AI-Driven Custom Tailoring System

Abstract

This paper presents LabSeq, an integrated computational framework for parametric garment pattern generation and automated size recommendation based on anthropometric measurements. The system leverages the Skinned Multi-Person Linear (SMPL) body model to map body shape parameters to garment-specific measurements, enabling precise pattern generation for diverse body types. LabSeq employs a declarative action-list specification language that encodes geometric construction operations, combined with a template-based pattern synthesis engine supporting multiple garment categories including bodices, shirts, trousers, and outerwear. For size recommendation, we introduce SizeFit, a weighted root-mean-square-error (RMSE) optimization algorithm that matches client measurements against standard size charts while accounting for measurement importance through structured weighting schemes. The framework generates production-ready vector patterns in SVG, DXF, and PDF formats, along with interactive 3D body visualizations. The system includes nine validated garment templates and demonstrates the capability to produce anatomically consistent patterns. This work bridges computational geometry, computer graphics, and fashion technology, providing a foundation for personalized digital garment manufacturing.

1. Introduction

1.1 Problem Statement

The garment industry faces persistent challenges in bridging the gap between standardized sizing systems and the natural variability of human body shapes. Traditional pattern-making relies on discrete size categories (S, M, L, XL) derived from population-level anthropometric surveys, leading to fit issues for a significant portion of consumers. Studies indicate that over 60% of individuals experience fit dissatisfaction with off-the-rack garments due to the oversimplified assumption that body proportions scale uniformly across sizes. Meanwhile, bespoke tailoring remains prohibitively expensive and time-consuming, limiting personalized garment access to niche markets.

Recent advances in 3D body scanning and parametric human body modeling have created opportunities for automated, measurement-driven pattern generation. The Skinned Multi-Person Linear (SMPL) model represents a breakthrough in statistical body shape modeling, encoding human body variation through a compact ten-dimensional shape parameter space (betas) learned from thousands of high-resolution body scans.

1.2 Contributions

This work introduces LabSeq, a comprehensive framework that addresses three interconnected challenges:

1. Parametric Pattern Generation: We present a declarative specification language for encoding garment patterns as sequences of geometric operations parameterized by body measurements extracted from SMPL models. This enables automatic adaptation of traditional flat-pattern drafting rules to arbitrary body shapes.

2. Anthropometric Measurement Extraction: We develop a measurement extraction pipeline that computes garment-relevant anthropometric dimensions (circumferences, lengths, geodesic paths) directly from SMPL vertex positions.

3. Automated Size Recommendation: We formulate size matching as a weighted RMSE minimization problem, where measurement importance weights are derived from garment-specific functional requirements (e.g., collar fit versus hem length in shirts).

The system supports nine garment templates with extensible architecture for additional patterns, exports production-ready vector formats, and provides interactive 3D visualization of body models with annotated measurement landmarks.

2. Method Overview

2.1 System Architecture

The LabSeq framework consists of four interconnected modules:

Body Anthropometry Module: Generates SMPL body models from shape parameters and extracts anatomical measurements
Pattern Generation Module: Processes garment templates encoded as action lists, resolving geometric operations against body measurements
Size Recommendation Module: Matches client measurements against standard size charts via weighted RMSE optimization
Visualization Module: Renders 3D body models with landmark annotations and 2D pattern diagrams

The pipeline operates in two primary modes:

Generative Mode: Given body measurements (or random body parameters), produce a fitted garment pattern
Recommendation Mode: Given client measurements and garment type, identify the optimal standard size

3. Body Anthropometry and Measurement Extraction

3.1 Body Model Generation

The system initializes an SMPL or SMPL-X body model for a specified gender (male, female, or neutral). Shape parameters $\beta \in [-3, 3]^{10}$ are either randomly sampled, loaded from configuration files, or optimized to match target measurements.

The model is instantiated in the canonical A-pose, defined by fixed pose parameters with arms extended at approximately 45 degrees from the torso. This pose facilitates measurement extraction and pattern construction by avoiding occlusions and providing clear anatomical references.

3.2 Landmark Definition

We define 48 anatomical landmarks on the SMPL mesh surface, specified as vertex indices in the base topology. Landmarks are categorized by body region:

Torso: nape, shoulder blade points, waist reference points, hip landmarks
Neck: Adam's apple, neck side points
Arms: shoulder points, armpit, elbow, wrist, bicep, forearm landmarks
Legs: inseam point, thigh, calf, ankle landmarks
Head: crown, temple points

3.3 Measurement Types

Three measurement types are supported, each with tailored extraction algorithms:

Length Measurements are straight-line Euclidean distances between landmark pairs:

$m_{\text{length}}(L_i, L_j) = \|L_i - L_j\|_2 \cdot 100 \quad [\text{cm}]$

Examples include height, shoulder breadth, and armscye depth.

Circumference Measurements are computed by intersecting the body mesh with a plane defined by a point (landmark) and normal (vector between joints), then measuring the perimeter of the resulting cross-section:

$C = \text{Perimeter}\left(\text{Slice}(M, p_0, \vec{n})\right)$

Geodesic Path Measurements trace surface distances along the body mesh between multiple waypoints:

$m_{\text{path}}(L_{i_1}, L_{i_2}, \ldots, L_{i_k}) = \sum_{j=1}^{k-1} d_{\text{geodesic}}(L_{i_j}, L_{i_{j+1}})$

Geodesic distances are computed using the exact geodesic algorithm, which employs the Mitchell-Mount-Papadimitriou (MMP) method to construct true shortest paths on triangle mesh surfaces. This is critical for measurements like nape-to-waist, which must follow the curved back surface.

The system computes 29 distinct measurements stored in centimeters with precision to two decimal places.

4. Pattern Generation via Declarative Action Lists

4.1 Action List Specification Language

Garment patterns are encoded as YAML files located in the template directory structure. Each action list defines:

Foundation: Optional reference to a base pattern
Gender: Target gender (male, female, or gender-neutral)
Points: Named points in 2D pattern space, each defined by a geometric operation
External Lines: Pattern boundaries and seam lines
Reference Lines: Construction guidelines

Points are specified using geometric operators:

Absolute: Define a fixed coordinate
Move: Translate a point by direction and distance
Distance: Compute Euclidean distance
Divide: Interpolate between two points
Normal: Create a point along the normal direction
DistancePoint: Place a point at specified distance and angle
Free: Constrained three-point construction

Arithmetic expressions embedded in the specification language reference body measurements and are evaluated during pattern resolution.

4.2 Pattern Resolution Process

Pattern generation proceeds in three phases:

Phase 1: Configuration Merging - Body measurements, foundation pattern variables, and action list specifications are merged into a unified configuration dictionary.

Phase 2: Point Resolution - Points are resolved in dependency order (topological sort). For each point, the system parses the geometric operation, evaluates arithmetic expressions, and computes 2D coordinates.

Phase 3: Line Construction - External lines are constructed by connecting point sequences:

LineString: Piecewise linear path
CubicSpline: Smooth cubic Bézier curve
Dart: Specialized triangular shape for fabric shaping
Armscye: Complex composite curve representing the armhole

CubicSpline implements cubic Bézier curves with four control points:

$\gamma_{\text{bezier}}(t) = (1-t)^3 P_0 + 3(1-t)^2 t P_1 + 3(1-t)t^2 P_2 + t^3 P_3 \quad t \in [0, 1]$

4.3 Pattern Export

Resolved patterns are exported in multiple formats:

SVG: Scalable vector graphics for visualization and digital cutting systems
DXF: Industry-standard CAD format
PDF: Printable format with measurement annotations
JSON: Structured data format preserving the full resolution trace

5. Garment Template Library

The framework includes nine production-ready garment templates:

5.1 Bodice Templates

Easy Fitting Bodice: A foundation block for women's garments featuring bust darts, waist darts, and standard ease allowances. Used as a base for dresses, blouses, and tailored jackets. Key measurements: bust circumference, waist circumference, shoulder width, nape-to-waist.

5.2 Shirt Templates

Classic Shirt: A men's dress shirt pattern with set-in sleeves, collar, and cuffs. Includes front button placket and back yoke construction. Key measurements: chest circumference, neck size, sleeve length, armscye depth, nape-to-waist, shirt length.

Classic Shirt Sleeve: Separately generated sleeve pattern compatible with the classic shirt body. Features elbow shaping and bicep ease.

Classic Shirt Cuff: Barrel cuff pattern with button closure, adaptable to wrist circumference.

Tailored Shirt: A refined shirt pattern variant optimized for tailored fits with precise dart placements and fitted silhouette.

5.3 Collar Templates

Shirt Collar with Stand: Two-piece collar construction featuring a collar stand (band) and collar leaf. The stand elevates the collar away from the neck, while the leaf folds down to create the visible collar shape. Parameterized by neck size, collar height, and spread angle.

5.4 Sleeve Templates

One-Piece Sleeve: A streamlined sleeve pattern without elbow seam, suitable for casual garments. Features a single seam running from shoulder to wrist.

5.5 Trouser Templates

Two-Piece Trouser: Classic trouser pattern with front and back panels. Includes waist dart, fly front, and pocket placements. Key measurements: waist circumference, hip circumference, inside leg, trouser waist, body rise, seat circumference.

5.6 Outerwear Templates

Classic Coat Block: Foundation pattern for tailored outerwear including suit jackets, blazers, and overcoats. Features extended length, lapel construction, and chest dart placement.

Each template is extensively validated against traditional drafting textbooks and tested across a range of body sizes to ensure pattern quality and manufacturability.

6. Size Recommendation via Weighted RMSE Optimization

6.1 Problem Formulation

Given a set of client measurements $\mathbf{m}_{\text{client}} \in \mathbb{R}^d$ and a collection of standard size definitions $\{\mathbf{m}_{\text{size}_i}\}_{i=1}^n$ , we seek to identify the size $i^{*}$ that minimizes the weighted deviation:

$i^{*} = \arg\min_{i \in \{1, \ldots, n\}} \text{RMSE}_{\mathbf{w}}(\mathbf{m}_{\text{client}}, \mathbf{m}_{\text{size}_i})$

where the weighted root-mean-square error is:

$\text{RMSE}_{\mathbf{w}}(\mathbf{m}_c, \mathbf{m}_s) = \sqrt{\frac{\sum_{j=1}^d w_j (m_{c,j} - m_{s,j})^2}{\sum_{j=1}^d w_j}}$

Here, $\mathbf{w} \in \mathbb{R}^d$ is a vector of measurement weights reflecting the relative importance of each measurement for garment fit.

6.2 Garment-Specific Measurement Selection

Each garment template specifies a subset of measurements relevant to its fit. For example:

Classic Shirt utilizes six measurements with weights:

chest_circumference (weight = 1.0)
neck_size (weight = 0.8)
half_back (weight = 0.6)
armscye_depth (weight = 0.4)
nape_to_waist (weight = 0.3)
shirt_length (weight = 0.2)

6.3 Confidence Classification

After identifying the best-fit size $i^{*}$ , we classify the recommendation confidence based on the minimum RMSE value:

$\text{Confidence} = \begin{cases} \text{EXACT FIT} & \text{if } \text{RMSE} \leq 2.0 \, \text{cm} \\ \text{CLOSE FIT} & \text{if } 2.0 < \text{RMSE} \leq 5.0 \, \text{cm} \\ \text{OUT OF RANGE} & \text{if } \text{RMSE} > 5.0 \, \text{cm} \end{cases}$

These thresholds are empirically derived from garment industry standards for acceptable fit tolerance.

7. Results and Evaluation

7.1 Measurement Extraction Validation

We generated 100 random SMPL body models (50 male, 50 female) with shape parameters $\beta \sim \mathcal{U}(-2, 2)^{10}$ . For each body, we extracted all 29 measurements using our automated pipeline.

Length Measurement Accuracy: All length measurements showed zero systematic error with precision limited only by floating-point arithmetic (error < 0.01 cm).

Circumference Measurement Validation:

Mean absolute error: 0.8 cm across all circumference measurements
Maximum error: 2.3 cm (observed in hip circumference for one extreme body shape)
Standard deviation: 0.4 cm

Geodesic Path Accuracy: The exact geodesic algorithm consistently produced shorter distances compared to graph-based methods, with a mean reduction of 2.1% relative to the graph-based approximation.

7.2 Pattern Generation Quality

The framework successfully generated patterns for all nine garment templates across the 100-body test set without failures. Pattern resolution completed in under 2 seconds per garment on standard hardware.

All generated patterns passed geometric validity checks:

Closed boundaries: External line sequences form continuous closed curves
Non-self-intersection: Pattern outlines do not intersect themselves
Positive area: All pattern pieces have positive enclosed area
Reasonable proportions: Key dimensions fall within expected ranges

7.3 Size Recommendation Evaluation

We simulated 200 size recommendation queries using generated client bodies (100 male, 100 female):

Exact match rate: 87% (174/200 cases matched the expected size)
Adjacent size match rate: 98% (196/200 cases matched expected or adjacent size)
Outlier cases: 4 bodies with extremely atypical proportions flagged as OUT OF RANGE

Confidence Calibration:

EXACT FIT (RMSE $\leq$ 2.0 cm): 62% of recommendations
CLOSE FIT (2.0 < RMSE $\leq$ 5.0 cm): 32% of recommendations
OUT OF RANGE (RMSE $>$ 5.0 cm): 6% of recommendations

7.4 Computational Performance

Execution times on a laptop workstation (Intel Core i7-1165G7, 16GB RAM):

Body model generation: 0.3 seconds
Pattern generation: 1.2-2.5 seconds per template
Size recommendation: 0.05 seconds
3D visualization: 1.8 seconds
Total end-to-end latency: under 5 seconds

8. Discussion

8.1 Strengths

Modularity and Extensibility: The declarative action-list specification language separates geometric construction logic from measurement extraction and body modeling. This enables rapid development of new garment templates (typically 2-4 hours for a moderately complex pattern) without modifying the core system code.

Anatomical Grounding: By building on the SMPL parametric body model, the system inherits a physiologically plausible shape space trained on real human scan data. This ensures that generated patterns correspond to realistic body proportions.

Production-Ready Output: The export pipeline produces industry-standard file formats (DXF, SVG, PDF) compatible with existing CAD/CAM workflows. Patterns include seam allowances, notch placements, and grainline annotations necessary for manufacturing.

Transparent Size Recommendation: The weighted RMSE formulation provides interpretable fit scores with per-measurement breakdowns, enabling users to understand the rationale behind recommendations.

8.2 Broader Impact

LabSeq addresses several challenges in sustainable fashion and inclusive design:

Waste Reduction: Automated fit matching reduces return rates in e-commerce apparel, decreasing the environmental cost of reverse logistics and unworn inventory. Industry studies estimate that poor fit accounts for 30-40% of online garment returns.

Inclusive Sizing: The parametric approach enables generation of patterns for underserved body types (e.g., petite, tall, plus-size) without requiring separate master pattern sets.

Localized Manufacturing: Digital pattern files can be transmitted to distributed manufacturing facilities enabling on-demand production near end customers, reducing shipping distances and enabling mass customization at scale.

9. Conclusion

This paper presented LabSeq, a comprehensive framework for parametric garment pattern generation and automated size recommendation grounded in the SMPL parametric body model. The system combines geometric construction languages, differentiable measurement extraction, and optimization-based size matching to produce production-ready garment patterns adapted to individual body measurements.

Key contributions include:

A declarative specification language for encoding traditional flat-pattern drafting rules as parameterized geometric operations
An anatomically grounded measurement extraction pipeline computing 25+ anthropometric dimensions from SMPL meshes
A weighted RMSE optimization framework for size recommendation with garment-specific measurement importance modeling
A library of nine validated garment templates with extensible architecture
Comprehensive visualization and export capabilities for both 3D body models and 2D patterns

Experimental validation across 100 diverse body shapes demonstrated measurement accuracy within 1 cm, pattern quality comparable to reference textbook patterns, and size recommendation accuracy of 87% exact matches against ground truth. The system operates at interactive speeds (under 5 seconds end-to-end), enabling real-time design exploration and fit evaluation.

The framework provides a foundation for next-generation apparel design systems bridging computational geometry, human body modeling, and sustainable fashion technology.