths must be exact and the UX may still warn. Stage 1 has none of this: planes are parallel to the picture plane and input is untouched.

Rendering and cost

Stage 1 is one extra affine per layer pass — the compositor already draws passes through optional matrices, so multiplane is a per-layer scale/translate derived from (camera, depth) where the identity is today's behaviour. It must hold the existing performance shape: parallax changes matrices, not allocation, so a camera move re-composites (which a camera move already does) and repaints nothing per-layer that a plain pan would not.

Stage 2 replaces the affine with a homography per plane and adds the footprint-driven raster scale. The known risks, written down before building:

• Sampling under projection. A tilted bitmap needs filtering; a filtered tile edge was the seam lesson of the tiled compositor. Planes are single surfaces (not tiles), so the seam risk is absent, but the filter choice must be pinned by a determinism test the way ATiledRenderIsBitIdenticalToAnUntiledOne pins the flat path.
• Depth sorting, not intersection. Planes composite painter's-algorithm by camera distance. Intersecting planes are out of scope, permanently — the simplification that keeps this a drawing tool rather than a renderer. Two planes at the same depth keep layer order, so today's documents (all planes at depth 0) keep today's compositing exactly.
• Onion skin and the light table show ghosts on the ghost's own plane; they ride the same projection and need no design of their own.

Export

Through the authored camera, as today — parallax (stage 1) and pose (stage 2) are camera-time transforms, so SequenceExporter picks them up where it already applies CameraTransform. Assets are untouched by construction: no camera, no parallax, no change.

Reach and defaults

The Q21 shape, unchanged by Q71's removal of the conflict machinery: the capability lives on the document (a depth or pose is authored data), a project supplies what a new document starts with, and no project type gates any of it. Shot-flavoured project types may default the Scene panel visible; nothing more is needed.

Not in scope, deliberately

• Meshes, lighting, materials. It is a drawing space, not a renderer.
• 3D stroke points. Priced and declined above.
• Intersecting planes. See rendering.
• Camera paths in 3D beyond keyed poses (cranes, rigs) — keyed poses interpolate exactly as 2D camera keys do today; anything richer is its own design.

Open questions

None blocking stage 1. Stage 2 will need: the pose interpolation space (quaternion slerp vs Euler — determinism says pick one and pin it), the grazing-angle drawing UX, and where the orbit gesture lives in ShortcutMap. Ask them when stage 2 starts, per the rules — not guessed here.