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.
Figures & tables
Command
Arguments
Description
M / m (MoveTo)
x , y
Move the cursor to the end-point (x,y) without drawing anything.
L / l (LineTo)
x , y
Draw a line to the end-point (x,y) .
Q / q (Quadratic Bézier)
qx , qyx , y
Draw a quadratic Bézier curve with the control point (qx,qy) and the end-point (x,y) .
Z / z (ClosePath)
∅
Close the path by moving the cursor back to the starting position (ignorable).
Table 1: SVG draw commands used in this work. We serialize each glyph as a single SVG path using only MoveTo , LineTo , Quadratic Bézier , and ClosePath . Uppercase commands use absolute coordinates; lowercase are relative to the current point. Coordinates are quantized to one decimal; other attributes (e.g., “fill”, “stroke”, “transform”) are unused.
Figure 2: Paradigm comparisons. a) Prior image-referenced pipelines use separate image and vector encoder–decoders and a geometry post-optimizer. b) Diffusion-based approaches cascade image diffusion with a vector decoder. c) VecGlypher unifies both text- and image-referenced conditioning within a single LLM: given a style description or reference glyph images plus a target character, the model autoregressively emits SVG path tokens that detokenize to a valid SVG path. This formulation removes raster intermediates and exemplar-sheet requirements while producing directly editable vectors. A practical workflow is to first generate a few reference glyphs from text descriptions, then bootstrap with those images to synthesize the full font.
Table 5: Dataset statistics. Word clouds (left) visualize tag vocabularies for Google Fonts (expert-curated, appearance-focused) versus Envato (noisier, marketing-oriented). The middle plots show the per-font number of tags (Google: concise; Envato: capped at < 15). Right: distributions of input token length (tags) and output token length (SVG path tokens), with Envato exhibiting heavier-tailed outputs. These differences motivate our two-stage recipe: large-scale continuation on Envato, then instruction alignment on Google Fonts.
Figure 3: VecGlypher pipeline and training recipe. a) A text tokenizer or image encoder condition the LLM ( ∣∣ denotes mutually exclusive choice), which predicts the next SVG token until the path is produced; detokenization yields SVG paths that we rasterize for display only. b) Training is two-stage: 1) SFT on Envato (text-referenced only) teaches SVG syntax and long-horizon geometry; 2) SFT on Google Fonts (text- or image-referenced) aligns geometry to appearance instructions. No raster denoisers or post-optimizers are used.
Figure 4: Text-referenced ablations. Representative text-to-glyph generations across model sizes and data regimens: ground truth (GT), Google-only 4B and 27B models, Envato-only 27B, mixed E+G 27B, and two-stage E → G 27B. Using the same style tags across columns, scaling and the two-stage recipe yield cleaner closures, stable counters, and more faithful style. Please refer to the supplementary materials for comprehensive results.
Data
Size
Repr.
R-ACC ↑
CD ↓
CLIP ↑
DINO ↑
FID ↓
Google
4B
R
73.96
4.29
26.02
90.27
15.57
A
66.66
3.75
26.04
89.92
14.69
27B
R
92.81
2.31
26.38
93.01
5.81
A
94.91
1.98
26.43
93.43
3.96
Envato
27B
R
93.99
3.89
26.43
89.79
30.68
A
93.84
3.63
26.45
90.24
20.43
Table 6: Text-referenced ablations on data and model size. We compare relative (R) vs. absolute (A) coordinate serialization and several data regimens: Google-only (G), Envato-only (E), mixed E+G, and two-stage E → G. Scaling and the two-stage E → G training with absolute coordinates yield the best geometry and fidelity.
Data
Size
Repr.
R-ACC ↑
CD ↓
CLIP ↑
DINO ↑
FID ↓
Google
4B
R
83.48
2.36
25.90
93.26
9.38
A
81.94
2.11
25.89
93.11
7.94
27B
R
94.42
1.53
26.03
94.88
4.07
A
96.46
1.41
26.04
95.15
2.84
E → G I
27B
R
98.90
1.17
26.06
95.69
2.51
A
97.88
1.16
26.06
95.84
2.55
Table 7: Image-referenced ablations on data and model size. With 1—8 reference glyphs, we study Google-only 4B/27B and two-stage variants that first SFT on Envato then fine-tune on Google (“E → G I” for image-only training; “E → G T,I” for mixed training). At 27B, absolute coordinates achieve the strongest overall results, outperforming single-stage training.
Model
R-ACC ↑
CD ↓
CLIP ↑
DINO ↑
FID ↓
GPT-5 Mini
4.17
10.70
24.75
79.65
63.86
Gemini-2.5 Flash
4.89
13.78
25.26
78.62
86.03
Claude Haiku 4.5
32.38
7.60
25.61
84.69
35.07
GPT-5
43.98
6.12
25.95
86.92
29.00
Gemini 2.5 Pro
24.04
8.22
25.57
83.77
47.82
Claude Sonnet 4.5
46.65
5.28
25.99
88.31
19.59
Table 8: Text-referenced comparisons with general LLMs. We evaluate several proprietary/budget LLMs against VecGlypher on cross-family OOD fonts. Budget models rarely emit valid SVG; flagship models improve but still trail our approach. VecGlypher-70B (A) attains state-of-the-art performance.
Model
R-ACC ↑
CD ↓
CLIP ↑
DINO ↑
FID ↓
DeepVecFont-v2
37.86
14.58
24.81
79.41
115.5
DualVector
49.20
16.45
25.07
79.57
105.5
VecGlypher 27B T,I,R
99.08
1.21
26.07
95.65
2.48
VecGlypher 27B T,I,A
99.12
1.18
26.07
95.82
2.32
Table 9: Image-referenced comparisons with vector-font baselines. Against DeepVecFont-v2 ( Wang et al., 2023 ) and DualVector ( Liu et al., 2023b ) , VecGlypher-27B (T,I) delivers substantially higher recognizability and drastically lower geometric and distributional errors, demonstrating strong generalization to unseen styles.
Figure 5: Image-referenced ablations. Given 1–8 reference glyphs from a font, we compare Google-only 4B/27B with two-stage variants (E → G I and E → G T,I). The two-stage settings at 27B best transfer style and preserve thin structures and closures; Google-only baselines underperform in geometry. Please refer to the supplementary materials for comprehensive results.
Figure 6: Text-referenced comparisons to general LLMs. Rows show GT, three flagship/budget multimodal LLMs, and our VecGlypher models. General LLMs often fail to output valid, closed paths or the requested character, whereas VecGlypher (27B/70B) consistently produces watertight, stylistically faithful vectors for the same prompts. More results can be found in the supplementary materials.
Figure 7: Image-referenced comparisons to vector-font baselines. On unseen font families, DeepVecFont-v2 ( Wang et al., 2023 ) and DualVector ( Liu et al., 2023b ) struggle with thin strokes and contour closure. VecGlypher-27B (T,I) preserves delicate structures, follows target content, and closes contours. Please refer to the supplementary materials for the comprehensive comparisons.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 13Table 14
Data
Size
Repr.
R-ACC
CD
CLIP
DINO
FID
FID (C)
R-ACC (U)
CD (T)
CD (ST)
L2
LPIPS
PSNR
SSIM
Google
4B
Rel.
73.96
4.29
26.02
90.27
15.57
1.84
75.51
3.83
4.40
0.196
0.245
8.30
0.643
4B
Abs.
66.66
3.75
26.04
89.92
14.69
2.19
68.68
3.36
3.85
0.205
0.251
8.14
0.631
27B
Rel.
92.81
2.31
26.38
93.01
5.81
0.440
93.40
1.96
2.32
0.158
0.212
9.21
0.679
27B
Abs.
94.91
1.98
26.43
93.43
3.96
0.250
95.30
1.69
2.00
0.152
0.204
9.50
0.686
Envato
27B
Rel.
93.99
3.89
26.43
89.79
30.68
1.63
94.86
3.35
3.90
0.211
0.265
7.46
0.621
27B
Abs.
93.84
3.63
26.45
90.24
20.43
1.31
95.18
3.22
3.64
0.197
0.254
7.78
0.636
Appendix
Table 10: Text-referenced ablations on data and model size.
Data
Size
Repr.
R-ACC
CD
CLIP
DINO
FID
FID (C)
R-ACC (U)
CD (T)
CD (ST)
L2
LPIPS
PSNR
SSIM
Google
4B
Rel.
83.48
2.36
25.90
93.26
9.38
1.08
84.66
1.97
2.45
0.155
0.206
9.60
0.684
4B
Abs.
81.94
2.11
25.89
93.11
7.94
1.07
83.53
1.77
2.19
0.156
0.205
9.58
0.682
27B
Rel.
94.42
1.53
26.03
94.88
4.07
0.293
95.30
1.24
1.50
0.127
0.179
10.65
0.716
27B
Abs.
96.46
1.41
26.04
95.15
2.84
0.147
97.03
1.12
1.34
0.121
0.172
10.79
0.722
E-G (I)
27B
Rel.
98.90
1.17
26.06
95.69
2.51
0.116
99.26
0.910
1.09
0.109
0.159
11.22
0.736
27B
Abs.
97.88
1.16
26.06
95.84
2.55
0.101
98.43
0.910
1.09
0.107
0.156
11.36
0.740
Appendix
Table 11: Image-referenced ablations on data and model size.
Model
R-ACC
CD
CLIP
DINO
FID
FID (C)
R-ACC (U)
CD (T)
CD (ST)
L2
LPIPS
PSNR
SSIM
GPT-5 mini
4.17
10.70
24.75
79.65
63.86
13.78
4.96
10.49
10.74
0.382
0.412
4.45
0.432
Gemini-2.5 Flash
4.89
13.78
25.26
78.62
86.03
12.77
7.15
13.38
13.78
0.364
0.379
4.85
0.480
Claude Haiku 4.5
32.38
7.60
25.61
84.69
35.07
5.85
38.80
7.17
7.60
0.289
0.332
5.96
0.545
GPT-5
43.98
6.12
25.95
86.92
29.00
3.71
48.03
5.26
6.26
0.301
0.346
5.68
0.521
Gemini 2.5 Pro
24.04
8.22
25.57
83.77
47.82
7.39
27.06
7.72
8.22
0.319
0.347
5.48
0.518
Claude Sonnet 4.5
46.65
5.28
25.99
88.31
19.59
2.81
52.99
4.58
5.34
0.253
0.305
6.62
0.580
Appendix
Table 12: Text-referenced comparisons with general LLMs.
Model
R-ACC
CD
CLIP
DINO
FID
FID (C)
R-ACC (U)
CD (T)
CD (ST)
L2
LPIPS
PSNR
SSIM
DeepVecFont-v2
37.86
14.58
24.81
79.41
115.5
16.45
49.29
14.58
14.58
0.243
0.320
6.70
0.599
DualVector
49.20
16.45
25.07
79.57
105.5
14.73
65.59
16.44
16.44
0.190
0.293
7.99
0.653
VecGlypher 27B T,I,R
99.08
1.21
26.07
95.65
2.48
0.120
99.44
0.950
1.13
0.112
0.162
11.38
0.734
VecGlypher 27B T,I,R
99.12
1.18
26.07
95.82
2.32
0.095
99.82
0.930
1.10
0.110
0.158
11.35
0.736
Appendix
Table 13: Image-referenced comparisons with vector-font baselines.
Model
R-ACC
CD
CLIP
DINO
FID
FID (C)
R-ACC (U)
CD (T)
CD (ST)
L2
LPIPS
PSNR
SSIM
Llama3.3 70B Instruct
0.08
38.28
24.84
75.85
143.73
21.49
0.08
38.26
38.30
0.374
0.388
4.64
0.480
Gemma3 27B IT
0.12
17.90
24.78
74.26
162.89
23.24
0.11
17.44
17.90
0.407
0.397
4.28
0.450
Qwen3 30B A3B Instruct 2507
0.39
26.15
24.66
74.03
148.78
22.47
0.55
26.07
26.15
0.418
0.422
4.219
0.427
OmniSVG
0.01
63.70
23.26
69.35
229.38
35.75
0.01
63.69
63.69
0.496
0.452
3.50
0.356
Appendix
Table 14: Performance of the open weight LLMs and the vector graphic LLM.
Direct generation of Chinese vector fonts is a challenging and ongoing problem. A Chinese vector glyph contains complex component structure, anchor layout, and Bézier curve details, which work at different scales, but a standard vector sequence writes them together in one long sequence, making the task of vector font synthesis challenging. Existing direct vector generators often fail on complex characters, while raster-domain methods must vectorize the synthesized glyph images afterward. To address the above-mentioned problem, this paper proposes VecFontLLM, an anchor-guided multimodal large language model for direct few-shot synthesis of Chinese vector fonts. Our key idea is to generate vector glyphs through anchors rather than a standard vector sequence. Specifically, the proposed VecFontLLM first predicts and refines an anchor scaffold that fixes the coarse layout of components and contours, and then completes Bézier control points to recover local curvature and style. At test time, a confidence-guided generation chain samples multiple component candidates and continues synthesis from the highest-confidence one, improving stability for complex glyphs. This work demonstrates, for the first time, high-quality few-shot synthesis of complex Chinese vector glyphs directly in the vector domain, without raster generation or vectorization. Experiments on several Chinese font datasets show substantial improvements over existing vector font synthesis methods, competitive glyph rendering quality against raster-domain baselines, and vector command distributions close to real fonts.
Hao Yuan, Yuxuan Luo, Xing Chen +1
Fuzhou University, China · Peking University, China
Recent text-to-image models can generate high-quality images from natural-language prompts, yet controlling typography remains challenging: requested typographic appearance is often ignored or only weakly followed. We address this limitation with a data-centric approach that trains image generation models using targeted supervision derived from a structured annotation pipeline specialized for typography. Our pipeline constructs a large-scale typography-focused dataset, FontUse, consisting of about 70K images annotated with user-friendly prompts, text-region locations, and OCR-recognized strings. The annotations are automatically produced using segmentation models and multimodal large language models (MLLMs). The prompts explicitly combine font styles (e.g., serif, script, elegant) and use cases (e.g., wedding invitations, coffee-shop menus), enabling intuitive specification even for novice users. Fine-tuning existing generators with these annotations allows them to consistently interpret style and use-case conditions as textual prompts without architectural modification. For evaluation, we introduce a Long-CLIP-based metric that measures alignment between generated typography and requested attributes. Experiments across diverse prompts and layouts show that models trained with our pipeline produce text renderings more consistent with prompts than competitive baselines. The source code for our annotation pipeline is available at https://github.com/xiaxinz/FontUSE.
We introduce VectorGym, a comprehensive benchmark suite for Scalable Vector Graphics (SVG) that spans generation from text and sketches, complex editing, and visual understanding. VectorGym addresses the lack of realistic, challenging benchmarks aligned with professional design workflows. Our benchmark comprises four tasks with expert human-authored annotations: the novel Sketch2SVG task (VG-Sketch); a new SVG editing dataset (VG-Edit) featuring complex, multi-step edits with higher-order primitives; Text2SVG generation (VG-Text); and SVG captioning (VG-Cap). Unlike prior benchmarks that rely on synthetic edits, VectorGym provides gold-standard human annotations that require semantic understanding and design intent. We also provide a multi-task reinforcement learning baseline that jointly optimizes across all four tasks using rendering-based rewards. This baseline, built on GRPO with curriculum learning, trains a Qwen3-VL 8B model that achieves state-of-the-art performance among open-source models, surpassing much larger models including Qwen3-VL 235B and matching GPT-4o. We also introduce a VLM-as-a-Judge metric for SVG generation, validated through human correlation studies. Our evaluation of frontier VLMs reveals significant performance gaps, positioning VectorGym as a rigorous framework for advancing visual code generation.
Joan Rodriguez, Haotian Zhang, Abhay Puri +14
Mila, Quebec AI Institute · ÉTS Montréal · QuiverAI +11