Srijika, a newly posted research paper on Hugging Face, proposes a different route to Indic font generation: instead of drawing glyphs from scratch, the system restyles the outlines of shaping-complete template fonts while keeping the template's layout tables intact.
The approach targets nine Brahmic scripts — Devanagari, Tamil, Bengali, Telugu, Kannada, Malayalam, Gujarati, Gurmukhi, and Odia — where hundreds to thousands of conjuncts, half forms, and matra variants must stay mutually consistent under OpenType shaping. Srijika preserves the template's cmap and GSUB closure along with its GPOS data under a documented metric policy, so each output is a complete font by construction.
The pipeline produced 66 TTFs: 57 curated presets and nine open-vocabulary showcase fonts. All pass the OpenType Sanitizer, and HarfBuzz and CoreText reproduce the template glyph-ID sequences on conjunct-heavy probes. A full-closure audit covering 80,915 glyphs and 54,812 anchors quantifies the metric changes.
Style selection is handled by Lipika, a retrieval index over roughly 650 open-license font families, driven by natural-language queries. A reference-conditioned latent diffusion model then redraws template glyphs in the chosen style, followed by content gating, harmonization, and shaped-cluster verification with fallback to template outlines.
The authors report a cautionary result: against no-learning baselines, template copying outperformed generation on 50 of 56 faces in diffusion-training-family-held-out SSIM gates. They note that style movement was measurable only with an internal same-model embedding whose training corpus includes the held-out families, and that a learned baseline, an independent style metric, and a human study remain outside the report's scope.
The paper's stated contributions are the layout-reusing formulation and pipeline, its nine-script audit and benchmark, and a negative-results catalogue covering failed conditioning, objective choices, and data-hull limits of reference-guided restyling.