It is built to support training and evaluation of methods that automate the creation of 2D sewing patterns, the construction of sewing relationships, and the synthesis of 3D garment initializations compatible with physics-based simulation.

This dataset facilitates versatile applications including scalable garment generation from multi-modal design concepts (text prompts, sketches, photographs), automatic modeling from raw flat sewing patterns, pattern recovery from unstructured point clouds, and progressive garment editing.

Data Statistics

GarmageSet comprises 14,801 professionally designed garments with detailed structural and style annotations. The following figures show mesh statistics of our dataset (first row of the figure) along with a comparison to the GarmentCodeData dataset (second row of the figure) through the following metric:

Avg. Panels per Garment: Average panels number per garment (Fig a).
Avg. Edges per Panel: Average edges number per panel (Fig b).
Avg. Stitch: Average stitches number per garmgent (Fig c) .
Avg. Vertices: Average number of vertices per garmgent mesh (Fig d).
Avg. Faces: Average number of faces per garmgent mesh (Fig e).

Dataset	Avg. Panels per Garment	Avg. Edges per Panel	Avg. Stitch	Avg. Vertices	Avg. Faces
GarmageSet	13.59 ± 7.89	8.62 ± 5.01	45.95 ± 25.56	72,560.5 ± 27,886.7	141,064.1 ± 55,899.8
GarmentCodeData	10.82 ± 6.29	6.75 ± 4.20	30.11 ± 17.68	13,998.9 ± 8,658.4	26,400.8 ± 16,751.8

Data Format

GarmageSet is organized using a prefix-based sharding strategy. Files are bucketed based on the first two characters (hexadecimal prefix) of their unique identifier (ID), where each bucket is stored as a separate .tar.gz archive (e.g., 00.tar.gz, 0a.tar.gz).

DATASET_ROOT
├── raw/                        # 3D Mesh Assets
│   ├── 00.tar.gz               # Shard containing IDs starting with '00'
│   │   ├── 00xxxx.obj          # Raw 3D geometry file
│   │   └── ...
│   ├── 0c.tar.gz               # Shard containing IDs starting with '0c'
│   └── ...
│
├── images/                       # Multi-view Renderings & Camera Data
│   ├── 00.tar.gz                 # Image shard matching the 'raw' file grouping
│   │   ├── 00xxxx.zip            # Per-asset container
│   │   │   ├── transform.json    # Camera intrinsics/extrinsics (NeRF format)
│   │   │   ├── xxxx_0.png        # Visual Sketch
│   │   │   ├── xxxx_0_image.png  # Grey-scale rendered view
│   │   │   └── ...
│   │   └── ...
│   └── ...
│
└── caption.csv                 # Metadata and Text Descriptions

We adopt a modified Wavefront OBJ schema to encode both the 3D geometry and specific construction data, such as 2D pattern panels and stitching topology. The field tags are reinterpreted as follows:

v : per-vertex 3D coordinate (world coordinate, unit mm).
vn : per-vertex normal.
vt : per-vertex 2D sewing pattern coordinates. Unlike standard UVs, these are unnormalized physical dimensions in mm (non-standard definition).
g : each group corresponds to a distinct cloth piece (non-standard definition).
l : stitched vertex pairs (non-standard definition).

# 3D vertex positions in the world coordinate system (x, y, z), measured in millimeters.
v -140.484 1086.47 33.9467
v ...
...

# per-vertex normals in (x, y, z) form (might not be unit vectors).
vn -0.742791 0.335566 -0.57936
vn ...
...

# UV coordinates aligned with the sewing pattern (2D world coordinate system, in millimeters).
vt -689.692 1233.65 0
vt ...
...

# Each group corresponds to a cloth piece (i.e., a sewing pattern panel in 2D).
g 7f9b2d8a-7088-4ad7-a408-ddcac965282e
usemtl default_fabric_front
f 101/101/101 99/99/99 95/95/95
f ...
...
g 830b1561-dfd1-436e-a054-52fc99438508
usemtl default_fabric_front
f 1646/1646/1646 1641/1641/1641 1648/1648/1648
f ...
...

# Sewing vertex pairs.
l 30369 29870
l ...
...

References

For further details, please refer to our project page, or check out GitHub to access the dataset toolkit.

Citation

If you find this work useful, please cite the original paper:

@article{li2025garmagenet,
  title     = {{GarmageNet}: A Multimodal Generative Framework for Sewing Pattern Design and Generic Garment Modeling},
  author    = {Li, Siran and Liu, Ruiyang and Liu, Chen and Wang, Zhendong and 
               He, Gaofeng and Li, Yong-Lu and Jin, Xiaogang and Wang, Huamin},    
  address   = {New York, NY, USA},
  articleno = {216},
  doi       = {10.1145/3763271},
  issue_date= {December 2025},
  journal   = {ACM Trans. Graph. (SIGGRAPH Asia)},
  keywords  = {garment modeling, garment dataset, diffusion Generation},
  month     = {December},
  number    = {6},
  numpages  = {23},
  publisher = {Association for Computing Machinery},
  url       = {https://doi.org/10.1145/3763271},
  volume    = {44},
  year      = {2025}
}

Downloads last month: 119