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

Anil Pai

cs.CVSep 4, 2026

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

Authors: Anil Pai

Organizations: Loopdesk Technologies LLP

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.

Figures & tables

Panel of Figure 1 — Figure 1: The Srijika serving pipeline. A prompt retrieves donor reference glyphs via Lipika ; the diffusion model restyles the template’s glyph closure tile-by-tile; four verification stages gate, harmonize, snap, and shape-verify the result, each able to revert individual glyphs to the template outline (dashed) — so every glyph in an output font either passed verification or is the template outline, and the layout program is inherited.

			SSIM				Fam. Lipika closure
Condition	Held-out family	Faces	Gen	Tpl	Morph	Geom	Gen	Geo	Morph
OOD (gen-v3.11)	Alkatra	5	0.6377	0.6997	0.6922	0.7101	0.866	0.009	0.008
	Lohit Oriya	1	0.7364	0.7619	0.7628	0.7325	0.034	0.020	$-$ 0.003
	condition macro		0.6870	0.7308	0.7275	0.7213	0.450	0.014	0.002
Genre (gen-v3.11b)	Noto Serif Oriya	5	0.7689	0.7917	0.7903	0.7608	0.025	0.006	0.007
	Lohit Oriya	1	0.7290	0.7619	0.7628	0.7325	0.026	0.020	$-$ 0.003

Table 1: Odia split-composition study, by held-out family (the independent unit): 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 (and, symmetrically, in training membership). All methods under the instrumented protocol of Sec. 7 : family-mean SSIM for generation, template copy, target-oracle morphology, and the reference-only tile-space geometric transform, plus family Lipika similarity-gap closures for generation, geometry, and morphology; condition macro rows average the two family rows unweighted (per-face pooled SSIM, the gate statistic of Table 3 , is 0.6541 OOD / 0.7622 genre-matched). For the held-out Alkatra family, generated closure is 0.866 (0.866/0.861/0.867 across three diffusion seeds), versus 0.009 for geometry and 0.008 for morphology; the two-family OOD-condition macro is 0.450. The decorative hold-out is in effect an out-of-distribution challenge; the genre-matched hold-out is the shipped gate. The re-split is post-hoc (Sec. 5 ); both conditions are part of the record, the delta is a benchmark-composition difference between two separately trained runs — the controlled single-model version of this comparison is Table 5 , which replicates the effect. The OOD row here is qualitative support , not controlled validation: SSIM ranks the visibly style-moving output below template-like structure while neither tested baseline reproduces the movement (Fig. 2 ).

Panel of Figure 2 — Figure 2: The OOD condition, qualitatively. Top: the held-out Alkatra wght700 face under gen-v3.11, rows labeled in-figure: template, target-oracle morphology, reference-only tile-space geometric transform, generated tiles, and the Alkatra ground truth. Morphology and geometry barely move from the template — consistent with their near-zero Lipika closures in Table 1 — while generation visibly adopts the donor’s heavy rounded style at a structural cost that SSIM penalizes. Bottom: a HarfBuzz-shaped specimen of a font assembled from the same held-out checkpoint and conditioning as the raster rows (gen-v3.11, seed 7, guidance 3, template Noto Sans Oriya wght400, the eval protocol’s six Alkatra-wght700 reference tiles): 210 of 562 closure glyphs passed the content gate and are restyled, the remainder revert to template outlines — the visible weight mixing is the pipeline’s honest OOD behavior, not a curated production preset. The artifact passes ots ; SHA-256 03560c71c59f9614 … in the supplement manifest.

Figure 3: Zero-shot reconstruction of held-out Odia faces (gen-v3.11b; Lohit Oriya and two Noto Serif Oriya weights). Each block shows, top to bottom: the template glyphs (model input content), the target-oracle morphology reference (template dilated/eroded to the target’s ink coverage), generated tiles conditioned on six reference glyphs from the unseen family, and the ground-truth target. Template and morphology are nearly indistinguishable; the generated row moves stroke weight, terminals, and contrast toward the target.

	v3.7 (donor)	v3.8 (+20k)	v3.8b (+30k)	v3.12 (+30k, wtd.)
Devanagari	0.7476	0.7458	0.7408	0.7408
Tamil	0.7570	0.7493	0.7532	0.7498
Bengali	0.7687	0.7558	0.7534	0.7587
Telugu	0.7691	0.7824	0.7790	0.7828
Kannada	0.6860	0.6714	0.6887	0.6803
Malayalam	—	0.7346	0.7401	0.7298

Table 2: 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 window lengths, while new scripts improve. The degradation is not monotone in window length — Tamil and Kannada partially recover from +20k to +30k, Kannada to just above the donor — so a longer window is not a reliable fix. Re-weighted sampling over the same window (v3.12, shares restoring each old script’s pre-graft proportion) shifts mean SSIM by

\approx 0

vs. v3.8b and relocates the damage to the down-weighted scripts: in our runs dilution is zero-sum. Old-script regressions in bold .

Script	Fam.	Faces	SSIM	Fam. macro	Tpl. copy	Morph	Geom	Serving model
Devanagari	3	12	0.7476	0.7371	0.7913	0.7919	0.7787	gen-v3.7
Tamil	2	3	0.7570	0.7674	0.7598	0.7628	0.6918	gen-v3.7
Bengali	2	4	0.7687	0.7694	0.7810	0.7781	0.7653	gen-v3.7
Telugu	4	8	0.7691	0.7542	0.7797	0.7972	0.8356	gen-v3.7
Kannada	2	11	0.6860	0.6980	0.7546	0.7545	0.7481	gen-v3.7
Malayalam	4	7	0.7401	0.7397	0.7816	0.7820	0.7991	gen-v3.8b

Table 3: Shipped zero-shot gate scores: an in-distribution, genre-matched operational gate (SSIM over held-out families, test split, guidance 3), read against three no-learning references — template copy (unrestyled template glyphs scored on the same targets), density-matched morphology (template glyphs dilated/eroded to each target’s ink coverage; a target-oracle diagnostic, since the generator never sees the scored target), and Geom the coherent reference-only geometric baseline (Sec. 7 ) — plus the per-script serving route. Fam. counts held-out families (the independent unit; faces of one family are correlated), Faces the held-out faces, and Fam. macro the family-level macro-average SSIM. Bars inherited from the first shipping generation (Devanagari 0.7445, Tamil 0.7551, Bengali 0.7687) are met or exceeded; for the remaining scripts the serving model is the first to ship them, so bar

=

gate. Per-face scores and ranges are in Appendix B ; family-level results for all methods and paired style scores are in Table 6 .

Figure 4: One generated font per script from a single showcase run, each rendered from its installable TTF with HarfBuzz shaping. Donors span native-script families (Amita, Ekushey Punarbhaba, Kumar One, Baloo Tamma 2, RIT Thaara) and system templates. The Tamil row (donor Oi, an ultra-black face at the data-hull edge, Sec. 8.3 ) visibly mixes restyled ultra-black glyphs with lighter template fallbacks — the defect-tolerant pipeline providing critic-mediated functional recovery at the cost of visible style consistency: evidence of fallback functionality, not of a coherent type family on this donor.

Figure 5: Redundant conditioning dies. Per-channel conv_in weight norms after 25k graft steps. The zero-initialized skeleton channel (v4.0) settles at weight-decay equilibrium because its signal is computable from the adjacent raster channel. Content-raster dropout (v4.1) makes the skeleton non-redundant and revives the pathway $17\times$ — and heavy-donor probes still got worse , falsifying the structure-deficit hypothesis rather than the mechanism.

Lever	Attack surface	Mechanism	Deva	AlfaSlab fb
v3.6 baseline	—	—	0.7445	54%
Aux content CE	objective	$\checkmark$	0.553 †	— †
Skeleton ch. (v4.0)	input cond.	$\times$ (dead)	0.7366	56%
+ raster drop (v4.1)	input cond.	$\checkmark$	0.7403	67%
Multi-scale (v5.0)	decoder cond.	$\checkmark$	0.7363	59%
Weight aug (v5.1)	data (synth)	$\checkmark$	0.7214	64%

Table 4: Six tested-and-unsuccessful lever families. Mechanism = did the lever verifiably do its job (channel alive, aug applied, loss learning)? Deva = SSIM on the deva test split under the protocol of record (bar 0.7445). AlfaSlab fb = template-fallback rate on an out-of-hull ultra-black donor (baseline 54%). Every working mechanism either left the fallback rate unchanged or made it worse, while five of eight broke the in-hull gate. † The aux-CE row predates the protocol of record and the AlfaSlab probe (its contemporary deva baseline was 0.729, so 0.553 is a 24% collapse); it is also the one lever that did cut content errors (Lobster fallbacks

157\!\to\!48

) — by pulling outputs toward classifier centroids, i.e. it bought content by destroying style. v4.0’s mechanism check failed (skeleton channel norm

0.0125

340\times

below siblings) — which is itself the finding.

Figure 6: Objective contamination has an instant loss fingerprint. Four runs resume the same gen-v3.7 checkpoint at step 120k with the same LR schedule; the only difference is a train-time content-identity CE term (weight 0.1). Contaminated runs (red/orange) sit 6–8 $\times$ above the clean diffusion-loss regime from the first logged step and never rejoin it; clean runs (blue/green) overlap almost exactly. Gates would not reveal the failure until $\sim$ 7 GPU-hours later.

Figure 7: The data hull. Per-font median bbox-normalized ink density of all 278 Devanagari training faces, ranked. The beyond-hull probe donor AlfaSlab (0.757) is blacker than every training target including Modak (0.732); its 54% content-gate fallback rate was immovable under all six unsuccessful levers. Lobster (0.543, $\approx$ p90) sits inside the hull and restyles with 11% fallback.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

			SSIM				Lipika style
Script	Held-out family	Faces	Gen	Tpl	Morph	Geom	$\bar{s}_{\mathrm{gen}}$	$\bar{s}_{\mathrm{tpl}}$	$\bar{s}_{\mathrm{geo}}$	$\bar{s}_{\mathrm{real}}$	$p$	$p_{\mathrm{geo}}$
Devanagari	Glegoo	2	0.6953	0.7182	0.7133	0.7206	0.081	0.105	0.127	0.833	$-$ 0.033	0.029
	Karma	5	0.7553	0.8178	0.8181	0.7881	0.143	0.008	0.036	0.917	0.148	0.030
	Khand	5	0.7608	0.7940	0.7972	0.7925	0.134	0.070	0.074	0.898	0.077	0.005
Tamil	Karla Tamil Upright	2	0.7361	0.7615	0.7644	0.6910	0.477	0.168	0.143	0.941	0.400	$-$ 0.032
	Meera Inimai	1	0.7988	0.7563	0.7594	0.6932	0.894	0.133	0.125	0.910	0.979	$-$ 0.010

Appendix

Table 6: Family-level results, all methods (test split, guidance 3). Faces are averaged within each held-out family first (the family is the independent unit;

n{=}2

–

4

families per script, so all statistics are descriptive). SSIM columns: generated, template copy, the target-oracle morphology diagnostic, and the coherent reference-only geometric baseline (one global ink-bbox scale and weight radius estimated from the six references). Style columns: mean Lipika cosine to the face’s six-reference centroid for generated, template, geometric, and real target tiles (

\bar{s}_{\mathrm{morph}}

tracks

\bar{s}_{\mathrm{tpl}}

within 0.08 everywhere and is in the CSV), and the Lipika similarity-gap closures

p=(\bar{s}_{\mathrm{gen}}-\bar{s}_{\mathrm{tpl}})/(\bar{s}_{\mathrm{real}}-\bar{s}_{\mathrm{tpl}})

and

p_{\mathrm{geo}}

(same formula for the geometric baseline; both unclipped, negative when the method lands farther from the centroid than the template). Generated style exceeds the template’s on 18 of 23 families, but the median closure is 0.047 (IQR 0.016–0.133; 8 families above 0.10, 3 above 0.25, 5 negative) — the typical effect is small and the large closures concentrate in a few families. The geometric baseline is the structural mirror image: higher SSIM than generation on 17 of 23 families, but median closure 0.006 with none above 0.25. Sampling-seed sensitivity: rerunning all nine gates with two additional diffusion seeds moves pooled per-script SSIM by at most

0.006

and family macros by at most

0.009

(median

0.004

); the closure medians are 0.047/0.054/0.047 across seeds with unchanged threshold counts and a single near-zero sign flip (Lohit Malayalam); template and morphology columns are seed-invariant. Values are the protocol-of-record seed; a machine-readable per-face, per-seed CSV (including the six reference-tile indices per face) accompanies the submission.

Script	Held-out face	SSIM	SSIM p25	Tpl. SSIM	Morph	Style ratio
Devanagari	Glegoo Bold	0.6699	0.6361	0.7057	0.6966	0.075
	Glegoo	0.7206	0.6652	0.7308	0.7300	0.123
	Karma Bold	0.7273	0.6699	0.8109	0.8012	0.088
	Karma Light	0.7793	0.7367	0.8268	0.8337	0.273
	Karma Medium	0.7805	0.7434	0.8287	0.8304	0.163
	Karma	0.7549	0.7087	0.8124	0.8167	0.174

Appendix

Table 7: Per-face gate scores (test split, guidance 3): the five originally shipped scripts (gen-v3.7).

Script	Held-out face	SSIM	SSIM p25	Tpl. SSIM	Morph	Style ratio
Malayalam	Lohit Malayalam	0.6996	0.6384	0.7304	0.7245	0.220
	RIT Panmana	0.7657	0.7244	0.8150	0.8192	0.422
	RIT TN Joy Bold	0.7482	0.6912	0.8006	0.8011	0.410
	RIT TN Joy ExtraBold	0.7042	0.6500	0.7959	0.7939	0.277
	RIT TN Joy	0.7332	0.6997	0.7865	0.7892	0.424
	RIT Uroob Italic	0.7635	0.7154	0.7700	0.7718	0.087

Appendix

Table 8: Per-face gate scores, continued: the four grafted scripts (Malayalam/Gujarati on gen-v3.8b; Gurmukhi/Odia on gen-v3.11b).

Feb 25, 2026

VecGlypher: Unified Vector Glyph Generation with Language Models

Vector glyphs are the atomic units of digital typography, yet most learning-based pipelines still depend on carefully curated exemplar sheets and raster-to-vector postprocessing, which limits accessibility and editability. We introduce VecGlypher, a single multimodal language model that generates high-fidelity vector glyphs directly from text descriptions or image exemplars. Given a style prompt, optional reference glyph images, and a target character, VecGlypher autoregressively emits SVG path tokens, avoiding raster intermediates and producing editable, watertight outlines in one pass. A typography-aware data and training recipe makes this possible: (i) a large-scale continuation stage on 39K noisy Envato fonts to master SVG syntax and long-horizon geometry, followed by (ii) post-training on 2.5K expert-annotated Google Fonts with descriptive tags and exemplars to align language and imagery with geometry; preprocessing normalizes coordinate frames, canonicalizes paths, de-duplicates families, and quantizes coordinates for stable long-sequence decoding. On cross-family OOD evaluation, VecGlypher substantially outperforms both general-purpose LLMs and specialized vector-font baselines for text-only generation, while image-referenced generation reaches a state-of-the-art performance, with marked gains over DeepVecFont-v2 and DualVector. Ablations show that model scale and the two-stage recipe are critical and that absolute-coordinate serialization yields the best geometry. VecGlypher lowers the barrier to font creation by letting users design with words or exemplars, and provides a scalable foundation for future multimodal design tools.

Xiaoke Huang, Bhavul Gauri, Kam Woh Ng +12

Meta AI · UC, Santa Cruz

Vector Graphics Generation Text-to-SVG Generation Multimodal Large Language Models

Ask PDF

Oct 15, 2026

CoDeFont: One-Shot Font Generation via Multi-Scale Factor-Specific Representation Learning

One-shot Font Generation (OFG) aims to synthesize a target glyph by transferring the stylistic characteristics of a single reference glyph while preserving the structure of the target character. However, accurately capturing font style from a single reference image remains challenging due to the limited stylistic information available in a single glyph, and this challenge is further amplified for complex scripts. We propose CoDeFont, a script-agnostic OFG framework that addresses this challenge through multi-level style-content factorization and progressive style-content reconciliation. CoDeFont independently aggregates complementary style and structural information across multiple spatial levels to obtain factor-specific representations, while a contrastive objective explicitly encourages separation between style and content embeddings. During reconstruction, the disentangled representations are progressively injected through multi-level style conditioning, enabling stylistic information to be incorporated at different decoding resolutions while preserving the underlying glyph structure. We evaluate CoDeFont on structurally distinct Bengali and Latin scripts across seen and unseen font styles. Extensive quantitative and qualitative analyses demonstrate consistent improvements over existing state-of-the-art methods, highlighting the effectiveness of the proposed framework for OFG across diverse writing systems.

Rejoy Chakraborty, Prasun Roy, Saumik Bhattacharya +1

Indian Statistical Institute Kolkata · Indian Institute of Technology Kharagpur

Image Generation Disentangled Representation Learning

Ask PDF

Jun 28, 2026

Can OCR-VLMs Read Devanagari? A Stress-Test Benchmark and Post-Correction Study

OCR systems, ranging from classical engines to specialised OCR vision-language models (OCR-VLMs) and frontier multimodal LLMs, report strong results on English and Chinese document benchmarks, yet their behaviour on Indic scripts is largely uncharacterised. We benchmark ten systems on Devanagari (Hindi): classical EasyOCR; open VLMs (Qwen2.5-VL-3B, Qwen3-VL-8B, olmOCR-7B); specialised OCR-VLMs (DeepSeek-OCR, Unlimited-OCR); and frontier closed models (Gemini 2.5 Flash, Claude Opus 4.7, GPT-5.5, Mistral OCR), across four synthetic degradation conditions and 300 real printed scans. We report four findings. First, on clean rendered text all ten cluster within chrF++ 91 to 98, so synthetic text does not separate them. Second, under degradation the specialised OCR-VLMs are the most fragile: DeepSeek-OCR suffers rare but catastrophic repetition failures (outputs up to 71 the reference length) that wreck its corpus mean even though its median is the best of any system, which is why we report median and catastrophic-rate instead of the mean. Third, on real scans nine of the ten systems collapse (EasyOCR falls from chrF++ 93.6 to 58.3) and the field spreads across a 76-point range, so synthetic renders badly overstate Devanagari quality. Fourth, strong English OCR does not predict Indic OCR: GPT-5.5 drops to chrF++ 58.5 (tying classical EasyOCR) and olmOCR-7B, the model behind olmOCR-Bench, falls to 40.5, while the open Qwen3-VL-8B (75.2, runnable on a single 24 GB GPU) beats GPT-5.5 and approaches Mistral; Gemini and Claude lead at 86.3 and 82.2. An error taxonomy separates surface errors (numerals, punctuation) from structural ones (conjuncts, matras, nukta), and a byte-level (ByT5) post-corrector improves a cheap engine on its own error distribution (chrF++ +1.2 to +1.5) but does not transfer across engines. We release the benchmark, code, and models.

Aditya Pratap Singh

VLM Evaluation VLM Robustness Optical Character Recognition

Ask PDF

Abstract

Explore similar work

VecGlypher: Unified Vector Glyph Generation with Language Models

CoDeFont: One-Shot Font Generation via Multi-Scale Factor-Specific Representation Learning

Can OCR-VLMs Read Devanagari? A Stress-Test Benchmark and Post-Correction Study