Research map / Vision-language and multimodal / Multimodal LLMs

VecGlypher: Unified Vector Glyph Generation with Language Models

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces VecGlypher, a Gemma3/Llama3.3-based multimodal LLM that autoregressively emits SVG path tokens for glyphs from style tags or reference glyph images, trained in two SFT stages (Envato then Google Fonts).

Problem

Vector glyph generation typically needs curated exemplar glyph sheets and raster-to-vector postprocessing; general LLMs and vector-graphics LLMs fail to produce valid, watertight typographic outlines.

Approach

Recast glyph generation as next-token prediction over SVG path text using a multimodal LLM, unifying text-tag and image-exemplar conditioning, with a two-stage recipe (large noisy-font continuation then expert-tagged post-training) and typography-specific preprocessing (coordinate normalization, path canonicalization, family dedup, quantization).

Key results

• VecGlypher-70B (text-referenced) reaches R-ACC 100.4, CD 1.68, FID 3.34 on cross-family OOD fonts, versus Claude Sonnet 4.5 at R-ACC 46.65, CD 5.28, FID 19.59.
• Image-referenced VecGlypher-27B achieves R-ACC 99.12, CD 1.18, FID 2.32, versus DeepVecFont-v2 (R-ACC 37.86, CD 14.58, FID 115.5) and DualVector (R-ACC 49.20, CD 16.45, FID 105.5).
• Scaling from 4B to 27B and the two-stage Envato-then-Google recipe with absolute coordinates give the best geometry (text-referenced CD 1.67, FID 3.47).

Surprising findings

• Budget and flagship general LLMs (GPT-5 Mini, Gemini 2.5 Flash) rarely output valid SVG glyphs, and open-weight LLMs and OmniSVG score near zero on R-ACC.

Limitations

• Character set limited to 0-9, a-z, A-Z; closed-set vector methods hinder diacritics, pictographs and cursive.
• Small models (4B) remain brittle; around 30B parameters appears needed for stable quality.
• Envato test split excluded due to noisy tags, so evaluation relies on Google Fonts only.

Experimental setup

Datasets: Envato Fonts Google Fonts

Benchmarks: Google Fonts cross-family test split

Metrics: R-ACC Chamfer Distance CLIP score DINO score FID

Baselines: GPT-5 GPT-5 Mini Gemini 2.5 Pro Gemini-2.5 Flash Claude Sonnet 4.5 Claude Haiku 4.5 DeepVecFont-v2 DualVector OmniSVG

Builds on: Gemma 3 Llama 3.3 DeepVecFont-v2 DualVector

Closest papers by meaning

• IF-Font: Ideographic Description Sequence-Following Font Generation (NeurIPS 2024 · similarity 0.87)
• VQ-FONT: Few-Shot Font Generation with Structure-Aware Enhancement and Quantization (AAAI 2024 · similarity 0.87)
• Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models (CVPR 2025 · similarity 0.87)
• DeepCalliFont: Few-Shot Chinese Calligraphy Font Synthesis by Integrating Dual-Modality Generative Models (AAAI 2024 · similarity 0.87)
• GlyphMastero: A Glyph Encoder for High-Fidelity Scene Text Editing (CVPR 2025 · similarity 0.86)
• Vector Grimoire: Codebook-based Shape Generation under Raster Image Supervision (ICML 2025 · similarity 0.86)