HindiText-Pilot

4,000 synthetic 512×512 image/text pairs for training diffusion models to render Hindi text in Devanagari script. Correct HarfBuzz shaping, exact ground-truth transcriptions, per-sample glyph masks.

Built as a feasibility pilot for a Stable Diffusion 1.5 LoRA, and released because the hard parts — correct shaping and the VAE size floor — are reusable regardless of what you train.

Note on scale. This is a pilot: 72 distinct words. It exists to answer "can a model learn Devanagari glyph shapes at all", not to teach the script broadly. See Limitations before using it as a training set.

Why this exists

Most Devanagari datasets target recognition (OCR) and are deliberately adversarial — blur, occlusion, low resolution. A generator trained on those learns to produce blurry text. This is the opposite: clean, sharp, and size-constrained so every sample is something a generator should aim to match.

What makes Devanagari hard

• Shirorekha — a continuous horizontal bar joins letters across a word. Latin-trained models produce broken or wavy bars.
• Conjuncts — क् + ष → क्ष. Hundreds of ligatures whose shapes are not compositional from their parts.
• Reordering — the matra ि is written before the consonant it follows logically.
• Three vertical zones — matras sit above and below the base line.
• Fine diacritics — nukta ़, anusvara ं, chandrabindu ँ are 2–8 pixel features that vanish under aggressive downsampling.

Contents

data/images/           4,000 × 512×512 JPEG (quality 95)
data/masks/            4,000 × 512×512 PNG, 8-bit glyph coverage masks
data/metadata.jsonl    one record per sample
data/train_words.txt   72 words used in the images
data/holdout_words.txt 20 words deliberately NOT rendered
fonts/                 22 OFL Devanagari fonts + fonts/licenses/
corpus/pilot_words.txt the source word list
src/                   the generator — this dataset is reproducible
FINDINGS.md            the measurements behind the design

metadata.jsonl

field	meaning
file	path relative to data/
text	the Hindi string rendered — exact ground truth
caption	training caption (varied template)
tokens	CLIP token count of the caption, template included
font	filename in fonts/
em	font em size in pixels
x, y, w, h	text bounding box
glyphs	shaped glyph count — differs from character count due to ligatures

{"file": "images/000000.jpg", "text": "रुकें",
 "caption": "Hindi text \"रुकें\" on a board, flat design", "tokens": 18,
 "font": "Khand-Regular.ttf", "em": 40, "x": 150, "y": 405,
 "w": 62, "h": 48, "glyphs": 3}

Statistics

samples	4,000
resolution	512 × 512
words	72 train, 20 held out (never rendered)
fonts	22, evenly sampled (156–211 samples each)
font em size	32–120 px, mean 75.5
text box	44–454 px wide, 30–161 px tall
caption tokens	11–29 of CLIP's 77
glyphs per sample	2–8

Vocabulary spans four groups: simple no-conjunct words, real signage terms, conjunct-bearing words (क्ष ज्ञ त्र श्र द्व ह्म), and diacritic stress cases (nukta, anusvara, chandrabindu, visarga).

How it was generated

Text is shaped with HarfBuzz (uharfbuzz) and rasterised per-glyph with FreeType (freetype-py).

⚠️ If you extend this, do not use Pillow for the text

Pillow's published wheels ship without Raqm, so ImageDraw.text() renders Devanagari unshaped: conjuncts break into separate letters and pre-base ि lands on the wrong side of its consonant. The output looks plausible to a non-reader while being wrong, and silently corrupts the dataset.

Check with PIL.features.check('raqm'). If it returns False, use src/render/shaper.py here instead.

Verified by src/diag/verify_shaping.py, which passes on all 22 fonts: क्ष / त्र / ज्ञ each collapse 3 codepoints → 1 glyph, and the base consonant lands second in कि, proving the matra reordered ahead of it.

Backgrounds are flat, gradient or lightly textured, with foreground/background luminance contrast enforced (≥ 0.35) rather than sampled and hoped for. Captions use 10 templates × 12 surface nouns with the quoted Hindi string as the single invariant anchor; the phrase "Hindi text" is dropped in ~20% of samples so a model does not become dependent on a trigger phrase.

The 32px rule

Every sample renders text at font em size ≥ 32px, asserted at generation time. SD1.5's 4-channel VAE downsamples 8×; measured encode→decode round-trips show Devanagari diacritics are destroyed below ~24px em and clean at 32px:

font em size	VAE reconstruction
12–20 px	destroyed — nuktas gone, conjuncts illegible
24 px	marginal — nukta merges into the stroke
32 px	clean — minimum safe size
48–64 px	near-perfect

Note this is em size, not ink-box height: the measured ink/em ratio varies 0.67–1.08 depending on which matras a word carries, so ink height is content-dependent and unsuitable as a threshold.

Consequence: this covers headline- and signage-sized text, roughly 1–4 words. It is not suitable for training small or body-sized text rendering on a 4-channel VAE at 512px — that is a reconstruction limit no training run can overcome.

Intended use

• Fine-tuning diffusion models (LoRA or full) for Devanagari text
• Evaluating text-rendering quality, with exact transcriptions as ground truth
• Glyph-conditioned approaches (ControlNet, GlyphControl, AnyText) — masks are included for this
• The 20 held-out words test generalisation rather than memorisation

Not suitable for OCR / scene-text recognition training (these are clean synthetic renders, not photographs), or small-text rendering.

Limitations

• Synthetic only. No photographic backgrounds, perspective, occlusion or real-world lighting. A model trained solely on this renders clean text but may not composite convincingly into photographs.
• 72 words is pilot scale — far too small to teach the script broadly.
• Partial conjunct coverage. श्च has zero coverage in the rendered samples.
• One script, one language. Devanagari as used for Hindi. Marathi and Nepali share the script with minor differences and are not specifically covered.

Licensing

• Images, masks, metadata, generator code — generated from the hand-written word list in corpus/pilot_words.txt using OFL fonts. No third-party data. CC BY 4.0.
• fonts/ — 22 families from Google Fonts under the SIL Open Font License 1.1, which permits redistribution. Full licence texts are included in fonts/licenses/. The OFL applies to those files, not CC BY.

Nothing in this repository is derived from any dataset whose terms prohibit redistribution.

Reproducing

pip install uharfbuzz freetype-py pillow numpy transformers
python src/diag/get_fonts.py          # fetch the 22 OFL fonts
python src/diag/font_audit.py         # coverage + conjunct-formation check
python src/diag/verify_shaping.py     # prove HarfBuzz shaping is active
python src/data/gen_pilot.py --n 4000 --out data/pilot --masks --seed 42

Citation

@misc{hinditext_pilot_2026,
  title  = {HindiText-Pilot: synthetic Devanagari text-rendering pairs},
  author = {Sinhal},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sinhal/HindiText-Pilot}
}