This is a free sample of a much larger licensed dataset of real textiles photographed in-house by Beglarian Fabrics, a European deadstock fabric house.

Full dataset: 50,000+ fabrics, 170,427 HD photos and 46,758 drape videos captured 2020–2026. Commercial, non-exclusive licences for model training — see data.beglarianfabrics.com.

Why this data

One capture protocol. Since 2023 every fabric is photographed to the same set of views on a light background: a swirl fold, a flat view with a ruler for scale, close-ups and the reverse side.
Structured labels, not scraped captions. Labels come from the product catalogue of the team that buys, measures and sells each fabric: fibre composition with exact percentages, colour, fabric type, weave, stretch, features, intended use, width and weight (g/m²).
Clean provenance. Every file was captured in-house; nothing is scraped. No faces, no personal data. Records use anonymous sample IDs.

What's in the sample

100 fabrics hand-picked across 33 fabric types for complete labels, image quality and visual variety, captured 2024–2026.
3–7 photos per fabric (different views of the same fabric), resized to 1024 px on the long side.
One row per image in data/train/metadata.jsonl; rows of the same fabric share a sample_id.

Fields

Field	Description
`file_name`	Image file
`sample_id`	Anonymous fabric ID (group by it to keep all views of one fabric in the same split)
`view`	Photo index for this fabric (0 = main view)
`captured`	Capture month (YYYY-MM)
`fabric_type`	Fabric type(s), e.g. `Jacquard`, `Woven`
`color`	Colour family and pattern, e.g. `Shades of blue`, `Plain`, `Printed` (multi-label)
`composition`	Fibres with percentage, e.g. `{"fiber": "Cotton", "pct": 47}`
`weave`	Weave / construction
`stretch`	Stretch class
`features`	Finish, pattern and other features (e.g. `Paisley`, `Coated fabric`)
`intended_use`	Typical end uses
`country_of_origin`	Country where the fabric was made
`width_cm`, `weight_gsm`	Roll width and fabric weight
`caption`	Short caption generated from the labels
`title_en`, `description_en`, `description_fr`	Catalogue title and description

Example record

{
  "file_name": "bf-f95ec0f58b_0.jpg",
  "sample_id": "bf-f95ec0f58b",
  "view": 0,
  "captured": "2024-11",
  "fabric_type": [
    "Tweed",
    "Woven"
  ],
  "color": [
    "Shades of gray"
  ],
  "composition": [
    {
      "fiber": "Polyamid",
      "pct": 85.0
    },
    {
      "fiber": "Cotton",
      "pct": 15.0
    }
  ],
  "weave": [
    "Twill Weave"
  ],
  "stretch": [
    "Non-stretch"
  ],
  "features": [
    "Panel Print",
    "Checked",
    "Houndstooth"
  ],
  "intended_use": [
    "Dress/Blouse",
    "Suit"
  ],
  "country_of_origin": [
    "Italy"
  ],
  "width_cm": 142.0,
  "weight_gsm": 270.0,
  "caption": "Gray tweed / woven, 85% polyamid, 15% cotton, 270 g/m², non-stretch",
  "title_en": "Checked Houndstooth Tweed Fabric",
  "description_en": "Italian shades of gray tweed fabric with polyamid and cotton for dress and suit, dense and non-stretch with panel print.",
  "description_fr": "Tweed italien en nuances de gris avec polyamide et coton pour robe et costume, dense et non extensible avec placée."
}

Usage

from datasets import load_dataset

ds = load_dataset("BeglarianFabrics/textile-dataset-sample", split="train")
print(ds[0]["caption"], ds[0]["composition"])

Licence

The sample is released under CC BY-NC 4.0: evaluation, research and other non-commercial use, with attribution to Beglarian Fabrics. Training or fine-tuning commercial models requires a commercial licence for the full dataset.

Contact

Licensing, custom subsets and the full corpus: kb@beglarianfabrics.com · data.beglarianfabrics.com

Downloads last month: -

Total file size:

130 MB