SwiftTailor: Efficient 3D Garment Generation with Geometry Image Representation

Phuc Pham; Uy Dieu Tran; Binh-Son Hua; Phong Nguyen

cs.CVMar 19, 2026

SwiftTailor: Efficient 3D Garment Generation with Geometry Image Representation

Authors: Phuc Pham, Uy Dieu Tran, Binh-Son Hua, Phong Nguyen

Organizations: Qualcomm AI Research · Trinity College Dublin, Ireland.

Abstract

Realistic and efficient 3D garment generation remains a longstanding challenge in computer vision and digital fashion. Existing methods typically rely on large vision- language models to produce serialized representations of 2D sewing patterns, which are then transformed into simulation-ready 3D meshes using garment modeling framework such as GarmentCode. Although these approaches yield high-quality results, they often suffer from slow inference times, ranging from 30 seconds to a minute. In this work, we introduce SwiftTailor, a novel two-stage framework that unifies sewing-pattern reasoning and geometry-based mesh synthesis through a compact geometry image representation. SwiftTailor comprises two lightweight modules: PatternMaker, an efficient vision-language model that predicts sewing patterns from diverse input modalities, and GarmentSewer, an efficient dense prediction transformer that converts these patterns into a novel Garment Geometry Image, encoding the 3D surface of all garment panels in a unified UV space. 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. Extensive experiments on the Multimodal GarmentCodeData demonstrate that SwiftTailor achieves state-of-the-art accuracy and visual fidelity while significantly reducing inference time. This work offers a scalable, interpretable, and high-performance solution for next-generation 3D garment generation.

Figures & tables

Figure 2: Preliminaries on geometry images [ 11 , 38 ] , an image-based 3D representation that parameterizes a 3D mesh into charts, each being stored as simple arrays of pixels. Our work integrates geometry images with semantic and stitching information to establish garment panels, yielding a novel garment geometry image representation suitable for 3D garment generation.

Figure 3: Overall pipeline. Our PatternMaker is a relatively small vision-language model (InternVL-3-2B [ 46 ] ) trained to output sewing patterns. The sewing patterns are constructed from discrete tokens and continuous parameters predicted by the VLM. Our GarmentSewer is a dense prediction transformer (DPT) that predicts a garment geometry image from the sewing patterns. In this step, we preprocess the sewing pattern to achieve the semantic and stitching map, which are then passed to the DPT to predict the geometry image, completing our garment geometry image representation (GGI). We then perform a postprocessing step to convert the GGI to a final 3D mesh.

Figure 4: ( Left ) We present how to prepare the three components (geometry, semantic and stiching) of our propose Garment Geometry Image (GGI); ( Right ) From the estimated geometry and stiching images of GarmentSewer and PatternMaker, two additional remeshing and stiching steps are performed to obtain the final 3D mesh result.

Method	Vertex L2 ( $\downarrow$ )	#Panel Acc ( $\uparrow$ )	#Edge Acc ( $\uparrow$ )	Rot L2 ( $\downarrow$ )	Transl L2 ( $\downarrow$ )	Stitch Acc ( $\uparrow$ )
AIpparel [ 32 ]	4.8 / 2.5	93.7 / 82.9	79.0 / 88.2	0.007 / 0.002	2.5 / 1.7	73.0 / 86.3
ChatGarment [ 1 ]	14.9 / 13.5	16.4 / 19.6	49.7 / 60.3	0.038 / 0.036	15.4 / 10.1	38.2 / 55.2
SewingLDM [ 27 ]	15.6	18.0	49.0	0.052	16.6	30.6
PatternMaker (Ours)	3.5 / 1.5	94.8 / 85.0	92.3 / 98.0	0.006 / 0.002	1.9 / 1.3	85.1 / 97.8

Table 1: Quantitative results on sewing-pattern generation (left) and editing (right). Best results are shown in bold .

Method	MMD $\downarrow$	COV $\uparrow$	#Sampling $\downarrow$
AIpparel [ 32 ] + GC [ 18 ]	6.94	0.52	4.27
ChatGarment [ 1 ] + GC [ 18 ]	12.27	0.22	1.20
SewingLDM [ 27 ] + GC [ 18 ]	11.33	0.34	5.87
PatternMaker + GC [ 18 ]	6.82	0.54	2.98
SwiftTailor (Ours)	5.31	0.68	2.98

Table 2: Quantitative results on mesh generation using multi-modal inputs (image and text). Best results are shown in bold .

Method	Stage 1	Stage 2	Post-proc.	Total
AIpparel [ 32 ] + GC [ 18 ]	10.20	49.55	3.99	63.74
ChatGarment [ 1 ] + GC [ 18 ]	18.85	33.92	5.76	58.53
SewingLDM [ 27 ] + GC [ 18 ]	5.44	46.87	2.16	54.47
PatternMaker + GC [ 18 ]	9.93	37.45	2.18	49.56
SwiftTailor (Ours)	9.93	0.02	4.83	14.78

Table 3: Running time comparison to obtain the final mesh (in seconds) between other baselines and our SwiftTailor. Stage 1 is generating patterns, while Stage 2 is constructing mesh from them.

Figure 5: Qualitative comparisons between SwiftTailor and recent state-of-the-art methods on 3D garment modeling [ 32 , 1 , 27 ] using an image, a text prompt, and both text and image as input, respectively.

Ablation	CD $\downarrow$	EMD $\downarrow$	MMD $\downarrow$	COV $\uparrow$
W/o Semantic UV Map	35.77	9.90	11.96	0.49
W/ Semantic UV Map
+ $\mathcal{L}_{\text{reg}}$	9.84	5.95	7.38	0.58
+ $\mathcal{L}_{\text{stitch}}$ (Ours)	3.40	4.48	3.36	0.88

Table 4: Ablation on semantic UV map and auxiliary losses

Figure 6: Qualitative results on generated 3D mesh using geometry vs. binary image as input to GarmentSewer.

Figure 7: Qualitative results on ablation study of training losses.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure B.1 : Effect of interpolation schemes on the geometry image. We first rasterize the mesh into a geometry image using different interpolation strategies and then remesh it back into 3D. Barycentric interpolation alone ( left ) introduces jagged and discontinuous boundary values that deviate from the true panel contour. Our hybrid interpolation ( right ), which applies linear interpolation along panel edges and barycentric interpolation only inside triangle interiors, produces smooth and consistent boundary signals, preventing these artifacts from propagating into GarmentSewer predictions.

Figure B.2 : Remeshing from the geometry image. Starting from the UV-aligned geometry image, we perform local triangular remeshing by examining each $2\times 2$ UV cell and generating one or two triangles depending on valid vertex occupancy. When all four vertices are present, the diagonal yielding the shorter 3D distance is selected. All faces are constructed in clockwise order to ensure consistent outward-facing normals across panels.

Figure B.3 : Stitching results before and after seam alignment. Using the stitching image, boundary edges are paired and aligned via Dynamic Time Warping, followed by vertex merging through a disjoint-set union. The zoomed-in wireframe views highlight how stitching resolves discontinuities and removes gaps between corresponding panel edges, achieving globally coherent garment mesh. The example is conducted on predicted sewing pattern from PatternMaker.

Method	Vertex L2 ( $\downarrow$ )	#Panel Acc ( $\uparrow$ )	#Edge Acc ( $\uparrow$ )	Rot L2 ( $\downarrow$ )	Transl L2 ( $\downarrow$ )	#Stitch Acc ( $\uparrow$ )
AIpparel [ 32 ]	5.18	89.94	75.76	0.007	2.51	71.04
ChatGarment [ 1 ]	16.47	14.05	39.08	0.057	19.99	30.58
SewingLDM [ 27 ]	19.41	15.42	42.77	0.107	25.04	28.17
Ours	3.70	91.04	88.96	0.006	1.91	83.69

Appendix

Table C.1 : Quantitative results on sewing-pattern generation with image condition only. Best results are shown in bold .

Method	MMD $\downarrow$	COV $\uparrow$	#Sampling $\downarrow$
AIpparel [ 32 ] + GC [ 18 ]	6.95	0.52	3.93
ChatGarment [ 1 ] + GC [ 18 ]	11.64	0.22	1.31
SewingLDM [ 27 ] + GC [ 18 ]	10.56	0.37	2.07
PatternMaker + GC [ 18 ]	6.82	0.56	2.93
Ours	5.23	0.68	2.93

Appendix

Table C.2: Quantitative results on mesh generation using image condition only. Best results are shown in bold .

Method	MMD $\downarrow$	COV $\uparrow$	#Sampling $\downarrow$
AIpparel [ 32 ] + GC [ 18 ]	8.59	0.39	3.76
ChatGarment [ 1 ] + GC [ 18 ]	12.89	0.20	1.15
SewingLDM [ 27 ] + GC [ 18 ]	19.97	0.26	4.80
PatternMaker + GC [ 18 ]	8.58	0.43	2.70
Ours	7.80	0.42	2.70

Appendix

Table C.3: Quantitative results on mesh generation using text condition only. Best results are shown in bold .

Method	MMD $\downarrow$	COV $\uparrow$	#Sampling $\downarrow$
AIpparel [ 32 ] + GC [ 18 ]	6.94	0.52	4.27
AIpparel [ 32 ] + GarmentSewer	6.03	0.63	4.27
ChatGarment [ 1 ] + GC [ 18 ]	12.27	0.22	1.20
ChatGarment [ 1 ] + GarmentSewer	12.56	0.23	1.20
SewingLDM [ 27 ] + GC [ 18 ]	11.33	0.34	5.87
SewingLDM [ 27 ] + GarmentSewer	10.96	0.41	5.87

Appendix

Table C.4: Quantitative results of all combinations between pattern generator and garment constructor on mesh generation using multi-modal inputs (image and text).

Figure D.1 : Qualitative comparison between GarmentCode [ 18 ] and GarmentSewer given the same sewing patterns predicted by PatternMaker. GarmentSewer produces stable initializations and consistent draping, while GarmentCode [ 18 ] often fails due to rule-based panel placement.

Figure D.2 : Additional qualitative results from our pipeline. Each example shows the re-draped garment on the SMPL [ 28 ] body together with its initial state constructed by GarmentSewer (the smaller mesh on the left). Textures are added to enhance visualization of garment geometry and structure.

Explore similar work

Jul 1, 2026

Stitched Embeddings: A Unified Latent Space for 3D Garments and 2D Patterns

While garments are essential for realistic digital humans, their topological variety makes them much harder to model than parametric bodies. Traditional tailoring relies on 2D sewing patterns, yet bridging these patterns to 3D geometry currently requires physical simulations. We present Stitched Embeddings, the first simulation-free framework to unify 3D garment reconstruction and sewing pattern inference within a single bidirectional latent space. By leveraging the geometric priors of a pretrained 3D foundation model, our approach overcomes the data scarcity typically associated with high-quality garment modeling. We propose to use the BoxMesh as a critical intermediate representation to align 2D panels into 3D configurations without the computational overhead of a simulator. This architecture achieves state-of-the-art accuracy in pattern reconstruction while significantly improving efficiency. Furthermore, our differentiable pipeline enables novel applications, including pattern recovery from meshes and 3D editing from 2D patterns. Finally, this work provides a scalable link between neural 3D vision and the physical garment manufacturing pipeline. Project Page: https://andreus00.github.io/stitchedembeddings

Andrea Sanchietti, Riccardo Marin, Bharat Lal Bhatnagar +2

University of Tübingen, Germany · Tübingen AI Center · Technical University of Munich, Germany +3

3D Reconstruction Multimodal Embedding

Ask PDF

Jul 25, 2026

Fashion-3DLR: A Controllable 3D Garment Generation Using Pairwise Fashion Elements for Intelligent Design

AI-generated content (AIGC) has made significant progress, with 2D generative models becoming ready-to-use tools for the digital fashion industry. However, 3D garment generation remains in its nascent stage, where in the realm of fashion, the semantic information of diverse design elements exhibits intricate coupling relationships in 3D representations, posing substantial challenges for generating diverse 3D garments. In this work, to handle the above problem, We introduce Fashion-3DLR, a novel 3D garment generation framework that utilizes diverse design elements to create high-quality, versatile 3D garment assets. Specifically, to bridge the semantic gaps between different fashion elements, we propose a Garment Feature Fusion Diffusion Transformer (GFF-DiT) module to integrate 2D fashion design elements, e.g., sketch and texture, into latent space. Within the latent space, we then employ a rectified flow transformer to generate geometry latents, which can be decoded into various 3D garment representations, including 3D Gaussians and meshes. Furthermore, we integrate Fashion-3DLR into downstream tasks, achieving the 3D Gaussian Splatting (3DGS)-driven cloth physical simulation and mesh-based virtual try-on. Experimental results indicate that Fashion-3DLR surpass the previous state-of-the-art methods, which verify that the proposed work can generate well-structured, non-watertight garments capable of physical simulation and virtual try-on, underscoring its potential as a versatile 3D garment design tool.

Shenghao Yang, Hongtao Zhang, Yuhan Yi +6

Donghua University Shanghai, China · Shanghai Artificial Intelligence Laboratory Shanghai, China

3D Asset Generation Conditional Generative Modeling

Ask PDF

Apr 3, 2026

InverseDraping: Recovering Sewing Patterns from 3D Garment Surfaces via BoxMesh Bridging

Recovering sewing patterns from draped 3D garments is a challenging problem in human digitization research. In contrast to the well-studied forward process of draping designed sewing patterns using mature physical simulation engines, the inverse process of recovering parametric 2D patterns from deformed garment geometry remains fundamentally ill-posed for existing methods. We propose a two-stage framework that centers on a structured intermediate representation, BoxMesh, which serves as the key to bridging the gap between 3D garment geometry and parametric sewing patterns. BoxMesh encodes both garment-level geometry and panel-level structure in 3D, while explicitly disentangling intrinsic panel geometry and stitching topology from draping-induced deformations. This representation imposes a physically grounded structure on the problem, significantly reducing ambiguity. In Stage I, a geometry-driven autoregressive model infers BoxMesh from the input 3D garment. In Stage II, a semantics-aware autoregressive model parses BoxMesh into parametric sewing patterns. We adopt autoregressive modeling to naturally handle the variable-length and structured nature of panel configurations and stitching relationships. This decomposition separates geometric inversion from structured pattern inference, leading to more accurate and robust recovery. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the GarmentCodeData benchmark and generalizes effectively to real-world scans and single-view images.

Leyang Jin, Zirong Jin, Zisheng Ye +4

Mohamed bin Zayed University of Artificial Intelligence · The Chinese University of Hong Kong, Shenzhen · School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, FNii-Shenzhen, and Guangdong Provincial Key Laboratory of Future Networks of Intelligence +1

3D Reconstruction Deformable Shape Reconstruction

Ask PDF