Each rib has its own width, depth and crest radius. Smooth transitions join its neighbours, while a consistent angular layout carries each groove from crown to base.

The generator describes three boundaries: outer rind, flesh beneath the skin and interior cavity. Each can vary independently, giving the crown, sides and base the thickness they need for cutting, scooping and light.

A saved recipe records proportions, ribs, surface settings, pattern seed and stem choices. It reproduces the same pumpkin through later tools. The same generator supports the flatter, taller, pear-shaped and uneven forms now seen in the app’s patch.

A seed is a repeatable starting number for a procedural pattern. The rind’s patternSeed accepts 4,096 values. In Colour mode, tapping the pumpkin advances the seed and lays out a new mottle pattern while keeping its shape and other recipe fields. Coverage, scale and contrast remain separate choices.

The stem has its own seeded appearance: four saved variants pair a stable grain seed with a dryness setting. A different rind pattern therefore does not unexpectedly reshape the stem or change its finish. These seeds select repeatable detail; the broader silhouette is described by explicit shape controls.

A few fields from the blue-grey pumpkin opened in the current app:

{
  "version": 3,
  "height": 0.9,
  "ribCount": 10,
  "stemWidth": 1.8,
  "stemVariant": 3,
  "patternSeed": 12
}

The full link carries every recipe field. Open this saved recipe in the app ↗

Pull, stretch, widen, lean or flatten the pumpkin while watching its silhouette. Height, fullness, bulges and rib expression become controls on the object.

The inner layers follow the same recipe. A volume-preserving height edit compensates in width; a broader shape must still leave a usable crown and cavity.

The generator rebuilds from the recipe instead of accumulating mesh edits. Before carving, the shape is frozen. Changing it afterwards would require rebuilding and replaying the cuts.

Broad pigment changes, elongated mottles, flecks and wax variation work at different scales. Fine relief changes the shading normal; actual geometry gives the ribs their shadows.

The shaders sample a fixed reference position, restPosition, so the texture stays aligned when a carved piece moves. Surviving faces keep their texture attributes; new cut walls get a separate flesh material.

Derivative-based filtering fades details smaller than a screen pixel, reducing shimmer. The materials use Three.js Shading Language node shaders, with values authored for this pumpkin family.

A small simulation moves connected guide particles. Distance and bending constraints give them structure; breakable attachments hold pulp to the wall. Many visible fibres follow those guides, separating visual detail from simulation cost.

The app uses a custom XPBD-style constraint solver with small substeps. Compliance lets connections yield; damping, bending limits and controlled tearing make fibres peel, bunch and settle. The established methods are XPBD and Small Steps in Physics Simulation.

Rapier 0.21 was explored separately for deformable pulp and rigid seeds. It remains an optional lab engine. The shipping scoop uses the custom solver.

The scoop, pulp and seeds must respect the cavity and actual opening, even during a fast gesture. When scooping removes flesh, it asks the shared carving system to update the wall too.

A checked outline becomes a signed distance field: a function describing what lies inside or outside the cut. It defines a three-dimensional tool. Manifold subtracts that tool from the remaining shell and intersects it with the shell to create the plug.

Draft is the inward cut angle; kerf is the gap left by the blade. Together they let the plug move through its opening. Overlap checks at five pull positions catch trapped pieces, though they do not prove continuous clearance.

Later cuts use the retained shell. Surviving faces keep their material coordinates; new walls inherit skin fraction, flesh depth, thickness and cut direction from the paired surfaces.

Cutting and removal are separate history events: undo reseats a piece before restoring its wall. Mirrored pieces share a cut but have separate removals. Validation at display precision catches thin-geometry failures; refusal preserves the pumpkin and history.

The studies below test the knife-cut finish, where mirrored cuts can go, how a pupil stays attached and what causes repeated corner fins. The images below are recorded lab results.

This was progress toward a hand-cut finish, with visible limits: fibres could bunch into clumps, facets interrupted the walls and some flesh stayed too dark under the interior lamp. The captures record those findings rather than a finished claim of realism.

One sampled rear gap shrank from 0.170 to 0.089 to zero in scene units. Connectivity does not establish bridge strength or transmitted light. A separate ring cut in a solid test block was refused; adding a bridge restored one connected result.

Trace cleanup fills internal holes while preserving outer concavities. The final solid still needs its own check. An explicit floating-island option was considered; it remains unimplemented and would need separate object and physics behavior.

The correction also survived a plain clay view, pull-clearance checks and saved-cut replay. Existing cuts keep their geometry version, so a cutter update cannot silently change earlier work.

Shaving leaves flesh behind. The slab study below isolates removal depth, exposed layers and light.

Try the scraping lab ↗

The backlight test reads that same thickness data: intact skin stays dark while thinner flesh brightens. A heightfield cannot represent an overhang or a scoop beneath intact skin. A three-dimensional distance field could retain an undercut; a pushed mesh stretched and lost volume. The heightfield was the cheapest useful option for shallow work. Joining that detail to curved-shell solid cuts remains separate integration work.

Transmission reads the actual point light, wall geometry, skin fraction and thickness. Colour-dependent exponential attenuation, from the Beer–Lambert family, absorbs more light along longer paths. Thin flesh glows more strongly; thick flesh darkens and shifts colour.

A through-cut exposes the cavity. A shave leaves a translucent wall. Moving the light changes both through the material model.

A shared diffuse cavity approximation adds bounced light to the lining, cut walls and shaves. It uses source power, surface area, opening area and authored albedo to estimate repeated illumination.

Local transmission and simplified cavity lighting make depth readable in real time. Full subsurface scattering and ray-traced global illumination remain outside this model.

A shared recipe, material coordinates, cavity queries and undo history connect the systems. The interior loads when scooping needs it; geometry work runs in a worker. The first interaction should not wait for every tool.

A link can carry the design.

The share payload includes the recipe, lid and carving outlines, removed-piece choices, paint strokes, lighting and display setting. The receiving app rebuilds from that description. Heavy meshes and the live simulation stay local.

Sharing supports a compact server-backed link, with a compressed, self-contained link as a fallback. The latter stores its data after the URL’s #. Opening a shared pumpkin creates a separate visiting draft, preserving the recipient’s own saved pumpkin.

Payloads have explicit versions and upgrade steps. Later engines replay the saved outlines against the recipe, so the design remains editable without promising identical pixels across future renderer changes.

The studio and lab captures record different stages of a developing project. Next are clearer motion sequences, broader geometry tests and continued touch and browser testing. The aim stays simple: handle the mechanics so the maker can shape the pumpkin and its expression.