Editable procedural vessels, saved as native Blender Geometry Nodes.
Project website: Local preview and GitHub Pages publishing — installation, usage, a vessel gallery and downloadable models. Ready to deploy; not published yet.
This iteration replaces the defective window surrounds and cabin joins, separates wall shape from deck setbacks, adds an optional section-based bulbous bow, and supports replaceable object/collection equipment.
Requires Blender 5.2.3+; tested on 5.2.3 LTS, Windows via Arch WSL.
- Install
dist/yard-2026.11.16.zipusing Preferences → Get Extensions → Install from Disk. - Open 3D View → N → Yard, select a preset, and Generate New Ship.
- Select its carrier, root or child attachment. The panel edits the actual saved Geometry Nodes modifier inputs.
- Edit dimensions, spacing, wall shape, bulb or equipment assignments in place. Shape the superstructure by moving, scaling and rotating the deck empties, or in the panel's Decks list. No rebuild callback is needed.
- Save the
.blend. Native controls, external references and geometry continue working without the extension.
Generate New Ship adds another vessel. Use live inputs or Apply Preset to Selected Ship to revise an existing one. Presets preserve custom nodes, material inputs, external asset assignments, attached objects and root transforms. Ship creation does not alter unrelated objects, units or render settings. No downloaded runtime assets or external dependencies are required. While the extension is enabled, debounced updates align deck-box displays and persistent equipment sockets after design edits; native ship geometry never depends on those display updates.
Existing 0.1/0.2 graphs are not automatically migrated. They retain their native modifier controls. The new panel reports missing inputs rather than replacing customized graphs. Create a new ship to use the revised construction. Schema-4 ships (no deck empties) are likewise not migrated.
Install dist/yard-2026.11.16-performance-v01.zip for the current performance and attachment improvements. New ships use independent, validated stock graphs automatically; incompatible Blender builds fall back to procedural construction. In the local compatible-build benchmark, CRUISE creation fell from 327.93 s to 0.97 s, with all 13 geometry cases preserved.
For faster editing, select the ship and open N → Yard → Performance: enable Viewport Preview and, for supplied native-mount assets, Instance Equipment. Both default off. Final renders and Yard Bake/GLB retain the requested final detail; existing graphs are not silently upgraded. Usage, compatibility, benchmarks and limitations.
The current source and dist/yard-2026.11.16-attachments-v02.zip add Attach / Rebind Equipment to Active Deck, world-preserving detach, and Regenerate Safely…. Stable deck-relative sockets follow deck resizing without stretching equipment. Regeneration stages and validates standard construction, retains every named input and the equipment hierarchy, and saves the previous node graph/input report as a backup. Replacing node construction requires explicit opt-in; custom node edits are not migrated.
Workflow, API, tests and limitations. Saved attachment poses reopen without Yard, but following further deck edits requires the updater. All cabin roof decks, the sheered main deck and raised foredeck are supported. Bake/GLB now includes socket equipment hierarchies and nested collection instances; four-vessel regeneration/export checks preserve 398 inputs per ship. The existing native 34-site equipment registry is unchanged.
- Full-height perforated cells replace the old open rings plus overlapping capped spandrel boxes. Separate reveals, frames, seals and 25 mm glazing remain.
- Planar walls stay planar. Chamfers and adjoining walls share intersections of the same offset planes, including independent rake and constant normal thickness. Matching boundaries are welded; recess returns are explicitly triangulated.
- Corner Treatment:
0opaque chamfer (default),1deliberately framed/glazed chamfer. Tiny facets remain solid. Rounded cabin corners are not implemented. - Curved hull cells are subdivided before mapping, then explicitly triangulated. Depth offsets follow the surface normal. Analytic per-corner normals on the skins do not average reveal faces into the hull. Mirrored winding is corrected.
- Regular arrays use explicit opening topology, not a Boolean per window. Existing exact Booleans remain for the cargo hold and stairwell penetrations.
The four isolated constructions and an actual extracted cruise-hull bay are shown with frames/glass hidden, under neutral grazing light, wireframe, and face-orientation shading. Blue indicates front faces; the thin orange border at a cut coupon edge is the exposed back of its inner skin, not an inverted exterior.
See diagnosis, evidence and limitations for measured original causes, checks and remaining constraints.
Coordinates: +X bow, −X stern, +Y port, +Z up; Z=0 nominal waterline. One model unit is one metre.
The superstructure no longer has cabin, level or setback numbers. Each deck is a cube empty Deck NN parented to the
ship root, with an optional window guide Deck NN Windows. The modifier reads both live through Object Info, so
moving, scaling or rotating them in the viewport is the edit:
| Transform | Effect |
|---|---|
| Deck box location X | Deck centre (+X bow) |
| Deck box scale X / Y | Deck length / width in metres |
| Deck box scale Z | Storey height, 2.3–4.5 m |
| Deck box rotation Y | Front lean: positive tips the top of the front wall forward (raked bridge), ±25° |
| Deck box rotation X | Side lean: positive leans both side walls inward, ±15° |
| Window guide scale X / Z, bottom edge | Window width / height / sill for that deck only |
A deck without a guide uses the global Window Width/Height scaled by its role. Smaller windows on the lower decks are therefore one smaller guide per deck. The graph keeps 0.10 m below and 0.15 m above every window, so a window that would reach the ceiling is moved down; the guide itself is not changed.
The Decks list in the Yard panel shows the same values as numbers: Role, Centre X, Length, Width, Storey Height,
Front/Side Lean and window width/height/sill. It also has Add, Remove and Align. Editing a storey height
restacks the decks above it. Deck Roles holds one letter per deck from the bottom (A auto, B bridge,
C cabins, P public, S service). Presets lay out deck empties that reproduce their earlier stepping in metres,
and Make Node Setup Unique gives a copied ship its own empties.
The modifier drops from 169 scalar inputs (schema 4) to 61. Gone: Levels, Cabin Position/Length/Width/Height,
the six rake angles, both setbacks and the 96 Deck NN Length/Width/Shift/Role values. New: Deck Roles. The 48
Object slots for the empties sit in a collapsed Deck Controls panel. Wall Taper, Corner Size/Treatment and Balcony
Depth remain global inputs.
Fixes in the same change, measured with tools/probe_controls.py and tools/check_stairs.py:
- Anchors no longer penetrate the hull. Anchor vertices inside the shell: cargo 290 → 0, ferry 372 → 0, cruise 384 → 0 (of 384).
- Stairs no longer pass through the slab above. Stair samples deeper than 1 cm inside a roof slab: cargo 521 → 0, cruise 1,284 → 0, ferry 663 → 2, research 392 → 0, workboat 129 → 0. The remaining hull/deck samples are stringers embedded in the deck and deliberate 5 cm sinks.
The local Rotterdam photograph was reviewed again: repeated accommodation walls are upright, balcony divisions are restrained, and upper terraces step independently. Yard is a generic reference-informed layout, not a measured replica. See references.
Existing hull dimensions, fullness, stern width, roundness, sheer, resolution and submerged rakes remain. Existing window dimensions, spacing, frame/radius/thickness, hull bands/portholes, railings, stairs, hatch, crane, seed/variation, detail budgets and human reference also remain. Full defaults and ranges: controls.
Enable Bulb Enabled, then adjust Projection, Width, Height, Z, Tip Shape and Transition independently. Z is relative to the nominal waterline. The disabled branch retains the original hull family.
The enabled hull uses 128 angular samples, at least 128 longitudinal sections, and 32 radial cap subdivisions. The bow cap is parameterized around the bulb centre instead of stretching the old keel-biased triangle fan. The bulb is part of the same closed hull mesh, not an intersecting ellipsoid. A compact vertical profile and quintic longitudinal blend control the transition.
Useful demonstrated cruise settings: projection 4 m, added width 5 m, height 4 m, Z −2.8 m, tip 0.65, transition 22 m. Bulb width is an added section-diameter control, not a guaranteed maximum beam measurement. Height ≤1.4×draft, width ≤0.5×beam, transition ≤0.4×length; centre is kept submerged and within the draft envelope. Requested input values are not overwritten by these effective clamps.
There are 24 stable sites for radar, winch, crane, two antennas, three vents, and eight lifeboat/davit pairs. Availability follows vessel type; the cruise example exposes 23 sites because its working crane is omitted.
Each site has native controls for Mode, Object, Collection, Offset, Heading, Scale and Footprint:
0useful built-in placeholder;1empty;2supplied object;3supplied collection.- An unassigned object/collection slot is empty. Placeholders are an explicit choice, not a silent replacement for missing assets.
- Append/link the source into a library collection in the scene; select it through the Yard panel. Hide source objects from direct display rather than excluding their library from the depsgraph.
- Object convention: apply rotation/scale, origin at the mounting foot, local +X forward and +Z up, metre units. Original-space Object Info deliberately ignores the object's library placement transforms. For collections, preserve child transforms and set the collection instance offset to its mounting origin.
- Supplied material assignments and local proportions are retained. The mounting Scale is uniform; no automatic bounding-box stretching occurs.
- Positions and headings reevaluate with hull/deck changes. Native
Mountsoutput andyard.assets.read_mounts()provide IDs, categories, positions, up/heading, offset, scale and advisory footprint data.
Read asset authoring and attachment contract before supplying models. The saved cruise example demonstrates the same radar.main site with an independently authored object and collection. No third-party asset is bundled or downloaded.
Beam ≤0.38×length; draft ≤0.65×beam; freeboard ≤0.55×beam. The lowest deck is ≤0.78×length, its centre stays between −0.46 and +0.43×length and its width fits the hull; each upper deck stays within the deck below. Decks are read from Deck 01 up to the first empty slot and at most min(Level Budget, 24) are built; empties above the budget are kept. Storey height is 2.3–4.5 m, front lean ±25°, side lean ±15°. Extreme wall rakes are bounded by the footprint to prevent collapsing the top perimeter. Hull bands are disabled below 1.5 m freeboard and their row count fits the available height. Tiny roofs and narrow footprints omit machinery/flights that do not fit. These are bounds, not collision or safety certification.
Detail 0 omits cabin openings and equipment; hull-band controls are separate. Detail 1 suppresses seals/heavy equipment and caps facade bays at 32. Detail 2 includes all details. Reduce Bay Budget, Level Budget, hull rows or disable the bulb while arranging the vessel.
Measured in artifacts/structural_report.json:
| Configuration | Triangles |
|---|---|
| Cargo | 74,962 |
| Ferry | 342,320 |
| Cruise, eight levels, no bulb | 795,072 |
| Cruise with demonstrated bulb | 843,732 |
Initial full graph creation: 236.1 seconds (0.3: 165.5 s; the increase has not been investigated). Warmed length edit on the detailed bulb-equipped cruise example: 188.8 ms median (0.3: 292.2 ms) over five edits, including assignment, depsgraph update and evaluated mesh access (excluding Python audits). This is a significant regression in setup cost and a heavier interactive graph than 0.2; it is not a low-poly/real-time asset. Subsequent edits reuse the graph without allocating objects or groups. Timings are local measurements, not guarantees.
The main target is a web simulation built on three.js, so the models must load as plain glTF PBR without Blender-only shader tricks. These changes are in the source tree but not yet released as a package.
Export. Use yard.export.export_glb(root, path, compression='NONE'|'DRACO'|'MESHOPT') or the CLI below. It writes <preset>[.draco].glb and a <preset>.yard.json manifest. The manifest holds the ship parameters and every mount with its ID, category, position, heading, scale and footprint in both ship space and glTF space. The exporter reads the file back and reports what GLTFLoader receives.
python tools/blender.py -b --factory-startup --python-exit-code 1 --python tools/export_web.py -- \
--blend artifacts/yard_cruise_v030.blend --tag slice1 [--presets CRUISE,FERRY,CARGO,WORKBOAT] [--compression NONE,DRACO]
glTF is +Y up. Ship (x, y, z) is exported as (x, z, −y): the bow stays at +X, up becomes +Y and port becomes −Z.
Materials. Every role is a plain Principled BSDF (base colour, metallic, roughness, optional emission), so it maps 1:1 to MeshStandardMaterial. material.userData.yard_role names the role. Glass is opaque, dark, tinted and reflective (no transmission), which avoids sorting and transmission passes in the browser. The roles are 13 in total; two are new:
- Antifouling: the red-brown underwater hull below the boot-top.
- Boot-top: the black band around the waterline.
Hull paint zones are real topology, not a shader mask. The hull cross-section is remapped so ring edges fall exactly on the boot-top limits at every station. As a result each zone is its own glTF primitive with clean straight edges. The face attribute yard_hull_zone stores the zone: 0 topside, 1 boot-top, 2 antifouling, 3 deck. Both hull ends are now closed with welded grid caps instead of triangle fans.
| New control | Default | Meaning |
|---|---|---|
| Boot-top Height | 0.45 m | Band top above the nominal waterline (capped at 0.75 × freeboard). |
| Boot-top Depth | 0.30 m | Band bottom below the waterline (capped at 0.6 × keel depth). |
| Window Corner Samples | 4 | Vertices per rounded corner for window frames, seals, glass and wall openings. 8 reproduces v0.3 exactly. |
Preset boot-top values: Cargo and Cruise 1.0/0.6 m, Ferry 0.8/0.5 m, Research 0.6/0.4 m, Workboat 0.45/0.3 m.
Lighter windows. With 4 corner samples a window cell has 444 triangles instead of 700. Wall cells keep their 32-vertex outer border, which they share with neighbouring bays. Their opening is resampled onto the coarse frame outline, so there is no gap between wall and frame (tools/probe_windows.py checks this). Hull window bands now subdivide only the skin before the curved mapping; frames, seals and glass keep their authored topology.
Equipment. Lifeboats always point bow-forward (+X) on both sides; only the davit turns to keep its posts inboard. Replacement models therefore need no per-side fix-up. The lifeboat.* mounts now report heading 0 on both sides.
Measured with tools/export_web.py (same saved cruise graph, presets applied with native set_values; artifacts/web/slice1/compare_v030.csv):
| Preset | v0.3 triangles | Now | v0.3 GLB / Draco | Now GLB / Draco |
|---|---|---|---|---|
| Cruise | 1,331,360 | 795,232 (−40%) | 61.0 / 2.86 MB | 34.4 / 1.90 MB |
| Ferry | 618,192 | 343,216 (−44%) | 27.3 / 1.39 MB | 13.8 / 0.85 MB |
| Cargo | 86,970 | 76,202 (−12%) | 3.3 / 0.26 MB | 2.7 / 0.22 MB |
| Workboat | 47,648 | 43,536 (−9%) | 1.7 / 0.16 MB | 1.5 / 0.14 MB |
Compatibility. The parameter schema is now 5 (deck empties). A ship saved with 0.3 or schema 4 still evaluates, because its graph is self-contained. Yard's compatibility check (parameters.compatible) does reject it, and the panel reports it as incompatible. Regenerate the ship to edit it with the new controls.
Validation tools:
tools/probe_hull.py: four presets with the bulb on and off, checking closed manifold, zones and band limits.--renderwrites close-ups;--save-bulbregeneratesartifacts/hull_bulb_on_v040.blend, whichtests/bulb_validation.pynow opens.tools/probe_windows.py: one window cell at every corner count, checking triangle count, the wall/frame gap and the v0.3 match at 8.tools/review_inspect.py -- --tag v040: close-up renders and per-role triangle counts.
nodes.py,bulb.py: bounded controls, hull sections, live Hull / Deck / Gunwale interface.architecture.py,walls.py,hull_windows.py: perforated cells, shared planar intersections, curved mapping/normals, reusable rails/stairs/doors.decks.py: deck-box and window-guide empties, their panel properties, Add/Remove/Align and display alignment.superstructure.py: deck reader (Object Info on the deck empties), roles, walls, balconies, circulation and roof lips.equipment.py,assets.py: retained placeholders, equipment placement, append-only mounting registry and native source selection.parameters.py,operators.py,scene.py: presets/UI, native binding, transactional ownership, deep graph uniqueness and independent baking.
The empty carrier remains an empty mesh: vessel topology, repetition, deformation and placement are native nodes. Python constructs/manages graphs, presets and assets. Custom instances are realized by Make Editable Copy, without replacing the procedural original. Make Node Setup Unique preserves/remaps internal references and keeps external assets external.
Stable face roles remain 0–17 as in 0.2. New inspection attributes include yard_surface (1 skin, 2 reveal), yard_facet, yard_design_normal and yard_mount_index. Material indices are not semantic identifiers. Mounts is a separate point-geometry output, not loose vertices mixed into the rendered vessel.
From the project root, use blender or this environment's python tools/blender.py:
blender -b --factory-startup --python-exit-code 1 --python tests/structural.py
blender -b --factory-startup --python-exit-code 1 --python tests/deck_controls.py
blender -b --factory-startup --python-exit-code 1 --python tests/assets.py
blender -b --factory-startup --python-exit-code 1 --python tests/structural_reopen.py
blender -b --factory-startup --python-exit-code 1 --python tests/native_interfaces.py
blender -b --factory-startup --python-exit-code 1 --python tools/render_bow.py
blender -b --factory-startup --python-exit-code 1 --python tools/probe_hull.py -- --save-bulb
blender -b --factory-startup --python-exit-code 1 --python tests/bulb_validation.py
blender -b --factory-startup --python-exit-code 1 --python tools/probe_windows.py
blender -b --factory-startup --python-exit-code 1 --python tools/render_structural.py -- after
blender -b --factory-startup --python-exit-code 1 --python tools/render_iteration.py
blender -b --factory-startup --python-exit-code 1 --python tools/render_hull_coupon.py
blender --factory-startup --window-geometry 50 50 1600 1000 --python-exit-code 1 --python tests/ui_test.py
blender --command extension build --source-dir yard --output-dir dist
blender --command extension validate dist/yard-2026.11.16.zip
blender -b --factory-startup --python-exit-code 1 --python tests/package_test.py
blender -b --factory-startup --python-exit-code 1 --python tools/check_stairs.py -- --tag after --no-render
python tools/write_controls_doc.py
tests/deck_controls.py -- --cached reuses artifacts/debug_workboat.blend (written by tools/debug_decks.py) instead of building a ship for about four minutes. tools/write_controls_doc.py regenerates docs/controls.md from yard/parameters.py and needs no Blender.
Older realism.py, realism_reopen.py, smoke_test.py/0.1 tests and render_release.py are historical contracts, not this release gate. Their n-gon-based glass sampling does not apply to the new explicit tessellation, and they drive the removed cabin/level inputs; tests/structural.py and tests/deck_controls.py replace them. New tests cover nine vessel configurations, aperture rays, skin planarity, miter sharing, all six bulb shape controls, object/collection/empty modes, material references, stable mounts, repeat evaluation, and reopening/editing without importing Yard. tests/deck_controls.py covers live deck moves, setbacks, front lean, window guides and the header clamp, storey restacking, roles, Level Budget, Add/Remove, presets, the web manifest and Make Unique. The reference closed bulb hull has no detected non-adjacent triangle crossings; this is not a proof for all parameters.
dist/yard-2026.11.16.zipartifacts/yard_cruise_v030.blend— native cruise graph and replaceable radar collection.artifacts/hull_bulb_on_v040.blend— schema-5 bulb example opened bytests/bulb_validation.py.artifacts/hull_bulb_on_v030.blend,hull_bulb_off_v030.blend— closed hull examples (schema 3).artifacts/diagnostics_before.blend,diagnostics_after.blend,actual_hull_coupon_v030.blend.previews/v030/— before/after diagnostics, finished and neutral close-ups, bow views and equipment views.artifacts/structural_report.json,asset_report.json,bulb_validation.json,structural_reopen_report.json,mounts_v030.json,deck_controls_report.json,ui_report.json.artifacts/probe_controls_{before,after}.json,check_stairs_{before,after}.json— anchor and stair measurements; close-ups inpreviews/probe/.
New Research ships now have a low aft working deck and a raised forward deck with paired connecting stairs. The cabin layout is unchanged. Main Deck controls set the break position, rise and stairs; Gunwale Height now defaults to 0.65 m while preserving the nominal railing-top height. See main-deck controls and checks. Existing saved scenes are not silently rebuilt.
This is visual modeling, not naval architecture or safety-design software. No hydrostatics, certified scantlings, stability or escape-route validation. The bulb remains one constrained section family. Its closed topology and controls pass the checks, but a visible stem-root crease and abrupt local curvature remain: the requested fully faired transition is not yet resolved. Unusual control combinations need visual review. Cabin corners are opaque or glazed chamfers, not continuously rounded. A deck box's location Z is display only (the graph stacks decks itself) and side lean rolls the box while the walls lean symmetrically. The hull profile is still driven by numbers, not by curves.
Window wall skins have intentional open perimeter borders beneath slabs and at diagnostic cuts. Some door/header interfaces are T-junctions; the entire passenger vessel is not one watertight welded solid. Hull bands and lower shell/decks remain separate surfaces. No all-pairs clearance solver, automatic asset fitting, arbitrary user-added mount UI, or general external rig/animation retargeting is provided. Footprints are advisory. Heavy supplied assets add their own evaluation cost. Existing simplified lifeboats, stairs, furnishings and circulation remain layout proxies.
The live Blender MCP connection timed out; all inspection and rendering used fresh installed Blender processes. The referenced chat review image was not present in the supplied transcript; the project’s previous cruise render, generated meshes and local reference photograph were inspected instead.
GPL-3.0-or-later. No reference photographs or downloaded models are included in the extension archive.




