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

AI summary

Overview

Research area: Computer vision and generative modeling applied to font synthesis, specifically OpenType font production for nine Brahmic (Indic) scripts.

Technical level: Advanced. The paper assumes familiarity with latent diffusion models, CLIP-style embeddings, OpenType tables (cmap, GSUB, GPOS), and shaping engines such as HarfBuzz.

Scope: Srijika produces installable OpenType fonts for Devanagari, Tamil, Bengali, Telugu, Kannada, Malayalam, Gujarati, Gurmukhi, and Odia by restyling the glyph outlines of shaping-complete template fonts rather than generating fonts from scratch, and reports both an evaluation protocol and a catalogue of negative results.

What This Paper Is About

A text-ready Indic font is not a drawing exercise but a shaping-engine contract: hundreds to thousands of conjuncts, half forms, and matra variants reachable through OpenType substitution must all be present and mutually consistent, which multiplies the cost of every new design. Prevailing generative approaches output glyph images, and an image of type is not a font. Srijika's goal is to take a natural-language style description and emit an installable TTF for any of the nine scripts by restyling the complete glyph closure of a professionally engineered open-license template font, so the template's substitution machinery is reused unchanged and the output is a font by construction.

Key Contributions

1. Template-layout-reusing restyling. The template's cmap and GSUB substitution closure are retained unchanged, and GPOS structure and feature routing are reused subject to documented metric and anchor-coordinate transformations (advances widen but never narrow; anchors of translated outlines move by the same vector; pair-positioning values are retained unchanged). Layout is reused, not learned, so restyling quality and layout logic decouple.

2. Retrieval-grounded styling via Lipika. Per-(family, script) CLIP-style embeddings turn free-text prompts into reference tiles, sidestepping the absence of Indic style-attribute datasets. Retrieval grounding guarantees the style is always realizable (it exists as a font) and that licensing is auditable per donor.

3. A graft recipe that scales one backbone to nine scripts. A single backbone is trained once and grafted per script by resuming a shipped checkpoint and appending new packs by uniform concatenation, with a documented "dilution" observation and per-script routing as the engineering response.

4. An evaluation protocol and a negative-results catalogue. Family-held-out zero-shot gates with genre-matched splits, script bars, content probes, a shaped-cluster critic, and a frozen two-strata (ID/style-OOD) benchmark — plus six tested-and-unsuccessful conditioning/objective levers, a forensic study of train-time objective contamination, and a data-hull boundary analysis.

Main Findings

• Template copy beats generation on SSIM. The do-nothing baseline that scores the unrestyled template glyph against the same held-out targets posts higher pooled SSIM than the model on every script (by 0.003–0.069) and on 50 of 56 held-out faces.

• Ink-coverage matching explains part of the SSIM gap. A density-matched morphology reference lands within ±0.018 of template copy on every script and above generated SSIM on 53 of 56 faces. It is a target-oracle diagnostic, not a comparable reference-conditioned baseline, because it reads each scored target's own ink coverage.

• Style movement is small and measured by a same-model metric. On the Lipika style metric, generated tiles attain 0.14–0.67 pooled normalized similarity to the held-out face's six-reference centroid versus 0.06–0.17 for the unrestyled template. Generated tiles sit closer to the reference centroid than the template on 18 of 23 held-out families, but the median Lipika similarity-gap closure is only 0.047 (IQR 0.016–0.133): 8 families exceed 0.10, 3 exceed 0.25, 1 exceeds 0.5, and 5 are negative (minimum −0.101). The abstract states the median gap closure as 0.05. Large closures concentrate in a few families (Karla Tamil Upright 0.40, Meera Inimai 0.98, RIT TN Joy 0.26), and template copy retains higher SSIM on 20 of 23 families.

• The style metric is not independent. The paper states plainly that the frozen benchmark's 30 families (including both Alkatra script renderings) all sit inside the 595-family corpus that trained both Lipika image models — the CLIP retrieval tower and index, and the recognizer whose frozen features are the generator's style encoder and the closure metric's embedding. The judge is a family-identity classifier trained to cluster the very families it scores.

• OOD style transfer is strongly family-specific, not category-general. In the controlled two-strata benchmark (gen-bench-v1, all nine packs uniformly concatenated, 168k steps, no resume, no auxiliary losses), the held-out multilingual Alkatra family produces high closure in both its Bengali (0.935) and Odia (0.853) renderings, seed-stable, with geometry and morphology at ≤0.011, while SSIM stays below template copy. The other seven unique OOD families (Modak, Coiny, two Akayas, Chilanka, Shrikhand, Langar) close a median 0.040 (range −0.034–0.062).

• Freezing both strata did not collapse in-distribution behavior. The frozen model's 22 ID families post a closure median of 0.118 with SSIM-minus-template deltas of median −0.023, matching the shipped models' band.

• The Odia split-composition study. Two training runs with the same recipe and initialization (gen-v3.11 / gen-v3.11b, both resumed from v3.7), differing only in which family is held out, gated at 0.6541 SSIM with the decorative Alkatra family in test and 0.7622 after re-splitting Alkatra into train. The +0.108 delta is a benchmark-composition difference between two separately trained runs, and the re-split is post-hoc.

• Dilution under grafting. Grafting Malayalam+Gujarati (+42% pairs, uniform sampling) onto gen-v3.7 (@120k) degrades most previously shipped scripts relative to the donor at both +20k and +30k windows, while new scripts improve. The degradation is not monotone in window length (Tamil and Kannada partially recover

Authors’ abstract

We present Srijika, a system for producing installable OpenType fonts for nine Brahmic scripts: Devanagari, Tamil, Bengali, Telugu, Kannada, Malayalam, Gujarati, Gurmukhi, and Odia. Rather than generating fonts from scratch, Srijika restyles glyph outlines from shaping-complete template fonts. It preserves the template's cmap and GSUB closure and its GPOS data under a documented metric policy, making every output a complete font by construction. This addresses a central challenge of Indic font generation: hundreds to thousands of conjuncts, half forms, and matra variants must remain mutually consistent under OpenType shaping. Srijika produces 66 TTFs: 57 curated presets and nine open-vocabulary showcase fonts. All pass the OpenType Sanitizer, while 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 metric changes. Natural-language style selection uses Lipika, a retrieval index over approximately 650 open-license font families. A reference-conditioned latent diffusion model redraws template glyphs in the selected style, followed by content gating, harmonization, and shaped-cluster verification with fallback to template outlines. We evaluate against no-learning baselines. On diffusion-training-family-held-out SSIM gates, template copying outperforms generation on 50 of 56 faces. Style movement is measurable only with an internal same-model embedding whose training corpus includes the held-out families, so these results require caution. A learned baseline, independent style metric, and human study are outside this report's scope. Our 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.