PatternMaker: Efficient Sewing Pattern Prediction

• PatternMaker is a vision-language model that predicts a serialized sewing pattern from text, images, or both, serving as the first stage in the SwiftTailor pipeline.
• The method uses joint token prediction and regression over geometric parameters to ensure both topological validity and detailed panel synthesis.
• Evaluations demonstrate high panel and edge accuracy with faster inference using a compact 2B-parameter InternVL-3-2B backbone compared to larger models.

PatternMaker is the first-stage module of the two-stage SwiftTailor framework, where it functions as an efficient vision-LLM for predicting structured sewing patterns from diverse input modalities (Pham et al., 19 Mar 2026). Within SwiftTailor, it performs sewing-pattern reasoning rather than dense 3D reconstruction directly: given text, images, or both, it outputs a serialized sewing pattern comprising garment panels and stitching pairs, which is then consumed by the downstream GarmentSewer module for geometry synthesis (Pham et al., 19 Mar 2026). In this sense, PatternMaker occupies the interface between high-level garment intent and simulation-oriented pattern structure, a role that is central in recent computational garment pipelines (Zhou et al., 2024, Korosteleva et al., 2023).

1. Definition and position within garment-generation research

PatternMaker is introduced in "SwiftTailor: Efficient 3D Garment Generation with Geometry Image Representation" as an efficient vision-LLM that predicts sewing patterns from diverse input modalities (Pham et al., 19 Mar 2026). SwiftTailor itself is described as a two-stage framework that unifies sewing-pattern reasoning and geometry-based mesh synthesis through a compact geometry image representation, and PatternMaker is explicitly designated as Stage 1 of that pipeline (Pham et al., 19 Mar 2026).

The module is situated in a research trajectory in which sewing patterns serve as an intermediate, editable, and simulation-ready representation for garment generation. "Design2GarmentCode: Turning Design Concepts to Tangible Garments Through Program Synthesis" frames sewing patterns as the bridge between design concepts and producible garments, while GarmentCode formalizes garment construction as a hierarchical, component-oriented DSL (Zhou et al., 2024, Korosteleva et al., 2023). PatternMaker belongs to the same general problem setting—mapping high-level design conditions to structured garment specifications—but differs in being a lightweight vision-LLM that predicts a serialized sewing-pattern representation for a downstream geometry-image pipeline rather than a pattern-making program (Pham et al., 19 Mar 2026).

A common misconception is to treat PatternMaker as a complete garment generator. The published description instead defines it as a pattern-prediction module whose output is subsequently transformed into semantic and stitching maps and passed to GarmentSewer, which predicts the 3D garment geometry in UV space (Pham et al., 19 Mar 2026). This suggests that PatternMaker is best understood as a structured reasoning component rather than as an end-to-end mesh generator.

2. Architecture and supported modalities

PatternMaker is designed as a lightweight, efficient vision-LLM based on the InternVL-3-2B backbone, described as a modern multimodal LLM that is significantly smaller than comparable models in the domain such as LLaVA-1.5V-7B (Pham et al., 19 Mar 2026). The architecture includes a token prediction head for discrete elements of the sewing pattern and MLP regression heads for continuous geometric parameters such as panel vertex coordinates and transformations (Pham et al., 19 Mar 2026).

The module supports multiple input modalities. The reported modalities are text, image, and multimodal fusion, with the input encoded through InternVL-3-2B (Pham et al., 19 Mar 2026). PatternMaker is described as reasoning over both high-level semantics, such as garment style, and low-level details, such as the number and types of panels and seam layout (Pham et al., 19 Mar 2026).

Its tokenization follows the tokenization strategy from AIpparel for sewing patterns, which is presented as ensuring a structured and interpretable representation (Pham et al., 19 Mar 2026). This choice is important because it preserves a symbolic decomposition of garment structure instead of collapsing prediction into a purely implicit latent code. A plausible implication is that this structured output is one reason the module is suitable for downstream assembly and evaluation.

3. Sewing-pattern representation and prediction objective

PatternMaker predicts a serialized sewing pattern denoted by

$\mathcal{P} = (\mathbf{P}, \mathbf{S}),$

where

$$\mathbf{P} = \big\{ P_i = (V_i, E_i, R_i) \big\}_{i=1}^{N}$$

and

$$\mathbf{S} = \big\{ s_k = (e_a, e_b) \mid e_a, e_b \in \cup_{i=1}^N E_i \big\}_{k=1}^M.$$

In this formulation, $\mathbf{P}$ is the set of panels, each with vertices $V_i$, edges $E_i$, and a rigid transformation $R_i$, while $\mathbf{S}$ is the set of stitching pairs connecting panel edges (Pham et al., 19 Mar 2026).

The prediction mechanism is joint token and parameter regression. Discrete structure is handled through autoregressive token prediction, while continuous geometric attributes are handled through regression losses (Pham et al., 19 Mar 2026). The reported training objective is

$$\mathcal{L}_{\mathrm{PatternMaker}} = \mathcal{L}_{\mathrm{token}} + \mathcal{L}_{\mathrm{reg}},$$

where $\mathcal{L}_{\mathrm{token}}$ is token-level cross-entropy over the sequence and

$$\mathbf{P} = \big\{ P_i = (V_i, E_i, R_i) \big\}_{i=1}^{N}$$0

is an MSE term over continuous geometric outputs (Pham et al., 19 Mar 2026).

This formulation makes clear that PatternMaker is not merely a language-model-style serializer. It is trained to preserve both topological validity and geometric fidelity. That distinction matters in garment pattern generation because panel count, edge count, rigid transformation, and stitch correctness are all evaluated explicitly in the reported experiments (Pham et al., 19 Mar 2026).

4. Role in the SwiftTailor pipeline

PatternMaker is the first stage in SwiftTailor, and its output is not the final garment mesh but pattern data for subsequent processing (Pham et al., 19 Mar 2026). After prediction, the sewing pattern is converted into semantic and stitching maps that define panel locations, panel types, panel edges, and sewing connections; these maps form the input to GarmentSewer, the second-stage model that predicts the Garment Geometry Image (Pham et al., 19 Mar 2026).

The Garment Geometry Image encodes the 3D surface of all garment panels in a unified UV space, and the final 3D mesh is reconstructed through an efficient inverse mapping process that incorporates remeshing and dynamic stitching algorithms to directly assemble the garment, thereby amortizing the cost of physical simulation (Pham et al., 19 Mar 2026). Within this division of labor, PatternMaker provides the explicit sewing-pattern prior on which the geometry stage depends.

This separation is significant in the context of prior work. SwiftTailor’s abstract states that existing methods typically rely on large vision-LLMs to produce serialized representations of 2D sewing patterns, which are then transformed into simulation-ready 3D meshes using garment modeling frameworks such as GarmentCode (Pham et al., 19 Mar 2026). GarmentCode itself formalizes garments as hierarchical, component-oriented programs with explicit interfaces and stitching rules (Korosteleva et al., 2023). By contrast, SwiftTailor uses PatternMaker for pattern reasoning and GarmentSewer for geometry synthesis, replacing iterative physics-heavy reconstruction with a learned dense prediction stage (Pham et al., 19 Mar 2026). This suggests a modular reinterpretation of the pattern-to-geometry problem rather than a rejection of structured patterns as the intermediate representation.

5. Training data, supervision, and empirical evaluation

PatternMaker is trained and evaluated on GCD-MM, described as a multimodal extension of GarmentCodeData containing image/text/pattern triplets and editing instructions (Pham et al., 19 Mar 2026). The training regimen uses the data split from GCD-MM, joint supervision for discrete and continuous elements, and fine-tuning of InternVL-3-2B for both modalities (Pham et al., 19 Mar 2026). The full pipeline is reported as trained on 4x A100 GPUs for 3 days (Pham et al., 19 Mar 2026).

In the reported sewing-pattern generation benchmark, PatternMaker is compared against AIpparel, ChatGarment, and SewingLDM (Pham et al., 19 Mar 2026). The published figures attribute the following values to PatternMaker: Vertex L2 of 3.5 / 1.5, panel accuracy of 94.8 / 85.0, edge accuracy of 92.3 / 98.0, rotation L2 of 0.006 / 0.002, translation L2 of 1.9 / 1.3, and stitch accuracy of 85.1 / 97.8 (Pham et al., 19 Mar 2026). The same table reports AIpparel at 4.8 / 2.5 Vertex L2, 93.7 / 82.9 panel accuracy, 79.0 / 88.2 edge accuracy, 0.007 rotation L2, 2.5 / 1.7 translation L2, and 73.0 / 86.3 stitch accuracy; ChatGarment at 14.9 / 13.5, 16.4 / 19.6, 49.7 / 60.3, 0.038, 15.4 / 10.1, and 38.2 / 55.2; and SewingLDM at 15.6, 18.0, 49.0, 0.052, 16.6, and 30.6 (Pham et al., 19 Mar 2026).

The efficiency claims are equally central. PatternMaker is reported as having a model size of approximately 2B parameters, contrasted with 7B+ prior vision-LLMs, and the overall Stage 1 + Stage 2 + postprocessing pipeline is reported at approximately 15 seconds versus 50–60 seconds for others (Pham et al., 19 Mar 2026). The paper also attributes to PatternMaker lower error, higher accuracy, and a higher proportion of directly convertible and simulation-ready patterns (Pham et al., 19 Mar 2026). These results are consistent with SwiftTailor’s broader claim of state-of-the-art accuracy and visual fidelity with significantly reduced inference time (Pham et al., 19 Mar 2026).

6. Interpretation, significance, and conceptual boundaries

PatternMaker’s significance lies in combining structured garment prediction with a smaller multimodal backbone and explicit joint supervision over topology and geometry (Pham et al., 19 Mar 2026). In the broader literature, Design2GarmentCode emphasizes parametric pattern-making programs as semantically meaningful representations that bridge multimodal design concepts and vectorized sewing patterns (Zhou et al., 2024), while GarmentCode emphasizes component abstraction, algorithmic manipulation, and free-form design parametrization through a DSL (Korosteleva et al., 2023). PatternMaker does not adopt those exact representations, but it shares their commitment to explicit structure: panels, edges, rigid transformations, and stitching pairs remain first-class prediction targets (Pham et al., 19 Mar 2026).

Its boundaries are also clear from the published description. PatternMaker is not presented as a universal garment authoring environment, a CAD system, or a standalone simulator. Its defined role is to produce the sewing pattern $$\mathbf{P} = \big\{ P_i = (V_i, E_i, R_i) \big\}_{i=1}^{N}$$1 from image, text, or multimodal input, after which SwiftTailor preprocesses that pattern into semantic and stitching maps for GarmentSewer (Pham et al., 19 Mar 2026). Nor is it described as training-free: it is fine-tuned on GCD-MM with joint supervision and uses explicit cross-entropy and regression losses (Pham et al., 19 Mar 2026).

A plausible implication is that PatternMaker exemplifies a design shift in digital fashion systems: instead of scaling ever larger multimodal models for the entire garment pipeline, it isolates sewing-pattern reasoning into a compact, structured predictor and delegates dense geometric reconstruction to a separate module. Within the evidence reported in SwiftTailor, this decomposition is associated with lower error, higher structural validity, and faster inference than prior baselines (Pham et al., 19 Mar 2026).