Skip to content

Next Gen STEP AP242 Native-ish BREP Geometric Modeling Kernel

8.6 KB
Raw Blame

Starting from GarmentCode sewing patterns

Experimental. Garment work is research, not part of the supported product. Its APIs, syntax and results may change or be withdrawn. See experimental modules.

Garment authoring can begin with an actual programmed sewing pattern rather than an approximate garment silhouette. Aetheris imports resolved GarmentCode JSON into ordinary schema Garment source: pure profile Features, arranged Panels and explicit Interface<Stitch> declarations. The existing profile resolver, analytic seam inspector, panel mesher and cloth solver retain authority. There is no Python dependency in the application runtime.

Credit: Maria Korosteleva and Olga Sorkine-Hornung, GarmentCode: Programming Parametric Sewing Patterns, SIGGRAPH Asia 2023. The upstream code is MIT licensed. See THIRD_PARTY_NOTICES.md and the complete preserved license in fixtures/Canonical/Garment/GarmentCode/LICENSE.txt.

Import, inspect, modify

dotnet Aetheris.CLI/bin/Release/net10.0/aetheris.dll garment import-garmentcode `
  fixtures/Canonical/Garment/GarmentCode/dress.json `
  --out-dir artifacts/local/garment/imported-dress --json

dotnet Aetheris.CLI/bin/Release/net10.0/aetheris.dll garment build `
  artifacts/local/garment/imported-dress/garment.firmament `
  --out-dir artifacts/local/garment/imported-dress/build --json

Import emits garment.firmament, the retained specification.json, an analytic patterns.svg preview and import-evidence.json. It does not claim to mesh, simulate or fit a new figure. Evidence records the input hash, original panel names, edge-index mapping, boundary reversals and measured seam lengths. Keep the original specification's license with imported derivatives.

The importer admits resolved relative curves, explicit units and panel transforms, one closed loop per panel, line/circular/quadratic/cubic edges and two-sided stitches. Quadratic curves are elevated exactly to cubic Beziers. Circular arcs retain radius, side and major/minor selection. Unsupported curves, unresolved parameter programs, invalid references and reused seam spans stop with diagnostics.

The source's local XY coordinates remain the flat material pattern. Upstream Y-up placement maps to Aetheris Z-up; Euler rotation order is retained. Clockwise loops are normalized to Firmament's counterclockwise outer boundary and seam correspondence is adjusted by the same reversal parity. right_wrong stitches retain their upstream same-direction pairing. Numeric edge IDs remain in the names even when semantic armhole/collar labels are present.

Upstream JSON does not carry Aetheris fabric properties or a material grain direction. The generated source exposes Aetheris defaults explicitly for review and replacement. Upstream unequal seam lengths become an explicit measured ease value; mismatches above Aetheris's 25% budget are rejected. Analytic cut lengths govern admission. Polygonal chord errors during meshing do not redefine authored ease.

Known templates

fixtures/Canonical/Garment/garmentcode-dress.firmament is the imported non-fitted Shirt block with sleeves disabled and upstream length multiplier 2.2. It has four half-torso panels and six seams, with real curved armholes and different front/back collar cuts. GarmentCode/shirt.json retains the original sleeved shirt control with eight panels and sixteen seams.

garmentcode-dress-template.firmament adds a native hemDrop: Length parameter to those same pure profile recipes:

Feature left_ftorsoCut = left_ftorsoPattern(hemDrop: 100mm);

Apply the same adjustment to all four cuts. Zero produces the same compiled cloth as the reference import; a positive value lowers the hem without changing armhole, collar or shoulder construction. Existing profile deltas remain the owner for other named-span edits. This is a resolved starter block, not a port of the entire upstream measurement-driven configurator.

Reproduce the research controls

The external checkout is pinned to d449629979028123a5c4dc9e732a2ec19b7fce31. scripts/research-garmentcode.py --upstream <checkout> --out-dir <ignored-directory> calls the real upstream Shirt program and its SVG/JSON exporter. The checked-in designs, measurement YAML, provenance and independently sampled curve oracles document the inputs. Canonical presentation ordering makes generated JSON repeat byte-for-byte despite upstream component discovery using a Python set.

Pattern-only Python dependencies used in the isolated research venv: numpy<2, pyyaml, scipy, svgwrite, svgpathtools, psutil, matplotlib, CairoSVG. The venv and checkout remain under artifacts/local/garment/research/.

For a matching external unrigged reference mesh, the shared GarmentDraping.SettleOnSurface API runs the same sewing/settling/replay path using IClothBodySurface3D, without inventing skeletons or skin weights. It rejects named figure/pose bindings and measured fits; those must use their actual humanoid binding. The bounded research harness accepts upstream Y-up, metre, triangle-only OBJ input:

dotnet run --file scripts/qualify-garmentcode-reference.cs -c Release -- `
  artifacts/local/garment/imported-dress/garment.firmament `
  <external-checkout>/assets/bodies/mean_all.obj `
  artifacts/local/garment/reference-body

The harness retains body/specification hashes, actual solver metrics, replay, contact/stretch outliers and OBJ artifacts. A failure returns code 2. These are discrete one-way CPU contact experiments without friction, CCD or self-contact; they do not qualify animation or production cloth. Body meshes and the separate GarmentCodeData dataset are not included or relicensed by this source adoption.

Current experiment results

Follow-up: garment draping research audits the retained mesh and strain distribution, compares primary research, and defines the next controlled experiments. It does not relax these qualification gates. The mesh and settling implementation records the new mesh, residual histories and solver split experiment separately from these historical controls. The dress still fails qualification.

The real upstream shirt and dress patterns match their independently generated curve oracles (five samples and length per edge, 0.00001mm tolerance), then build through the normal CLI mesher. A 100mm native hem extension keeps collar and armhole geometry unchanged; zero reproduces the reference compiled cloth.

The first drape controls use the same 60mm mesh request, cotton defaults, 30 sewing ticks and 30 gravity ticks. Blender renders only the retained solver OBJ output. These controls fail the existing 15% peak strain budget:

Control Peak stretch Maximum seam gap Maximum surface penetration Replay
Original upstream sizing/arrangement on Anatolia 42.35% 3.51mm <0.001mm Exact
Original cut on matching upstream mean_all reference body 85.62% 11.30mm <0.001mm Exact
Matching reference body, arrangement raised 100mm, cut unchanged 154.72% 4.20mm 1.49mm Exact
Anatolia, same raised arrangement and unchanged cut 50.82% 1.28mm <0.001mm Exact

The unraised dress slips below the shoulders on both bodies. Raising arrangement improves seam closure but worsens peak strain, so an arrangement offset alone is not a qualified fix. The raised control's largest outlier is a 1.48mm rest link near the front neckline, despite the 60mm mesh request. This isolates mesh quality and sewing/body-contact convergence as the next investigation; it does not prove that short links alone caused the failure. Neither changing the reference cut nor relaxing the acceptance budget is justified by these results.

Retained evidence: artifacts/local/garment/garmentcode-dress/ contains imports, sources, antonia-reference-sizing/, reference-body/, raised-reference-body/ and their metric/outlier reports. The upstream simulator was not run. No production garment fit, animation or new-body automatic measurement mapping is claimed by these pattern/interchange controls.

The raised arrangement stays on the shoulders in the reference-body render; its improved silhouette does not override the failed strain check. The same arrangement is retained as anatolia-raised.firmament and anatolia-raised/, with an unchanged flat cut and no hidden supports. Both still need solver/mesh convergence work. Both Release solution builds passed; the full Aetheris suite passed 4,479 tests, and the Aurelian consumer suite passed 1,217 with its installed Naga selected through AURELIAN_NAGA.