README
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-x4-48followed by Lanczos resizing; 11 are at their original size. Each image's note insources.csvsays 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.
Similar repos
README
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-x4-48followed by Lanczos resizing; 11 are at their original size. Each image's note insources.csvsays 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.