README.md

metadata

license: other
license_name: mixed-provenance-unverified
pretty_name: Indian Matchbox Labels
task_categories:
  - text-to-image
language:
  - en
size_categories:
  - n<1K
tags:
  - matchbox
  - india
  - vintage
  - style
  - lora-training
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - '*.jpg'
          - metadata.csv

Indian Matchbox Labels

197 vintage Indian matchbox labels, one label per image, each with a hand-written English caption. It was assembled to train the two style LoRAs in spearb0lt/Indian-Matchbox-Art-Style-Text-to-Image-Generator (trigger word phlmx); the training notebooks are on GitHub.

Contents

file	what
mb_001.jpg to mb_197.jpg	the labels, JPEG, short side 1 039 to 2 048 px (134 portrait, 63 landscape or square)
mb_001.txt to mb_197.txt	the caption of each image, same name (the kohya convention, read by the notebooks)
metadata.csv	file_name, text (the caption), source, source_url: what the dataset viewer and load_dataset read
sources.csv	per image: size, caption, source, source URL, original file name, and processing notes

from datasets import load_dataset
ds = load_dataset("spearb0lt/Indian-Matchbox-Labels", split="train")   # columns: image, text, source, source_url

Captions

Each caption describes the content and the printed text, not the style, so that a LoRA's trigger word, not ordinary words, absorbs the look. The shape is a matchbox label, <what is shown>, <background colour>, headline '<main word>', text '<other printed text>':

a matchbox label, a white lion lying on a blue ribbon in front of a yellow sunburst, red background, headline 'LION', text 'Safety Matches Best Quality' and 'Pugalgiri Match Works, Dindigul'

Text in Indian scripts is named (for example "Malayalam lettering at the top") rather than transcribed.

Sources and processing

source	images
Wikimedia Commons (scans from the MatchboxPosters collection)	131
Internet Archive (matchboxpostersarchive)	46
collected from other websites (original site not recorded)	20

• Upscaling. Most originals are small scans, so 186 images were upscaled 2x or 4x with caidas/swin2SR-compressed-sr-x4-48 followed by Lanczos resizing; 11 are at their original size. Each image's note in sources.csv says which.
• Cropping. Two images were cropped to the label.
• Duplicates. Checked with SHA-256, a perceptual hash and DINO image embeddings; none remain.

Licence and intended use

The labels are historical commercial prints. Where the source file name gives a date, it is mostly "circa 1940s" (103 images), with some "circa 1920s" (17), "mid-century" (15) and "1960s" (2); 60 have no date. These dates come from the uploaders and were not verified. Their copyright status has not been verified: the Wikimedia Commons and Internet Archive items carry the uploaders' own licence statements, and the 20 images from other websites have unknown rights. The dataset is shared for research and education only; check the rights of individual images before any other use. The per-image source_url lets you do that.