← sources

Made an open-source AI model that generates vintage Indian matchbox labels (Mayur, Love Birds, Tiger, Jumbo...)

source: reddit (reddit:IndiaTech)published 2026-10-03 17:27ingested 2026-10-04 02:36

retained text

Remember the matchbox labels with a peacock and "MAYUR" in big letters, or a tiger in a yellow circle, from Sivakasi and Kovilpatti match works? I trained image models to make new ones from a text prompt.

**How it works**

* I collected 197 vintage Indian matchbox labels (Wikimedia Commons, Internet Archive, a few from other sites), upscaled the small scans, removed duplicates and wrote a caption for each one: what's shown, the background colour, the headline
* Then I fine-tuned two open models with LoRA: **SDXL** and **Stable Diffusion 3.5-medium**, on Google Colab with an L4 GPU (about 2 hours and 1.25 hours of training)
* Trigger word: `phlmx`. Prompt shape: `phlmx, a matchbox label, a roaring tiger's head inside a yellow circle, red background, headline 'TIGER'`

**Results**

* SD3.5-medium gets short English headlines right most of the time and has the bright, flat look of a fresh print
* SDXL looks more like an old, worn label but misspells more
* Neither can write real Devanagari, Tamil or Malayalam yet: they draw script-like shapes. That's the next problem I want to solve

Everything is open: notebooks, weights and the dataset.

* GitHub: https://github.com/spearb0lt/Indian-Matchbox-Art-Style-Text-to-Image-Generator
* Weights: https://huggingface.co/spearb0lt/Indian-Matchbox-Art-Style-Text-to-Image-Generator
* Dataset: https://huggingface.co/datasets/spearb0lt/Indian-Matchbox-Labels

If you have old labels from your family's collection (good scans, especially regional-language ones), I'd love to add them.

---

document fetched 2026-10-04 00:25 · canonical https://reddit.com/r/IndiaTech/comments/1wwst3p/made_an_opensource_ai_model_that_generates