Towards automatic digital garment assembly

arXiv:2610.20651 · cs.CV, cs.GR · Submitted 2026-10-12 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Towards automatic digital garment assembly".

Jane: The gist: The authors present two learning-based approaches that automate both tasks—sewing and arrangement—using a 2D vector pattern as their sole input to create simulation-ready 3D garments without manual adjustments.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Let's talk about what this paper actually does in detail. It tackles the problem of automatically assembling garments from a 2D pattern, which used to require people to manually define all the sewing operations and arrange those pattern panels for the drape simulation <ref:2610.20651#pg1>.

Jane: So, essentially, they are proposing a system that uses deep learning models to figure out exactly where every seam goes and how every panel is positioned before we even start the draping part.

Lu: The core of their automatic sewing model operates at the segment level, using a Transformer encoder with multiple prediction heads to handle many-to-many, ordered segment correspondences, plus things like hems and overlapping seams that were previously neglected by prior works one <ref:2610.20651#pg1,automatic sewing model operates at the segment level>.

Meng: That's what I'm interested in—handling those complex things like facing seams and hems. Most models only focused on the basic topology of the seam.

Lalam: And they do predict regular seams, as well as facing seams, which are operations that join two panels one on top of the other to create layered garments three <ref:2610.20651#pg1>.

Tom: They also handle construction properties like seam allowances and stitch types, which were ignored by previous methods one <ref:2610.20651#pg1>. That's a lot of detail for a machine to predict.

Jane: So, when you look at the input for that segment-level model, it takes panel features like its area and perimeter, plus per-segment features including things like the degree of the Bezier segment and coordinates of up to four control points four.

Lu: They concatenate those two sets of features to get a final feature vector with twenty-one values, which implicitly groups same-panel segments by their shared panel features.

Meng: Twenty-one features sounds like a lot for one segment, but they say this structure lets the Transformer encoder process all N segments at once.

Lalam: The backbone is a standard four-layer Transformer encoder with four attention heads per layer, and they use full self-attention without masking because the basic segment features implicitly encode enough information to deduce topological connections four.

The paper's summary: Tom: Moving on to what the paper actually summarizes, they propose two architectures. First, the automatic sewing model at the segment level handles the sewing operations and construction properties.

Jane: And then they have a complementary panel-level model that predicts body-region assignments, layers, and placement coordinates specifically to initialize the drape simulation from a favorable configuration two <ref:2610.20651#pg1>.

Lu: This arrangement part is crucial because it sets up the starting point for the three dee drape simulation after the sewing is done <ref:2610.20651#pg1>. It's not just about making seams; it's about getting the initial panel placement right.

Meng: So they are tackling both sides of garment creation: how to sew it correctly and how to position it so that it drapes realistically in a simulation.

Lalam: They show how these two methods can be integrated with existing workflows, allowing them to go from a 2D pattern to the final three dee asset without manual adjustments required one <ref:2610.20651#pg1,3D asset without manual adjustments required>.

Tom: The authors are showing that by separating the segment-level sewing prediction from the panel-level arrangement prediction, they can achieve full automation from the input pattern.

Jane: It really shifts how you think about garment creation, moving it away from purely manual steps toward a fully automated pipeline driven by learned models.

Lu: It’s a way to address the complexity of garment construction by breaking it down into two manageable, interconnected learning tasks.

The paper's improvements: Tom: Now for the improvements they highlight. They specifically point out that their segment-level model can address many-to-many, ordered segment correspondences and handle hems and overlapping seams needed for multilayer garments one <ref:2610.20651#pg1,many-to-many, ordered segment correspondences>.

Jane: That means it handles those complex sewing operations we talked about—the ones that involve things like facing seams and hems—which were previously neglected by prior works three <ref:2610.20651#pg1>.

Lu: They also introduce specialized prediction heads. For instance, there's a parity score head to compute logits for whether a segment pair belongs to opposing sides of a seam, and an order score head for whether one segment goes immediately before another on their shared seam side four.

Meng: That’s interesting because it breaks down the prediction into specific tasks. It's not just one big prediction; it’s checking pairings, checking order, checking if something is a hem.

Lalam: They also have binary classification for hem prediction and two-layer MLP classification heads for categorical properties, and regression heads for numerical ones, acting directly on the embeddings four.

Tom: The total loss combines losses for parity, order, hem membership, and property predictions to guide the learning process.

Jane: And they show how they reconstruct seams by creating a graph with all segments as nodes and positive pairs in that parity matrix as edges to detect connected components using the Union-Find algorithm four.

Conclusion: Tom: So, wrapping up on this paper, the main implication is that these two learning-based approaches automate both sewing and arrangement using only a 2D pattern as input <ref:2610.20651#pg1>.

Jane: It means we can go from a 2D pattern to a finished three dee asset with full automation in mind, which is really something when you think about the time it used to take to manually define everything <ref:2610.20651#pg1>.

Lu: The results show that the automatic sewing model achieves a perfect recovery of sixty-three point twenty-eight percent of garments when looking at topology alone nine <ref:2610.20651#pg2>.

Meng: And for arrangement, they get sixty-six point eighteen percent of perfectly arranged garments over the whole test set ten <ref:2610.20651#pg2>.

Lalam: The paper shows that the integration into existing apparel CAD software allows users to inspect and correct prediction errors using their current editing tools eleven.

Tom: It sounds like a very robust system because it’s designed to be integrated seamlessly, even if it doesn's get every single detail perfect on its own.

Jane: So, the overall message is that this paper demonstrates how enabling full inter-panel and intra-panel attention improves overall performance in these kinds of tasks four.

Lu: And they suggest that a larger dataset with more examples of underrepresented sewing operations would improve model performance eleven.

Meng: I'm worried about the specific limitations they mention; for instance, the paper excludes more complex, although arguably rare, overlapping constructions involving three or more layers nine <ref:2610.20651#pg2>.

Lalam: They also noted that some F1 scores for properties are lower due to the stylistic nature of certain construction choices like stitching applied to regular seams in visible areas nine <ref:2610.20651#pg2>.

Tom: So they're not claiming perfection on every single detail right away, but they show a clear path toward fully automated assembly.

Jane: And that path is made possible by these two learning-based approaches described in "Towards automatic digital garment assembly".

cs.CV, cs.GR

Submitted: 2026-10-12

Updated: 2026-10-12

Comments: 11 pages, 8 figures

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 84/100

The gist: The gist: The authors present two learning-based approaches that automate both tasks—sewing and arrangement—using a 2D vector pattern as their sole input to create simulation-ready 3D garments

Key concepts

Automatic Sewing Model: This model operates at the segment level, using a Transformer encoder with multiple prediction heads to determine many-to-many correspondences between pattern segments. It predicts complex construction properties like seam allowances, stitch types, and hem membership required for the final garment.
Automatic Arrangement Model: This complementary model takes 2D panels and uses a 1D cyclic residual network combined with a Transformer to predict body-region assignments, fabric layers, and precise 2D placement coordinates. This initializes the drape simulation needed for creating the virtual garment's shape.
Segment Input Features: The input features for the sewing model combine per-panel data (like area and perimeter) with per-segment data (like Bezier segment degree and control points). This combined vector is processed by a Transformer to infer topological connections between pattern pieces.
Panel-Level Model: This model focuses on the arrangement task. It processes each garment panel as a 1D signal, using a ResNet and Transformer layers to predict which body region and layer each panel belongs to, along with its required 2D position for accurate drape simulation.

Summary

The gist: The authors present two learning-based approaches that automate both tasks—sewing and arrangement—using a 2D vector pattern as their sole input to create simulation-ready 3D garments without manual adjustments.

How it works

The automatic sewing model operates at the segment level, leveraging a Transformer encoder with multiple prediction heads to support many-to-many, ordered segment correspondences, as well as hems, overlapping seams needed for multilayer garments, and construction properties such as seam allowances and stitch types required for the finished virtual garment but neglected by previous methods<ref:2610.20651#pg2>.
A complementary panel-level model predicts body-region assignments, layers, and placement coordinates to initialize the drape simulation<ref:2610.20651#pg2>.

Automatic Sewing Model Details

** Segment input features include a set of per-panel features like the center of the panel, its area, its perimeter, and its centered axis-aligned bounding box minimum, maximum, and size for a total of 10 features<ref:2610.20651#pg4>. **

** Per-segment features consist of a one-hot encoding of the degree of the Bezier segment and the panel-centered 2D coordinates up to four control points (two points in the case of a linear Bezier, four points in the case of a cubic Bèzier), for a total of 11 features<ref:2610.20651#pg4>. **

** These segment features are concatenated with corresponding panel features to form a final feature vector of 21 values, implicitly grouping same-panel segments by their shared panel features<ref:2610.20651#pg4>. **

** The backbone is a Transformer encoder that processes all N segments at once, projecting the input feature vectors up to an embedding tensor E through a learnable linear layer and then passing them through a standard four-layer Transformer encoder with four attention heads per layer<ref:2610.20651#pg4>. **

** Full self-attention is enabled without masking or additional positional encodings because basic segment features implicitly encode enough information to deduce topological connections, and panel awareness is introduced by the concatenated panel features<ref:2610.20651#pg4>. **

** Specialized prediction heads are used for different tasks: a parity score head computes logits for whether a segment pair belongs to the opposing sides of a seam, and an order score head computes logits for whether a segment goes immediately before another one on their shared seam side<ref:2610.20651#pg4>. **

** For hem prediction, a binary classification head is used, which processes the input embedding row-wise to output whether or not a segment belongs to a hem<ref:2610.20651#pg4>. **

** Seam property heads are added to predict categorical properties using 2-layer MLP classification heads and numerical properties using 2-layer MLP regression heads, acting directly row-wise on the embeddings H<ref:2610.20651#pg4>. **

** The total loss is a sum of losses for parity, order, hem membership, and property losses<ref:2610.20651#pg4>. **

** Seam reconstruction involves creating a graph with all segments as nodes and positive pairs in the corresponding parity matrix as edges to detect connected components using Union-Find algorithm<ref:2610.20651#pg4>. **

Automatic Arrangement Model Details

The model architecture for automatic arrangement combines a one-dimensional cyclic residual network with a Transformer encoder that processes all garment panels at once<ref:2610.20651#pg6>.
The input is prepared by discretizing each of the M garment panels as a multichannel 1D signal, uniformly sampling its boundary, where each sample consists of panel-centered position, normal, and local curvature<ref:2610.20651#pg6>.
Each panel signal is processed with a 13-layer 1D ResNet to generate per-panel embeddings Ep<ref:2610.20651#pg6>.
Positional information is added by linearly projecting the center of each panel into pi in Rdp, i ∈ one to M, stacking them together and adding these projections to the panel embeddings Ep to get Hp<ref:2610.20651#pg6>.
Hp is then processed by a two-layer Transformer encoder with full self-attention to produce Hc<ref:2610.20651#pg6>.
Three independent linear prediction heads process the final embeddings Hc to output the categorical body region assignment logits, the layer assignment logits, and the 2D arrangement position<ref:2610.20651#pg6>.
The model is trained end-to-end using a total loss given by the sum of cross-entropy losses for body region and layer assignments and mean squared error (MSE) loss for the 2D arrangement position<ref:2610.20651#pg6>.

Automatic Arrangement Evaluation

** A panel is considered correctly arranged if it gets assigned to its ground truth body region and layer, and its arrangement position is within a 10 cm radius from the ground truth position<ref:2610.20651#pg10>. **

** The model achieves 66.18% of perfectly arranged garments over the total in the test set<ref:2610.20651#pg10>. **

** For imperfectly arranged garments, 73.4% have fewer than 6 panels that would require manual adjustments<ref:2610.20651#pg10>. **

Integration and Performance

The proposed methods are integrated into existing apparel CAD software, allowing the process to go from 2D pattern to finished asset in a fully automatic manner<ref:2610.20651#pg11>. The percentage of garments in the test set that are both perfectly reconstructed and correctly arranged drops to 54.34%<ref:2610.20651#pg11>. The automatic sewing model achieves a perfect recovery of 63.28% of the garments, while regular seams achieve a perfect recovery of 91.52%<ref:2610.20651#pg9>. The arrangement model achieves 66.18% correctly arranged panels and 82.61% perfectly arranged garments<ref:2610.20651#pg10>. The results support the hypothesis that enabling full inter-panel and intra-panel attention improves overall performance<ref:2610.20651#pg4>.

Further Work

The proposed sewing model excludes more complex, although arguably rare, overlapping constructions with three or more layers, and operations involving only part of a boundary segment<ref:2610.20651#pg9>. Extending parametric models to support more complex overlapping constructions, partial-segment operations, and their construction properties could provide broader and more scalable supervision<ref:2610.20651#pg11>. The results suggest that a larger dataset with more examples of underrepresented sewing operations would improve model performance<ref:2610.20651#pg9>. The F1 score for some of the properties is notably lower due to the stylistic nature of some construction choices, such as stitching applied to regular seams, particularly in visible areas of the garment<ref:2610.20651#pg9>. The results support our hypotheses that the raw Bezier features contain sufficient information for the Transformer model to infer relevant topological relationships and that enabling full inter-panel and intra-panel attention improves overall performance<ref:2610.20651#pg4>. The percentage of garments in our test set that are both perfectly reconstructed and correctly arranged drops to 54.34%, but since both methods integrate seamlessly with the existing seam creation and panel arranging tooling exposed by the CAD software, users can inspect and correct prediction errors using the existing editing tools<ref:2610.20651#pg11>. The hoodie (c) is correctly arranged, but it is missing the seam that joins the hood to the body pieces, which must be created manually<ref:2610.20651#pg11>. The sweatshirt (d) has a spurious pair connecting the left and right armhole seams, causing both seams to be bundled in the same connected component with disconnected order chains, getting discarded<ref:2610.20651#pg11>. The sweatpants (e) have front pieces assigned to the front waist region instead of their respective front thigh regions, and one pocket panel was assigned to the incorrect layer<ref:2610.20651#pg11>. The tank top (f) has all seams correctly created and all pieces are assigned to their correct body region and layer, but the arrangement position of one of them deviated ∼16 cm from its ground truth position<ref:2610.20651#pg11>. The simulation is robust to that mistake given the simplicity of the garment and no manual adjustment is needed<ref:2610.20651#pg11>.

--- Page 4 ---

Fig. 2. Example of a finished garment with hems and facing seams. The collar has two layers of fabric joined together by facing seams before being sewn to the torso panels with a regular seam. The sleeves and the shirt bottom are finished with hems<ref:2610.20651#pg4>.

--- Page 8 ---

TABLE II

RESULTS OF THE AUTOMATIC SEWING EXPERIMENTS ON THE TEST SET, WITH PERFECT RECOVERY PERCENTAGES FOR WHOLE GARMENTS AND FOR REGULAR SEAMS, FACING SEAMS AND HEMS. CORRESPONDING RESULTS FOR THE ABLATION EXPERIMENTS ARE ALSO REPORTED<ref:2610.20651#pg9>.

--- Page 10 ---

TABLE IV

RESULTS OF THE AUTOMATIC ARRANGEMENT EXPERIMENTS ON THE TEST SET, LISTING PERCENTAGE OF CORRECTLY ARRANGED PANELS AND GARMENTS. RESULTS FOR THE ABLATION EXPERIMENT ARE ALSO REPORTED<ref:2610.20651#pg10>.

--- Page 11 ---

[27], suggesting that a larger dataset with more examples of underrepresented sewing operations would improve model performance<ref:2610.20651#pg9>.

--- Page 5 ---

Fig. 3. Proposed automatic sewing model architecture. Per-segment features X, including panel-level features, are input to the model.

Improvements for AI systems

Bold segment-level learning architecture for automatic sewing: The system can now address many-to-many, ordered segment correspondences, as well as hems, overlapping seams needed for multilayer garments, which was previously neglected by prior methods.
Panel-level arrangement prediction: The system can automate the initialization of drape simulation by predicting body-region assignments, layers, and placement coordinates to initialize the drape simulation, moving beyond manual placement in parametric spaces.
Unified sewing operation prediction: The model predicts regular seams, as well as facing seams and hems, along with seam allowances, folding directions, and stitch types, addressing operations previously ignored by previous solutions that only focused on seam topology.
Exact recovery of garment topology: The system achieves a perfect recovery of 63.28% of the garments when considering only topology (correct per-side ordered segments), indicating robust inference even when properties are imperfectly recovered, saving significant manual work.
Automated garment assembly pipeline: The integrated system can go from 2D pattern to finished 3D asset without manual adjustments required, streamlining the workflow from pattern input to simulation-ready configuration.

Abstract

Creating digital garments from existing two-dimensional patterns remains a time-consuming process. Although patterns are widely available in digital form, their assembly in garment CAD software involves manual definition of sewing operations and arrangement of the pattern panels for drape simulation. We present two learning-based approaches that automate both tasks using the 2D vector pattern as their sole input. Our automatic sewing model operates at the segment level, leveraging a Transformer encoder with multiple prediction heads. Our novel architecture supports many-to-many, ordered segment correspondences, as well as hems, overlapping seams needed for multilayer garments, and construction properties such as seam allowances and stitch types, required for the finished virtual garment but neglected by previous methods. A complementary panel-level model predicts body-region assignments, layers, and placement coordinates to initialize the drape simulation. We evaluate both models on a curated dataset of 4,035 professionally assembled garment instances. The sewing model exactly recovers all seam and hem topologies of 63.3% of test garments, while the arrangement model meets the specified placement criteria for all panels in 66.2% of test garments. We also integrate both methods into garment CAD software, demonstrating automatic assembly of successfully predicted garments.