Research libraryFrom Mood Board to Specification
Chapter 04 of 14 · Part II · Draft web edition
A structured representation makes errors inspectable, but validity of a representation does not guarantee a good garment.
The cloth is woven from this chapter's 5 exhibits: each weft is one exhibit's own data, normalised; each warp group is a section. The probe reads the exhibit's own value where it crosses. Procedural visualisation, not a measurement.
Read-out
Illustrative GFT placement of published representation families on two qualitative engineering axes: directness of named parameter editing by a technician (horizontal) and explicitness of built-in structural schema constraints before CAD export (vertical). Mechanism definitions are sourced from S02, S08, S09, S10, S14 and S16; coordinate positions are illustrative GFT analytical judgements for workflow design, not empirical benchmark measurements.
Unit: illustrative control score (0 to 10). Status: illustrative. Sources: S02, S08, S09, S10, S14, S16. Illustrative figures are teaching examples, not measured factory results.
Overture · 1 min
A greenfield apparel manufacturer should accept generative pattern assistance only when the system emits an explicit, editable two-dimensional geometry and seam-relationship record rather than a flat picture or an unindexed three-dimensional mesh.[S02][S09] In Cameroon, where approximately 95 percent of raw cotton fibre is exported unprocessed and roughly 5 percent undergoes local transformation against a 50 percent target by 2030, capturing value in domestic cut-make-trim plants depends on converting creative briefs into producible cutting instructions without wasting cloth or sample-room time.[R06][R07] An approved concept from Chapter 3 defines visual intent and permitted feature changes for an illustrative woven cotton shirt, yet no spreading table or cutter acts on pixels alone.[S09][S22] Between the selected concept and the pattern technician's computer-aided design (CAD) workstation sits a representation choice: how panels, boundary curves, seam pairings, grain directions and dimensional parameters are encoded, validated and edited before cutting.[S02][S14] Published research from 2021 to 2026 represents sewing patterns as domain-specific programs, JSON parameter files, discrete token sequences, explicit specification languages or implicit signed-distance fields.[S02][S08][S10][S14][S0401][S0403] Because each representation exposes different control points and fails differently when parameters conflict, GFT recommends placing a deterministic validation and technician-editing boundary between model inference and CAD release.[S16][S20][S22]
4.1
Exhibit 136 paragraphs step it
Read-out
| Panel identity | panel_id / role | SHIRT-01-FRONT-L (Left front bodice) | Unique ID per piece; no orphan panels |
|---|---|---|---|
| Cutting rule | cut_qty / mirror_pair | 1 self, mirror with SHIRT-01-FRONT-R | Left-right symmetry or declared placket asymmetry |
| Coordinate basis | unit / closed_loop | millimetres; tolerance 0.1 mm | Unit readback in CAD; zero open contours |
| Fabric alignment | grainline_vector | 0 deg to warp selvedge | Prevent off-grain twist during spreading |
| Boundary indexing | edge_id / curve_type | E01..E07 (line, cubic Bezier) | Every stitchable edge addressable by ID |
| Seam mapping | seam_pair_id / allowance | SP-SHOULDER-L; 10 mm net-to-cut | Pass to Chapter 5 topology and Chapter 9 DXF |
| Material link | material_id / interlining | MAT-COT-POPLIN-01; collar fusible | Pass to Chapter 6 physical swatch record |
Illustrative GFT teaching schema for the running cotton-shirt example. Geometric and topology concepts draw on S02 and S14; CAD exchange fields draw on S22. Field codes are a GFT design proposal, not a standard adopted by those authors.
Unit: proposed schema fields. Status: illustrative. Sources: S02, S14, S22. Illustrative figures are teaching examples, not measured factory results.
A garment image records projected light, shading and folds from one viewpoint, whereas a sewing pattern defines the intrinsic two-dimensional rest shape of flat fabric pieces and the topological rules joining their boundaries into a three-dimensional shell.[S02][S03][S0402] When an image generator renders a cotton shirt on a figure, it resolves occluded rear panels, yokes, collar stands and underarm seams by visual plausibility rather than planar area conservation.[S03][S09] Receiving only that rendered image leaves the technician to draft every panel boundary by hand, place darts or seams, assign seam allowances and verify that opposing stitching lines match within tolerances.[S02][S22] Treating image synthesis and pattern generation as interchangeable confuses visual ideation with geometric engineering.[S09][S17]
Computational pattern research models garments as structured two-dimensional panels rather than pixel arrays or deformed meshes.[S02][S0401][S0402] Early synthetic dataset generation introduced parametric templates across 19 base garment categories to produce over 20,000 design variations paired with three-dimensional draped models, establishing flat panel geometry as an essential prior.[S0401] NeuralTailor then framed pattern recovery from three-dimensional point clouds as set regression with variable cardinality, using point-level attention to disentangle flat panel shapes and stitch links from deformations caused by gravity, body contact and fabric properties.[S0402] In 2023, Sewformer extended structured recovery to single two-dimensional photos via a two-level Transformer trained on the synthetic SewFactory dataset of roughly 1 million image-and-pattern pairs, predicting parametric panel shapes alongside stitch relationships.[S03]
While reconstruction models infer patterns from visual observations, programmatic systems construct patterns forward from explicit rules.[S02] GarmentCode introduced an object-oriented domain-specific language (DSL) organising sewing patterns into hierarchical, interchangeable components.[S02] Instead of dragging vertices on a flat canvas, a GarmentCode program defines modular sub-assemblies such as a bodice, sleeve, collar or skirt, and computes two-dimensional contours algorithmically from design parameters and body measurements.[S02] That separation matters in product development because changing a high-level dimension, such as neck circumference or armhole depth, must propagate consistently across connected boundaries without breaking loop closure or creating orphan stitch links.[S02][S30]
Academic representations encode panel outlines as ordered straight segments, Bezier curves or signed distance fields in a two-dimensional plane, paired with stitch correspondences for simulation.[S02][S03][S0403] A simulation-ready research file remains narrower than a factory pattern record.[S14][S22] In GFT's Cameroon textile programme base case, centred on cut, make, trim and finishing operations with optional partnered upstream supply, every pattern piece entering the sample room must state cut quantity, warp and weft alignment, fabric or interlining assignment, and notch and seam-allowance placement relative to the net stitching line.[R06][S22]
To bridge research geometry and sample-room execution, GFT proposes an illustrative panel identity schema for the running cotton-shirt example ({{xr:wp10-ch04-pattern-record-schema}}).[S02][S14][S22] Each piece carries a stable panel identifier, component role, cut quantity, mirror rule, explicit warp grainline vector, material and interlining link, and indexed boundary edges separating net sewing lines from outer cutting lines.[S02][S22] These cutting-room fields wrap the geometric output so technicians inspect a complete specification candidate rather than bare coordinate arrays.[S19][S22]
The schema in {{xr:wp10-ch04-pattern-record-schema}} clarifies what a geometric representation can and cannot establish.[S02][S14][S22] It organises the metadata required before CAD export, ensuring coordinates travel with unit declarations, grain orientation and edge identifiers.[S22] What the table cannot establish is whether the dimensions fit a target wearer group or whether the cotton poplin drapes without puckering along the armhole; those questions require seam-assembly verification in Chapter 5, physical material testing in Chapter 6 and size-aware fit trials in Chapter 7.[S12][S27]
4.4
Technical teams evaluating structured pattern generation must choose among intermediate representations published between 2023 and 2026.[S02][S08][S09][S10][S14][S16] Multimodal models do not natively output cutting files; they emit an intermediate representation that a decoder or compiler translates into flat two-dimensional panels.[S09][S10][S14] That intermediate structure determines whether a technician can inspect, edit and repair the output when a model misinterprets a brief.[S09][S15]
GFT note
For a greenfield maker establishing its first digital pattern workflow, GFT recommends prioritising representations emitting inspectable parameter tables or schema-checked component programs ([^S02], [^S09], [^S10], [^S16]) over opaque coordinate regressions. When a collar stand or placket fails a dimensional check, the technician must edit an explicit parameter value and recompile deterministically rather than re-prompting a neural network.[S15][S20]
As {{xr:wp10-ch04-representation-matrix}} illustrates, representations exposing named parameters or hierarchical schema branches ([^S02], [^S09], [^S10], [^S16]) give technicians stronger control than flat token sequences that must be decoded before their geometry is visible ([^S08]).[S14][S16] Still, {{xr:wp10-ch04-representation-matrix}} classifies edit surfaces rather than model accuracy; it cannot prove that a DSL or JSON pipeline will succeed on a Cameroonian buyer's specification until tested locally under Chapter 13.[S19][S20]
4.7
Loop closure
Loop closure: Open panel contour 100, Self-intersecting dart 0, Wrong unit scale (mm vs cm) 0, Duplicate or redundant panel 0 (illustrative severity index (0 to 100)) · illustrative
Valid syntax in an intermediate representation does not guarantee a geometrically valid or sewable pattern.
Simple polygon
Simple polygon: Open panel contour 0, Self-intersecting dart 100, Wrong unit scale (mm vs cm) 0, Duplicate or redundant panel 0 (illustrative severity index (0 to 100)) · illustrative
Unit bounds
Unit bounds: Open panel contour 0, Self-intersecting dart 0, Wrong unit scale (mm vs cm) 100, Duplicate or redundant panel 0 (illustrative severity index (0 to 100)) · illustrative
Part count
Part count: Open panel contour 0, Self-intersecting dart 0, Wrong unit scale (mm vs cm) 0, Duplicate or redundant panel 100 (illustrative severity index (0 to 100)) · illustrative
Edge ratio
Edge ratio: Open panel contour 50, Self-intersecting dart 50, Wrong unit scale (mm vs cm) 0, Duplicate or redundant panel 50 (illustrative severity index (0 to 100)) · illustrative
The method behind it: Deterministic pre-CAD geometry check protocol.
Illustrative comparison of four structural failure modes across five pre-CAD check dimensions, grounded in the quality-control exclusions of S30 (self-intersecting panels, empty/trivial components, out-of-bound lengths) and CAD unit/contour requirements of S22. Severity values (0, 50, 100) are illustrative rule-trigger indicators showing which check catches each defect, not observed defect frequencies.
Unit: illustrative severity index (0 to 100). Status: illustrative. Sources: S22, S30. Illustrative figures are teaching examples, not measured factory results.
4.7 · the bodyExhibit 15 above; its figures are illustrative where marked.
Valid syntax in an intermediate representation does not guarantee a geometrically valid or sewable pattern.[S14][S16][S30] A JSON file can parse cleanly while specifying a cuff wider than its sleeve opening, and curve tokens can decode into a self-crossing polygon.[S16][S30] GarmentWeaver's authors note that flat long sequences entangle structure with detailed parameters, causing redundant components, inaccurate local details and poor simulation compatibility.[S16] Similarly, GarmentCodeData's creators report that incompatible parameter combinations remain a common pitfall of complex parametric systems, including their own.[S30]
Recent research addresses these failures during training and inference via hierarchical schemas and regularisation ([^S16]), and during decoding via deterministic validity handling ([^S14]). GarmentWeaver activates only structural branches relevant to the garment category and applies feasibility-aware regularisation to penalise invalid predictions.[S16] PatternGSL encodes panel boundaries, parameterised seams and stitch topology in a standardised language and applies lightweight deterministic validity handling during decoding, preserving simulation readiness without expensive optimisation cleanup.[S14]
Even when a decoder enforces closure, industrial pattern engineering requires auditable invariants before a file enters CAD.[S20][S22] Allowing an invalid panel into a 3D simulator or 2D marker queue wastes compute and technician hours on defects that computational geometry rules catch in milliseconds.[S20][S30]
Drawing on GarmentCodeData's exclusions ([^S30]) and DXF exchange parameters ([^S22]), GFT proposes a mandatory deterministic validation gate between pattern generation and technician review ({{xr:wp10-ch04-geometry-failures}}). Before any generated cotton-shirt pattern is opened for refinement, an automated validator should check six rules:
mm, cm or inch), and bounding boxes must lie within category bounds so an illustrative 74-centimetre shirt front is never imported as 74 millimetres or 74 inches.[S22][S30][^S30]) and collar-stand neckline length matching the neck opening within an illustrative plus or minus 2.0 millimetres, unless a technician flags an intentional gathering, pleating or ease exception.[S12][S27][S30]The profile in {{xr:wp10-ch04-geometry-failures}} shows why one validity flag cannot diagnose failures.[S20][S22][S30] An open contour fails loop closure while passing unit-scale and part-count checks; centimetre-versus-millimetre confusion produces closed polygons ten times too small in CAD; and a duplicate sleeve passes single-panel tests while corrupting the bill of materials.[S16][S22][S30] What {{xr:wp10-ch04-geometry-failures}} does not provide is an empirical defect rate; failure shares must be measured in a local pilot.[S19][S20]
Interlude · the cloth half-woven
Woven: sections 4.0 to 4.7Still to weave: 4.10 to 4.19
Read-out
Sourced from AutoSew full-text sections 8.1 and 8.2. Total and multi-edge metrics concern stitch correspondences on the authors' test set. Garment sewn percentage is fully resolved test patterns divided by all test patterns. The abstract rounds total F1 to 96 percent. These are digital benchmark results, not physical sewing yield or GFT performance.
Unit: percent, with different denominators. Status: sourced. Sources: S12.
Read this exhibit in its chapter ↗4.10
A recurring question in the 2023 to 2026 literature is whether generative garment systems should rely on parametric templates or operate in a template-free regime.[S02][S14] In computer vision, template dependence limits open-world reconstruction: when recovering arbitrary clothing from street photographs, a fixed library of programmatic blocks cannot represent novel cut-outs, drapery or seams absent from the library.[S03][S14] PatternGSL targets this gap, preserving pattern-based rigour while removing template dependence across 300,000 training examples.[S14]
preserving pattern-based rigour while removing template dependence across 300,000 training examples
S14
What solves open-world reconstruction in computer vision does not automatically solve quality control in a greenfield garment factory.[S19][S20] In industrial production, a base block (or sloper) is an accumulated record of verified fit, tested grading increments, sewing-line balance and buyer-approved tolerances.[S02][S15][S22] When a plant manufactures woven cotton shirts, uniforms or school garments, starting every style from an unconstrained, template-free coordinate prediction introduces unnecessary variance in armhole scye depth, shoulder slope and neck-drop balance.[S02][S27]
Even within the literature, systems oriented toward controllable authoring and multi-layer refinement build on structured component libraries.[S02][S09][S10][S18] GarmentCode's component abstraction lets users swap, combine and parameterise modular elements while maintaining validity.[S02] ChatGarment refines GarmentCode to translate vision-language JSON parameters into valid patterns ([^S10]), and DressAgent extends GarmentCode to recover multi-layer outfits with closed-loop simulation refinement ([^S18]). TailorCoPilot's version-control backend, TailorTrace, captures expert practice by recording discrete state transformations and operation sequences over panels, edges, vertices and stitches inside a conventional interface.[S15]
For GFT's staged Cameroon programme, distinguishing template modification from template-free synthesis yields a clear rule.[R06][S02][S14] During plant launch and qualification, the factory should anchor its workflow in an owned, technician-verified library of base blocks for its core product families, including the illustrative woven cotton shirt.[S02][S15] Generative tools should first map a constrained brief from Chapter 3 onto controlled parametric variations of those blocks, such as altering collar point length, switching between single-piece and two-piece yokes, adjusting sleeve length or resizing a chest pocket.[S09][S10]
Template-free generation ([^S14]) and schema-aware synthesis ([^S16]) remain valuable to test in bounded pilots where a designer requests a silhouette outside existing blocks, but they should never silently overwrite an approved block.[S14][S16][S20] Whenever a technician accepts a departure beyond an approved block's parameters, the pattern record should log an exception reason, model version and physical fit-sample gate before cutting is authorised.[S15][S19][S20]
4.13
Exhibit 1610 paragraphs step it
Read-out
Derived by GFT arithmetic from Section 3.4 of S30: each generation batch comprises 5,000 design samples that have already passed 2D pattern quality checks; after XPBD draping and collision/equilibrium quality control, approximately 3,200 final designs pass per batch. Derived exclusion = 5,000 - 3,200 = 1,800 designs (36 percent of the 5,000 batch). Pattern-stage rejections prior to the 5,000 batch are discarded continuously until 5,000 valid patterns are collected and are not quantified as a fixed count in S30.
Unit: designs per 5,000-sample batch. Status: derived. Sources: S30.
Learning-based pattern generators and stitch predictors depend on synthetic datasets because paired industrial archives of garment photographs, graded DXF patterns, measured fabric physics and multi-size fit approvals are rarely public.[S03][S08][S14][S30][S0401] Understanding how synthetic corpora are constructed and filtered is essential when judging whether a published model will generalise to local garments and wearers.[S19][S20][S30]
Public datasets illustrate both the scale and built-in assumptions of synthetic garment data: the 2021 NeurIPS dataset generated over 20,000 variations from 19 base types with 7 held-out topologies ([^S0401]); SewFactory synthesised roughly 1 million image-and-pattern pairs ([^S03]); GarmentCodeData constructed 115,000 samples across three textile presets fitted to a neutral body and 5,000 sampled body shapes ([^S30]); AIpparel curated over 120,000 multimodal garments ([^S08]); PatternGSLData generated 300,000 image-to-GSL pairs ([^S14]); AutoSew annotated over 18,000 multi-edge patterns ([^S12]); and DressAgent built over 15,000 multi-layer outfits ([^S18]).
Because the full text of GarmentCodeData ([^S30]) discloses its generation and quality-control mechanics, it reveals what synthetic training data contains. First, naive uniform parameter sampling produced non-representative garments, such as skirts or sleeves longer than body dimensions.[S30] To prevent this, the authors aligned parameter bounds with body measurements and introduced manual coin-flip probabilities $p_i$ to favour daily-wear conventions, dividing the corpus into roughly one third upper-body, one third lower-body and one third full-body garments, and setting $p_i = 0.8$ on neckline and sleeve asymmetry to favour left-right symmetry.[S30]
Second, the body shapes were generated by registering a template mesh to the European subset of the CAESAR database (1,700 three-dimensional scans), applying principal component analysis (PCA) to sample 5,000 body shapes, and computing 25 body measurements automatically along balance planes.[S30] Compared against manual tailor measurements in CAESAR, automatic convex-hull algorithms overestimated bust circumference by 2.81 centimetres, hip circumference by 1.99 centimetres and leg circumference by 3.49 centimetres, while underestimating shoulder width by 4.57 centimetres by measuring across shoulder ball joints rather than deltoid muscles.[S30]
Third, the pipeline relied on two quality-control filtering stages ({{xr:wp10-ch04-synthetic-filtering-waterfall}}).[S30] At the two-dimensional pattern stage, the authors excluded samples with self-intersecting panels, full-body lengths exceeding floor length, empty patterns, belt-only outputs, and lower garments prone to sliding off.[S30] Passing patterns were generated in batches of 5,000 designs (10,000 patterns per batch across random and neutral bodies) and draped in an Extended Position-Based Dynamics (XPBD) simulator with temporary half-space constraints, capped at 5 minutes and 2,400 frames per drape.[S30] Drapes failing static equilibrium or showing high cloth-body or self-collision intersections were discarded: out of each batch of 5,000 simulated drapes, approximately 3,200 final designs passed simulation quality control, a batch retention of 64 percent (and an overall reported simulation success rate of about 72 percent).[S30]
The reconciliation in {{xr:wp10-ch04-synthetic-filtering-waterfall}} carries an operational warning.[S30] Training only on the surviving 3,200 designs per batch that passed two-dimensional geometric filtering and three-dimensional simulation filtering yields a curated survivor distribution rather than unconstrained parametric sampling.[S20][S30] Moreover, {{xr:wp10-ch04-synthetic-filtering-waterfall}} reflects only simulation-stage attrition inside the authors' pipeline; it cannot predict how often a generative model emits invalid geometry when prompted with local Cameroonian styles, non-European body proportions or fabrics outside the three presets in GarmentCodeData.[S19][S30]
Because synthetic corpora encode specific body-shape priors (the European CAESAR subset in [^S30]), strong symmetry preferences ($p_i = 0.8$ in [^S30]) and three textile presets, a greenfield maker must test distribution shift explicitly.[S19][S20][S30] GFT recommends splitting any local test suite into two buckets:
Reporting accuracy separately on these buckets prevents strong performance on symmetric shirts from masking failures on asymmetric or locally proportioned garments.[S19][S20]
4.16
Brief and block input
Brief and block input: Approved style brief (Ch 3), Frozen base-block ID, Target measurement table, Permitted variant flags · illustrative
When a system produces sewing patterns via program synthesis ( [^S09] , [^S18] ), JSON parameter files ( [^S10] ) or specification languages ( [^S14] )…
Multimodal synthesis
Multimodal synthesis: Program synthesis (S09), Parameter JSON (S10), Schema branches (S16), PatternGSL spec (S14) · illustrative
Deterministic validator
Deterministic validator: Whitelisted DSL/JSON parser, Closed-loop and area checks, Self-intersection sweep, Frozen-parameter diff gate · illustrative
Technician editor
Technician editor: Inspectable parameter table, 2D panel contour view, Operation trace (S15), Signed exception codes · illustrative
CAD export and handoff
CAD export and handoff: AAMA/ASTM DXF export (S22), Explicit unit and notch rules, Seam map to Chapter 5, CAD readback to Chapter 9 · illustrative
The method behind it: Unverified code repositories and security boundaries.
Illustrative GFT five-layer architecture separating probabilistic model generation from deterministic validation, technician editing and CAD exchange. Informed by program/JSON/schema generation in S09, S10, S14 and S16, risk controls in S20 and DXF exchange options in S22. This is a proposed engineering control architecture, not an installed GFT software deployment or a security certification.
Unit: system responsibility layers. Status: illustrative. Sources: S09, S10, S14, S16, S22. Illustrative figures are teaching examples, not measured factory results.
4.16 · the bodyExhibit 17 above; its figures are illustrative where marked.
When a system produces sewing patterns via program synthesis ([^S09], [^S18]), JSON parameter files ([^S10]) or specification languages ([^S14]), software must parse and execute that output to build two-dimensional geometry and export a CAD file.[S02][S22] The NIST Generative Artificial Intelligence Profile ([^S20]) notes that generative models can emit syntactically plausible but hazardous or unverified outputs, recommending capability checks, input/output validation and lifecycle monitoring ([^S19], [^S20]).
In a program-synthesis pipeline such as Design2GarmentCode ([^S09]), the multimodal model writes executable Python code targeting the GarmentCode library ([^S02]). Executing model-generated scripts directly in the technician's primary environment with file-system and network privileges creates two risks.[S20] First, an unconstrained interpreter could run arbitrary system calls, overwrite approved blocks or exfiltrate tech-pack records.[S20] Second, generated code could import unapproved helpers, alter unit settings, enter an infinite loop during curve subdivision or silently modify parameters on components marked as frozen.[S02][S20]
To contain both failure modes, GFT proposes a deterministic architectural boundary between the generative model and the factory's CAD environment ({{xr:wp10-ch04-deterministic-boundary}}).[S09][S10][S14][S16][S20][S22] The model never interacts directly with the pattern database or CAD application; outputs enter an isolated, network-disabled geometry worker restricted to a whitelist of declarative primitives and bounded schemas.[S10][S14][S20]
The five-layer separation in {{xr:wp10-ch04-deterministic-boundary}} ensures probabilistic inference stops at the second layer.[S09][S10][S14][S16] The third layer acts as an airlock: it parses only schema-compliant JSON dictionaries ([^S10]) or syntax trees containing whitelisted GarmentCode or PatternGSL constructors ([^S02], [^S09], [^S14]), enforces timeouts, runs the geometric validators from Section 4.3, and blocks unrequested parameter changes against the frozen base block.[S20] Only patterns passing this validator reach the technician's editor and DXF export utility.[S15][S22] While {{xr:wp10-ch04-deterministic-boundary}} provides an integration blueprint, it is an illustrative GFT proposal rather than a security certification or a substitute for receiving-CAD readback tests in Chapter 9.[S20][S22]
The validator layer should also write an immutable execution log for every candidate.[S15][S20] When the illustrative cotton shirt starts from an approved base block and the Chapter 3 brief permits only collar-point and chest-pocket modifications, the geometry worker computes a parameter diff between the base block and the generated specification.[S02][S15] If the model also altered chest width by an illustrative 12 millimetres or sleeve cap height by 8 millimetres without authorisation, the diff gate rejects the candidate before technician inspection.[S20] Retaining the model identifier, weight version, brief hash, emitted code or JSON, validator flags and parameter diff log aligns the workflow with companion paper wp14.[S20]
4.19
Even when a generated pattern passes every syntax, closure, self-intersection and parameter-diff check, it remains a candidate specification rather than an authorised cutting instruction.[S19][S22] Computational validity proves only that polygons close and parameters lie within mathematical bounds; it does not prove that the collar roll sits cleanly, that armhole pitch balances movement, or that seam allowances match factory folders.[S02][S22][S27] Therefore, pattern output must arrive in two linked forms that a technician can inspect and modify: a structured intermediate parameter or operation view, and an industrial 2D CAD exchange specimen.[S09][S15][S22]
At the intermediate level, TailorCoPilot ([^S15]) shows the value of representing pattern making as discrete, version-controlled state transformations over panels, edges, vertices and stitches. When a technician adjusts a generated shirt pattern, recording the operation sequence and edit reason turns a one-off repair into a reusable training and governance record, developed in Chapter 10.[S15]
At the CAD exchange level, the pattern must cross into industrial pattern-making and marker-making software.[S22] Vendor documentation for CLO's two-dimensional pattern DXF import and export ([^S22]) illustrates the settings governing exchange survival:
Before a pattern candidate leaves Chapter 4, the technician must sign off a pattern-geometry record with seven mandatory elements: 1. Verified panel inventory: Complete panel count with panel_id, role, cut_qty, mirror rules and warp grainline_vector matching the Chapter 3 brief ({{xr:wp10-ch04-pattern-record-schema}}).[S02][S22] 2. Canonical unit and coordinate basis: Confirmed unit (millimetres), bounding-box dimensions and net enclosed area, verified by the geometry worker.[S22][S30] 3. Intermediate snapshot and diff: Archived DSL program, JSON file or PatternGSL specification, parameter diff against the approved base block and recorded technician edits.[S09][S10][S14][S15] 4. Indexed boundary and seam-pair map: Edge identifiers and proposed stitch partners handed to Chapter 5 ([^S12]) so edge correspondence and whole-garment topology are verified separately from single-panel geometry.[S12][S14] 5. Material and interlining links: Material identifiers for self fabric and fusibles handed to Chapter 6 for physical swatch calibration.[S30] 6. Base-size points of measure and ease intent: Key garment dimensions handed to Chapter 7 ([^S27]) for size-aware fit and grading evaluation.[S22][S27] 7. Exported DXF specimen and settings: Saved AAMA or ASTM DXF file with declared unit, seam-allowance, cutting-line swap and notch settings, handed to Chapter 9 for receiving-CAD readback and to wp20 for digital sampling.[S22]
Enforcing this explicit structure between creative intent and specification ensures generative pattern tools act as inspectable assistants under technician authority rather than black-box generators of unverified cutting files.[S15][S19][S20] Nothing in this chapter constitutes investment, legal or tax advice, and no pattern representation is production-ready until its seam assembly, material behaviour, physical fit and receiving-CAD readback pass their local gates.[S12][S19][S22]
Check the record
Every knot is a source this chapter cites; select one to read its record, or filter the register by grade. Publication dates and inherited-source limits stay explicit. Read the paper's full evidence summary.
23 sources · 288 citation marks on this pageMost cited: S22, 36 times
Grade AGrade BGrade CKnot size: citations on this page
All 23 records shown
R06 · Grade C · inherited record
Inherited evidence R06
About 95 percent of cotton fibre exported raw against 5 percent processed locally; industrialisation target of 50 percent local processing.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R07 · Grade C · inherited record
Inherited evidence R07
SND30 restates the 50 percent local-processing target by 2030.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
c04-S02 · Grade B
GarmentCode: Programming Parametric Sewing Patterns
Realistic representation of garments with valid sewing patterns is key to their accurate digital simulation and eventual fabrication.
2023-06-06. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S03 · Grade B
Towards Garment Sewing Pattern Reconstruction from a Single Image
Then, we propose a two-level Transformer network called Sewformer, which significantly improves the sewing pattern prediction performance.
2023-11-07. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S04 · Grade B
DressCode: Autoregressively Sewing and Generating Garments from Text Guidance
We first introduce SewingGPT, a GPT-based architecture integrating cross-attention with text-conditioned embedding to generate sewing patterns with text guidance.
2024-01-29. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S08 · Grade B
AIpparel: A Multimodal Foundation Model for Digital Garments
To simplify this process, we introduce AIpparel, a multimodal foundation model for generating and editing sewing patterns.
2024-12-05. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S09 · Grade B
Design2GarmentCode: Turning Design Concepts to Tangible Garments Through Program Synthesis
Sewing patterns, the essential blueprints for fabric cutting and tailoring, act as a crucial bridge between design concepts and producible garments.
2024-12-11. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S10 · Grade B
ChatGarment: Garment Estimation, Generation and Editing via Large Language Models
These sewing patterns can then be draped on a 3D body and animated.
2024-12-23. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S12 · Grade B
AutoSew: A Geometric Approach to Stitching Prediction with Graph Neural Networks
AutoSew achieves 96% F1-score and successfully assembles 73.3% of test garments without error, outperforming existing methods while relying solely on geometric input.
2026-02-25. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S14 · Grade B
PatternGSL: A Structured Specification Language for Template-Free and Simulation-Ready 3D Garments
PatternGSL preserves the physical rigor of pattern-based models while removing template dependence, elevating sewing structure as a first-class target for generative modeling.
2026-06-23. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S15 · Grade B
TailorCoPilot: Enabling Agentic Pattern Making with Version-Controlled State Tracking
Ultimately, TailorCoPilot demonstrates a viable pathway to capture practice-based expertise, operationalizing it to support both generative AI advancements and human apprenticeship.
2026-08-26. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S16 · Grade B
GarmentWeaver: Schema-Aware Structured Synthesis for Multimodal Sewing Patterns
In this paper, we present GarmentWeaver, a schema-aware framework for multimodal Sewing pattern generation.
2026-08-31. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S17 · Grade B
EasyFashion: A Human-AI Co-Creation System for Personalized Fashion Design and Sewing Pattern Generation
Recent generative artificial intelligence (AI) systems can visualize garment ideas but often stop short of supporting downstream production.
2026-09-16. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S18 · Grade B
Inverse Garment Design with Visual Agents
Carefully designed tools and an evolving memory make this refinement converge on the target geometry.
2026-10-11. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S19 · Grade A
Artificial Intelligence Risk Management Framework (AI RMF 1.0)
While the AI RMF can be used to prioritize risk, it does not prescribe risk tolerance.
2023-01-26. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S20 · Grade A
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
This document defines risks that are novel to or exacerbated by the use of GAI.
2024-07-26. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S22 · Grade C
2D Pattern (DXF) Import/Export
DXF is a standard file format of industrial CAD patterns and stores 2D pattern information.
2026-07-15. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S27 · Grade B
EASE: Parametric garment design with explicit and local ease control
Garment fit and comfort depend critically on ease, the local allowance of excess material relative to the body.
2026-06-28. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S30 · Grade B
GarmentCodeData: A Dataset of 3D Made-to-Measure Garments With Sewing Patterns
For each generation step, we implement quality control conditions to exclude bad samples from the data.
2024-05-27. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S0401 · Grade B
Generating Datasets of 3D Garments with Sewing Patterns
With this pipeline, we created the first large-scale synthetic dataset of 3D garment models with their sewing patterns.
2021-09-12. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S0402 · Grade B
NeuralTailor: Reconstructing Sewing Pattern Structures from 3D Point Clouds of Garments
We propose to use a garment sewing pattern, a realistic and compact garment descriptor, to facilitate the intrinsic garment shape estimation.
2022-01-31. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S0403 · Grade B
ISP: Multi-Layered Garment Draping with Implicit Sewing Patterns
Their 2D shape is defined by a Signed Distance Function and 3D shape by a 2D to 3D mapping.
2023-05-23. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c04-S0404 · Grade B
Learning Sewing Patterns via Latent Flow Matching of Implicit Fields
Despite progress in automated pattern generation, accurately modeling sewing patterns remains difficult due to the broad variability in panel geometry and seam arrangements.
2026-01-25. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗S02 Authors via arXiv; ACM Transactions on Graphics. 2023-06-06.
S09 Authors via arXiv; CVPR 2025. 2024-12-11.
R06 Cameroon Trade Hub. Publication date unrecorded; inherited source-text check.
R07 Ecofin Agency. Publication date unrecorded; inherited source-text check.
S22 CLO Virtual Fashion. 2026-07-15.
S14 Authors via arXiv. 2026-06-23.
S08 Authors via arXiv. 2024-12-05.
S10 Authors via arXiv; CVPR 2025. 2024-12-23.
S0401 Authors via arXiv; NeurIPS 2021 Datasets and Benchmarks. 2021-09-12.
S0403 Authors via arXiv. 2023-05-23.
S16 Authors via arXiv. 2026-08-31.
S20 National Institute of Standards and Technology. 2024-07-26.
S03 Authors via arXiv; ACM Transactions on Graphics. 2023-11-07.
S0402 Authors via arXiv; ACM Transactions on Graphics. 2022-01-31.
S17 Authors via arXiv. 2026-09-16.
S30 Authors via arXiv; ECCV 2024. 2024-05-27.
S19 National Institute of Standards and Technology. 2023-01-26.
S12 Authors via arXiv; WACV 2026. 2026-02-25.
S27 Authors via arXiv; Computers & Graphics. 2026-06-28.
S15 Authors via arXiv; UIST 2026 listed. 2026-08-26.
S0404 Authors via arXiv. 2026-01-25.
S18 Authors via arXiv. 2026-10-11.
S04 Authors via arXiv. 2024-01-29.
Keep the thread · chapter 04 of 14
Move the shuttle along the rack to preview any chapter: its sections, exhibits and reading time. Draft web edition. This research is not investment, legal or tax advice; illustrative models are not measured factory results.
The paper's rack · 14 boltsBolt length: reading time
01Part I: Production readiness is the test
GFT should buy creative assistance selectively and release garments only through an accountable specification process.
02Part I: Production readiness is the test
The 2023 to 2026 record shows multiple technical paths advancing together, with no basis here for an industry adoption rate.[^S02][^S18]
03Part I: Production readiness is the test
Image generation can broaden exploration when designers retain the constraints and decide which concepts deserve technical work.
04Part II: Specifications require geometry, materials and fit
A structured representation makes errors inspectable, but validity of a representation does not guarantee a good garment.
05Part II: Specifications require geometry, materials and fit
Whole-garment assembly must be checked separately from edge-level prediction metrics.
06Part II: Specifications require geometry, materials and fit
Predicted material properties can guide a test, but a simulation must remain traceable to the actual material and its measurement basis.
07Part II: Specifications require geometry, materials and fit
Fit review needs actual garment and wearer dimensions, plus explicit ease and movement requirements.
08Part II: Specifications require geometry, materials and fit
A coherent digital pattern still needs repeat, colour and provenance checks before it becomes a product specification.
09Part III: Adoption requires people, rights and measured economics
Generated specification fields must resolve to approved geometry, materials, instructions and revisions in the actual receiving workflow.
10Part III: Adoption requires people, rights and measured economics
AI assistance should expose and preserve pattern-making judgement rather than obscure who is responsible for a garment.
11Part III: Adoption requires people, rights and measured economics
Procurement must establish what may be uploaded, generated, retained and used before sensitive design work enters a model.
12Part III: Adoption requires people, rights and measured economics
Savings exist only where the cost of review and correction is lower than the work genuinely avoided.
13Part IV: GFT should scale only what passes its gates
A greenfield maker should buy proven components and qualify a bounded assisted workflow before increasing dependence on it.
14Part IV: GFT should scale only what passes its gates
Adopt creative assistance conditionally, preserve production authority and expand only when the specification workflow demonstrates value.