Srijika: OpenType-Layout-Reusing Font Restyling for Nine Indic Scripts

Srijika. OpenType-Layout-Reusing Font Restyling for Nine Indic Scripts. By Anil Pai, from Loopdesk Technologies. Published on the fourth of September. No code repository is linked to it.

Making a font for an Indic script is not a matter of drawing an alphabet. The central challenge is combinatorial. Hundreds to thousands of conjuncts, half forms and matra variants all have to stay mutually consistent under OpenType shaping, the process that turns a string of characters into the right sequence of glyphs on screen. A generator that draws each glyph beautifully can still produce a font that shapes text wrongly, or is not a complete font at all.

Srijika produces installable OpenType fonts for nine Brahmic scripts: Devanagari, Tamil, Bengali, Telugu, Kannada, Malayalam, Gujarati, Gurmukhi and Odia. Its central decision is not to generate fonts from scratch. It restyles the glyph outlines of template fonts that are already shaping-complete, and keeps their layout machinery. The template's cmap, its GSUB closure, and its GPOS data are preserved under a documented metric policy, so every output is a complete font by construction. The style changes; the rules that make the script work are reused. The choice trades freedom for guarantees. A redrawn glyph can change how a letter looks, but not how the font behaves, because the part of the font that governs behaviour is carried over from a template that already works.

Style is chosen in natural language, through Lipika, a retrieval index over roughly six hundred and fifty open-license font families. A reference-conditioned latent diffusion model then redraws the template glyphs in the selected style. What it draws is checked before it is kept. Content gating, harmonization, and verification of shaped clusters follow, and where a redrawn glyph fails, the pipeline falls back to the template outline. The fallback is what keeps the guarantee intact. A glyph the generator cannot draw acceptably is simply left as the template drew it.

Srijika produces sixty-six TTF files: fifty-seven curated presets and nine open-vocabulary showcase fonts. All of them pass the OpenType Sanitizer, and HarfBuzz and CoreText reproduce the template's glyph-ID sequences on probes heavy with conjuncts. A full-closure audit covering about eighty-one thousand glyphs and about fifty-five thousand anchors quantifies how the metrics changed.

The evaluation is against baselines that involve no learning. On SSIM gates using diffusion-training families held out from training, simply copying the template outperforms generation on fifty of fifty-six faces. Style movement can only be measured with an internal embedding from the same model, whose training corpus includes the held-out families, so the authors say these results require caution. A learned baseline, an independent style metric, and a human study are outside the report's scope.

The contributions are the layout-reusing formulation and pipeline, its nine-script audit and benchmark, and a catalogue of negative results covering failed conditioning, objective choices, and the data-hull limits of reference-guided restyling. What the report establishes firmly is the structural half: every output is a working font. How far the style moves is the half it flags as still open.